全球 AI 发展趋势

在 Artificial Analysis 的智能指数上，Claude Opus 5.5 以 58 分领先，GPT-6 Astra 和谷歌的 Gemini 4 Argon 以 53 分并列。在由用户投票产生的 Arena 默认排名中，Argon 以 1525 分领先，几款 Claude 模型以约 1505 分紧随其后，OpenAI 最好的 GPT-5.6 Sol 排第 20。中国最好的开源模型，小米的 MiMo-V2.6-Pro，在该指数上得 46 分。Argon 目前只向部分用户开放。

On Artificial Analysis’s Intelligence Index, Claude Opus 5.5 leads at 58, and GPT-6 Astra and Google’s Gemini 4 Argon tie at 53. On Arena’s default ranking, built from people’s votes, Argon leads at 1525, a tight pack of Claude models follows at about 1505 and OpenAI’s best, GPT-5.6 Sol, ranks #20. China’s best open-weights model, Xiaomi’s MiMo-V2.6-Pro, scores 46 on the index. Argon is so far only available to selected users.

在 52 周中，Anthropic 有 51 周在 Arena 上至少有一款模型进入前五，在 Artificial Analysis 上为 44 周。只有 Anthropic 和 Google DeepMind 在两个榜单上都超过全年的三分之一。OpenAI 只在 Artificial Analysis 上超过这一门槛，xAI 只在 Arena 上超过。

Anthropic had at least one model in the top five for 51 of 52 weeks on Arena and 44 on Artificial Analysis. Only Anthropic and Google DeepMind cleared one-third of the year on each leaderboard. OpenAI cleared that threshold only on Artificial Analysis, and xAI only on Arena.

我们分析了 arXiv 上提及 21 个模型系列的 AI 论文。美国闭源模型（GPT、Claude、Gemini）在 2026 年占提及量的 42%，但份额正在被中国模型夺走。Zeta Alpha 的分析显示，在开源模型中，中国模型系列的提及占比从 2024 年的 9% 升至 31%，美国模型从 31% 降至 23%，Qwen 超过了 Llama。

We analyzed AI papers on arXiv mentioning 21 model families. While closed US models (GPT, Claude, Gemini) account for 42% of mentions in 2026, they are losing ground to Chinese models. Zeta Alpha analysis reveals that among open-weight models, Chinese families surged from 9% of mentions in 2024 to 31%, US models fell from 31% to 23%, and Qwen overtook Llama.

更便宜的智能正在扩大需求的覆盖面，这意味着模型竞争并非零和博弈。恰恰相反，分层的性能和定价让企业有更多机会在价值链上下游开拓应用场景。从更大的图景看，开源模型的 token 份额在上升，但 token 总支出也在增长，前沿模型在前沿问题上仍能获得溢价。

Cheaper intelligence is expanding the surface area of demand, which means that model competition is not a zero-sum-game. To the contrary, tiered performance and pricing is giving firms more opportunities to unlock use-cases, up and down the value chain. In the bigger scheme of things, open-weight tokens are increasing share, but overall token spend is increasing, as well, with frontier models still capturing a premium for frontier problems.

OpenAI 和 Anthropic 占 OpenRouter 请求量的 20.7%。10 月 4 日，开源模型处理了 Vercel 上 62.7% 的 token，却只获得 26.9% 的支出。由于客户样本、工作负载和价格各不相同，这些快照不能当作统一的全市场排名来解读。

OpenAI and Anthropic account for 20.7% of requests on OpenRouter. On October 4, open-weight models handled 62.7% of Vercel tokens but received 26.9% of spending. Different customer samples, workloads and prices mean these snapshots cannot be read as a single market-wide ranking.

在本报告选取的基准中，软件类占已核实引用量的 57%，仅 Terminal-Bench 一项就占 45%。统计范围是 2025 年 10 月以来新发布或更新的 58 个基准中，可查到谷歌学术引用数的 46 个。八个类别依次为：写软件、操作电脑、学习与记忆、改进 AI、做科研、经营企业、健康与安全、视听与行动。

Software accounts for 57% of verified citations in this selection; Terminal-Bench alone contributes 45%. Totals sum Google Scholar counts for 46 of 58 benchmark releases introduced or refreshed since October 2025.

Epoch AI 在首批 15 个基准审查中发现 9 个存在实质缺陷。评分逻辑出错、环境可被钻空子、评测条件不统一，都会扭曲用来比较模型的结果。审查方法是：检查任务、评分逻辑、提示词、工具和资源限制；每个基准随机抽 50 个任务（不足 50 个则全查），排查隐藏要求、错误评分和捷径；抽样中 ≥20% 有问题或评分存在系统性问题即判为不合格。结果是 9 个有缺陷（如 Terminal-Bench 4.0.0、SWE-bench Verified），4 个通过但有保留（如 PostTrainBench v1.1），2 个信息不足（CritPt、FrontierCode）。

Epoch AI finds substantive flaws in nine of its first 15 benchmark reviews. Broken scoring, exploitable environments, and uneven evaluation conditions can distort the results used to compare models. • Epoch inspects tasks, scoring logic, prompts, tools, and resource limits. Insufficient access means “Not enough info.” • Reviewers sample 50 random tasks, or all if fewer, checking for hidden requirements, wrong scoring, and shortcuts. • The default failure threshold is ≥20% faulty sampled tasks or systemic grading issues. Unfair setups and inconsistent versions can also fail.

2025 年报告中列出的每一项头条评测，都已达到或接近满分。在 Epoch AI 的独立测试中，AIME 最高分 100%，GPQA Diamond 95.8%，FrontierMath 第 1–3 级 93.7%，ARC-AGI-2 在 11 个月内从 18% 升到 95%。GPQA Diamond 第 1 名与第 5 名只差 1 分，且约 5%–10% 的题目标注有误，满分本身达不到；ARC-AGI-2 的单题成本同期从 7.14 美元降到 1.12 美元，GPT-6.1 Sol 以 0.25 美元拿到 94.2%。2026 年 1 月，Artificial Analysis 因区分度不足剔除了 MMLU-Pro、AIME 2025 和 LiveCodeBench。

Every headline eval in the 2025 State of AI Report has hit or neared its ceiling. On Epoch AI's independent runs the best AIME score is 100%, GPQA Diamond 95.8% and FrontierMath Tiers 1 to 3 93.7%, and ARC-AGI-2 went from 18% to 95% in eleven months. • GPQA Diamond: the #1 and #5 models are 1 point apart, and Epoch estimates 5 to 10% of questions are mislabeled, so the ceiling sits below 100. • ARC-AGI-2: 18.3% (GPT-5 Pro, Oct 2025) to 95.0% (GPT-6 Astra, Sep 2026), as cost per task fell from $7.14 to $1.12. GPT-6.1 Sol now scores 94.2% for $0.25. • In January 2026, Artificial Analysis dropped MMLU-Pro, AIME 2025 and LiveCodeBench because they no longer discriminate.

FrontierMath 第 4 级按研究级数学设计，原本预期能撑好几年。在 Epoch 修订后的 v2 题集上，GPT-5 在 2025 年 8 月得 22%，GPT-5.4 Pro 在 3 月得 59%，Claude Fable 5 在 6 月得 90%；9 月 GPT-6 Astra 达到 98%，GPT-6.1 Sol 解出了全部 41 道私有题。OpenAI 资助了 FrontierMath，能接触最初 50 道第 4 级题中的 30 道，但 Claude Opus 5.5 同样得到 95%。Epoch 追踪的 49 个未解研究问题中，有 9 个被认定由 AI 解出，其中 4 个没有人类帮助；GPT-6 Astra 为其中一个被评为“重大进展”的问题给出了主要证明。Surge 的私有 Riemann-Bench 仍有空间，GPT-5.6 Sol 以 74.4% 领先。

FrontierMath Tier 4 was built as research-level math that should hold out for years. On Epoch's corrected v2 set, GPT-5 scored 22% in August 2025, GPT-5.4 Pro 59% in March and Claude Fable 5 90% in June. In September GPT-6 Astra hit 98% and GPT-6.1 Sol solved all 41 private problems. • OpenAI funded FrontierMath and has access to 30 of the original 50 Tier 4 problems, but Claude Opus 5.5 also scores 95%. • Epoch now tracks 49 unsolved research problems and credits AI with nine solutions, four without human help. GPT-6 Astra supplied the main proof of one rated a major advance. • Surge's private Riemann-Bench still has headroom: GPT-5.6 Sol leads with 74.4%.

2026 年 3 月推出的 ARC-AGI-3 包含数百个完全可交互、类似回合制电子游戏的环境，难度远高于 ARC-AGI-2 的静态视觉谜题。人类得分 100%，前沿模型起初只有 0.5%。7 月 Claude Opus 5 达到 30.2%，9 月 GPT-6 Astra 改变了局面，GPT-6.1 Sol 随后达到 52.7%。Astra 在标准 ARC Prize harness 上得 62.7%，换用可保存状态的适配器后得 99.9%，速度快 3.66 倍、token 少用 49%；在 96% 的关卡中，它的操作效率高于人类中位数。ARC 的开发者认为 Astra 的新型符号建模能力是“重大突破”。

In March 2026, ARC-AGI-3 introduced hundreds of fully interactive, video game style turn-based environments, meant to be much harder than ARC-AGI-2’s static visual puzzles. Although humans score 100%, initial scores by frontier models were a measly 0.5%. But then Claude Opus 5 reached 30.2% in July, GPT-6 Astra changed the game in September and GPT-6.1 Sol followed at 52.7%. • Astra achieved 62.7% on the standard ARC Prize harness and 99.9% with a state-persistent adapter, which ran 3.66x faster on 49% fewer tokens. • It outperforms the median human in action efficiency on 96% of levels. • Astra is considered a "major breakthrough" by ARC developers for its novel symbolic modeling capabilities.

哪些评测还能把模型区分开？图中每个点是一个基准：横轴为已公布的最高分，纵轴为第 1 名与第 5 名的差距。ARC-AGI-3 和 MirrorCode 仍是区分度最大的两项，差距分别为 55 分和 46 分。Terminal-Bench 4.0、APEX-Agents 1.1 和 FrontierSWE v2 的差距为 16、14、11 分。CritPt 仍然很难（最高分 32%），但前三名相差不到 0.6 分。APEX 最高分已到 82%，前三名差 8.7 分，不再紧挨在一起。DeepSWE 前五名在 70%–74% 之间，FrontierCode 1.1 的分布只有 2.6 分。这些都是点估计，测试的模型名单也不同，不是显著性检验。

Which evals still tell models apart? We plot each benchmark’s best published score against its first-to-fifth model gap. ARC-AGI-3 and MirrorCode remain the widest separators, at 55 and 46 points. • Terminal-Bench 4.0, APEX-Agents 1.1 and FrontierSWE v2 have first-to-fifth gaps of 16, 14 and 11 points. • CritPt remains hard (top score 32%), yet its top three are within 0.6 points. APEX now tops out at 82%, with an 8.7-point top-three gap: its leaders are no longer as tightly clustered. • Other leaders remain close: DeepSWE’s top five score about 70–74%. FrontierCode 1.1 spans 2.6 points. These are point estimates across different tested model rosters, not significance tests.

FrontierChallenge 和 EarthVerse 测试科学分析和自然灾害调查，OSWorld 2.0 测试报销等办公流程。部分得分按满足的单项要求给分：计算或文档正确就能得分，哪怕报告没写完、报销单没提交。在 FrontierChallenge 的 97 个科学工作流中，GPT-5.6 Sol 平均得 87.9 分，但完整完成的任务只有 20.6%；在 EarthVerse 的 405 项调查中，它答对 82.1% 的必答项，只有 34.8% 的任务达到 95% 的完成门槛；在 OSWorld 2.0 的 108 项电脑任务中，Opus 5 拿到 77.7% 的部分分，完成 44.3%。漏掉审批或没提交表单，会让原本有用的工作停在半路。

FrontierChallenge and EarthVerse test scientific analyses and natural-hazard investigations, while OSWorld 2.0 tests office workflows such as expense claims. Partial scores credit individual requirements met. A correct calculation or document can earn points even when the report is incomplete or the claim is never submitted. • On 97 scientific workflows in FrontierChallenge, GPT-5.6 Sol scores 87.9/100 but fully completes only 20.6% of tasks. • On EarthVerse's 405 natural-hazard investigations, the same model gets 82.1% of required answer items right. Only 34.8% of tasks reach the 95% threshold. • On OSWorld 2.0's 108 computer tasks, Opus 5 earns 77.7% partial credit but completes 44.3%. Missed approvals or unsubmitted forms can leave otherwise useful work unfinished.

METR 的“50% 时间跨度”指智能体有一半把握完成的任务长度，以人类专家所需小时数计；它衡量任务难度，不是无人监督的运行时长。最强的模型已经到达量表上测试样本稀少的一端，更长的时间跨度越来越难可靠测量。估计值从 Opus 4.5 的 4.9 小时，升到 Opus 4.6 的 12.0 小时，再到 Mythos Preview 早期版本的 17.4 小时，后者的 95% 置信区间为 8.5–55.1 小时。人类需要超过 16 小时的任务只有 5 个，METR 自己提示 16 小时以上的结果不可靠。GPT-6 Astra、Fable 5.1 和 Opus 5.5 都还没有估计值；6 月 GPT-5.6 Sol 的估计值随作弊行为的计分方式大幅变化，METR 认为所有结果都不稳健。

METR’s 50% time horizon is the length of task, measured in human expert hours, that an agent is predicted to complete half the time. It measures task difficulty, not time without human oversight. The strongest models now reach the sparsely tested end of the scale, so longer horizons are becoming harder to measure reliably. • Estimates rose from 4.9h for Opus 4.5 to 12.0h for Opus 4.6 and 17.4h for early Mythos Preview. The latter's 95% confidence interval spans 8.5-55.1h. • Only five tasks take humans more than 16h, so estimates up there are sensitive to scoring choices. METR warns that results above 16h are unreliable. • METR's dashboard has no estimate for GPT-6 Astra, Fable 5.1 or Opus 5.5. June's GPT-5.6 Sol estimate swung with how cheating was scored, and METR called none of the results robust.

FrontierSWE 给智能体 20 小时，在统一 harness 中开发软件；MirrorCode 要求根据文档和可运行的参考程序重建软件，最多允许 7 天和 100 亿 token。两个测试选出的领先者不同。FrontierSWE 上，Astra 得 65.5%，Opus 5.5 得 62.3%，单次运行成本分别为 1,030 美元和 99 美元。MirrorCode 上，Opus 5.5 在最高推理强度下 77% 的任务通过全部隐藏测试，Astra 在高强度下为 47%。两项分数在不同预算下衡量不同结果，都不能确立一个通用的编程排名。

FrontierSWE gives agents 20 hours to build software in a common harness. MirrorCode asks them to rebuild programs from documentation and a runnable reference, allowing up to seven days and 10B tokens. These different tests produce different leaders. • FrontierSWE: Astra scores 65.5% versus Opus 5.5’s 62.3%, at $1,030 versus $99 per trial. • MirrorCode: Opus 5.5 passes every hidden test on 77% of tasks at max effort, versus Astra’s 47% at high effort. • The scores measure different outcomes under different budgets. Neither establishes a universal coding ranking.

前沿模型需要学会在变化的环境中完成长周期任务。General Reasoning 为此构建了 KellyBench：一个模拟英超博彩市场，智能体需要自己搭建机器学习模型，对比赛下注，在一个赛季内追求收益最大化。所有模型平均都亏钱，说明前沿模型在超长周期任务上仍有能力缺口。智能体按时间顺序获得 2023–24 赛季的阵容、赛果、公开赔率和比赛数据，起始资金 10 万英镑。12 个模型在 5 次运行中平均全部亏损，6 个至少破产一次；Opus 4.7 最终剩 9.6 万英镑，GPT-5.5 剩 9 万，多数模型差得多。按 52 项专家评分标准，策略复杂度最高的 GPT-5.5 也只得 30%。

Frontier models must learn to solve long-horizon tasks in changing environments. To test this, General Reasoning built KellyBench, a simulated English Premier League market where agents build ML models to trade on matches and maximize gains over a season. Every model lost money on average, highlighting a remaining capability gap for frontier models on ultra-long-horizon tasks. • Agents receive the 2023-24 season’s lineups, results, public odds and match statistics in sequence, with £100k to build models, find an edge and size bets. • All 12 models lose money on average over 5 seeds, and six go bankrupt in at least one. Opus 4.7 ends with £96k and GPT-5.5 with £90k, but most do far worse. • GPT-5.5’s strategies score highest on a 52-point expert rubric for sophistication, at just 30%.

E-CommerceBench 给 18 个模型各 10 万元人民币，在 365 个模拟日内经营最多 4 家网店。它由阿里巴巴和香港科技大学基于淘宝、天猫数据构建，测试采购、定价和现金管理。高回报可能掩盖供应商筛选薄弱、从经验中学得很少的问题。GPT-5.6 Sol 五次运行平均年末资产 143 万元，但18.48% 的采购支出流向了欺诈供应商，Opus 4.7 只有 0.12%。环境中有 6,900 种商品、576 家供应商和变化的需求。在重复采购中，18 个模型里有 16 个没有表现出压价能力的明显进步，一年的经营很少转化为更好的采购。

E-CommerceBench gives 18 models CNY 100k to run up to four online stores for 365 simulated days. Built by Alibaba and HKUST using Taobao and Tmall data, it tests purchasing, pricing and cash management. A high return can hide weak supplier screening and little learning from experience. • GPT-5.6 Sol averages CNY 1.43M over five runs, but sends 18.48% of order spending to fraudulent suppliers. Opus 4.7 sends 0.12%. • Prices and deals follow rules, unlike Vending-Bench’s LLM-driven suppliers. Agents face 6.9k products, 576 suppliers and changing demand, with an LLM voicing replies. • On repeat purchases, 16 of 18 models show no clear progress in bargaining prices down. A year of trading rarely translates into better buying.

DiG-bench 给智能体 70 个从未在网上出现过的文字游戏，用纯文本把“发现”单独作为考察对象：每一回合，智能体能看到当前局面和可选动作，但看不到规则和获胜条件，必须自己试出来。70 个游戏中 49 个不公开，以减少数据污染和针对性调优；每个游戏都至少有一名人类首次接触就解出，说明规则可以被发现。Opus 5 解出 50/70，但前沿模型在最难的两个等级合计只解出 9/20，通用智能体框架也没有帮助。瓶颈在发现规则，不在执行：给出真实规则但不给策略时，Gemini 3.1 Pro 从 18/70 升到 69/70。

DiG-bench gives agents 70 text-based games that have never appeared online. Text is used to isolate discovery as • Forty-nine of the 70 games are private to limit contamination and tuning. Every game was solved by at least one human on first exposure, so its rules are discoverable. • Opus 5 solved 50/70, but frontier models together solved only 9/20 in the two hardest tiers. General-purpose agent harnesses did not help. • Rule discovery, not execution, is the bottleneck: Gemini 3.1 Pro rose from 18/70 to 69/70 when given the true rules but no strategy.

详细的技术报告、代码库和训练日志，让构建强模型比三年前透明得多，从预训练到推理和智能体行为的大部分路径已经公开。但规模定律的指导仍不完整，数据配比仍靠试验；训练强模型仍需要研究判断（所谓的“品味”）和足够的算力来验证大量想法。具体做法上，训练变得更有选择性：英伟达 Nemotron 3 Super 先用 20 万亿 token 覆盖广度，再用 5 万亿 token 强调质量；阶跃星辰 Step 3.5 Flash 后期侧重代码、推理和工具使用，同时保留部分通用数据。更长的上下文可能削弱其他能力：百万 token 训练期间数学成绩下降，英伟达加入了 1M 与 4K token 序列交替的阶段来缓解。损失曲线正常也可能掩盖组件故障：阶跃星辰发现部分 MoE 专家失活或出现极端数值，而整体指标看起来正常。

Detailed technical reports, codebases and training logs make building capable models less opaque than three • Training becomes more selective: Nemotron 3 Super uses 20T tokens for broad coverage, then 5T emphasizing quality. Step 3.5 Flash later emphasizes code, reasoning and tool use, retaining some general data as it specializes. • Longer context can weaken other skills. After math scores dipped during million-token training, NVIDIA added a phase alternating 1M-token and 4K-token sequences to mitigate the decline. • Healthy training loss can hide failing components. StepFun found MoE experts becoming inactive or developing extreme internal values while aggregate metrics looked normal. Monitoring individual experts exposed these problems.

今年出现了几份文档完整的智能体训练流程开源复现，效果接近最先进水平。开源复现的智能体模型仍落后前沿几步，但它们公开的技巧降低了新实验室训练智能体的门槛，Thinking Machines 和 OpenThoughts Agent 就是例子。Meta 的 ScaleRL 报告了超过 40 万 GPU 小时的消融实验，涉及精度修正（fp32 logits）、RL 目标函数和批大小，许多发现推翻了小规模实验的结论。Cursor 的 Composer 2 技术报告公开了大量长程 rollout 细节，例如在 rollout 中途更新行为策略权重，以及修复采样器与训练器不一致的方法（如 MoE 路由重放）。DeepSeekMath-v2 详细公开了智能体数据生成流程；Kimi K3 介绍了如何训练不同推理强度、兼容多种 harness，以及防止奖励作弊的机制。

Several well-documented open reproductions of agentic training pipelines were published this year, achieving • Meta’s ScaleRL reports 400k+ GPU hours of ablations covering precision fixes (fp32 logits), RL objectives and batch size. Many findings reverse conclusions from smaller scales. • Cursor’s Composer 2 technical report includes many long-horizon rollout details such as updating behaviour policy weights mid-rollout, and fixes for sampler/trainer inconsistency such as MoE router replay. • DeepSeekMath-v2 shares their agentic data-generation pipeline in detail. Kimi K3 details ways to train reasoning efforts, compatibility with multiple harnesses, and anti-reward-hacking machinery.

实验室会隐藏模型的推理过程，因为竞争对手可以用它做蒸馏训练。Panfilov 等人发现，更便宜的模型能还原更强模型被隐藏的推理，厂商随后修补了这个漏洞。限制获取推理轨迹，会不会拉大开源模型与闭源模型的差距？具体做法是：把 Opus 4.8 返回的加密推理块交给 Haiku 4.5，Haiku 会把推理过程以明文写出；同样的攻击在 OpenAI 和谷歌也有效。推理轨迹展示的是老师的解题步骤，不只是答案；此前的研究中，还原出的轨迹让一个 7B 学生模型在 MATH500 上从 68.4% 提高到 76.0%。关闭这条获取途径，可能让前沿推理更难被用于训练。论文没有证明近期开源模型的进步在多大程度上依赖它，也无法判断差距是否会拉大，报告作者认为可能会。

Labs hide their models' reasoning because rivals can train on it using a process called distillation. Panfilov and colleagues found that cheaper models could reveal stronger models' hidden reasoning. Providers patched the flaw. Could restricting access to these traces widen the gap between open-weight and closed models? • Given Opus 4.8's encrypted reasoning block, Haiku 4.5 wrote out the reasoning in plain text. The same attack worked at OpenAI and Google. • A reasoning trace shows the teacher's steps, not just its answer. In prior work, reconstructed traces lifted a 7B student from 68.4% to 76.0% on MATH500. • Closing this extraction route could make frontier reasoning harder to obtain for training. The paper does not establish how much recent open-model progress depended on it, or whether the capability gap will widen - we suspect it could.

长程智能体 RL 的大部分算力花在推理采样（rollout）上，而不是反向传播；它要在真实的沙箱环境中运行有状态的采样，还要不断把新权重同步到分布式推理集群，同时避免轨迹过时。推理占据了 RL 算力的大头：1T 参数、35B 激活的 MoE 模型 MAI-Thinking-1 在 4,864 块 GB200 中用了 4,096 块做推理，约为学习端的 5 倍，单次生成最长 12.8 万 token。CPU 和内存需求也在上升：Kimi K3 使用了 5,120 万个支持暂停、恢复、分叉、快照和亚秒级启动的有状态沙箱；DeepSeek 的 DSec 在一个部署单元中用 3 万个 CPU 核和 250 TB 内存，每天服务 300 万个沙箱实例，但约 90% 的沙箱平均只用了申请 CPU 的 5% 以下，资源共享和空闲内存回收因此非常关键。训练环境越来越像真实工作：Kimi K3 模拟了 Gmail、Notion、Slack 和 Canvas，并接入 Kimi Code、Claude Code 和 Codex 等 harness。推理端也可以分散部署：Composer 2 每个训练步通过 S3 同步权重，增量压缩让 1T 参数模型的更新只有几 GB，使跨地域的 RL 推理成为可能。

Long-horizon agentic RL spends much of its compute on inference (rollouts) rather than backprop, runs stateful • Inference dominates RL compute: MAI-Thinking-1, a 1T-A35B MoE, uses 4,096 of 4,864 GB200s for inference, roughly five times its learner allocation, with generations up to 128k tokens. • CPU and memory demand rises too: Kimi K3 used 51.2M stateful sandboxes with pause/resume, fork, snapshot, and sub-second launch. DeepSeek’s DSec reports 30,000 CPU cores and 250 TB of RAM serving 3M sandbox instances daily in one deployment unit. Yet ~90% of sandboxes average ≤5% of requested CPU capacity, making resource sharing and idle-memory reclamation essential. • Training environments increasingly resemble real work: Kimi K3 combines mocks of Gmail, Notion, Slack, and Canvas with harnesses including Kimi Code, Claude Code, and Codex. • Inference can be distributed: Composer 2 synchronizes weights every training step through S3. Delta compression reduces updates for a 1T-parameter model to a few GB, enabling geographically distributed RL inference.

RL 智能体从总是成功或总是失败的任务中学不到什么。微软的 TaskPilot 在真实代码仓库中生成编程题和隐藏测试，再根据当前模型的尝试调整难度；随着 FrogNano 变强，新一批任务让训练始终处在它刚好能解决的边缘。Agent-World 生成了 1,978 个环境和 1.98 万个工具，用于训练 8B 和 14B 智能体，在 23 个基准上提升了工具使用、规划和编程能力。FrogNano 用更简单的工具接口和 5 批共 1,500 个任务训练，Qwen3.5-4B 在 SWE-bench Verified（验证集）上解出 61.5%，在未见过的 SWE-bench Pro 问题上解出 37.6%。检验仍然重要：在 GPT-5.2 固定的情况下，Universal Verifier 把 1% 的失败电脑操作误判为成功，WebJudge 为 7%，WebVoyager 为 10%。

RL agents learn little from tasks they always pass or always fail. Microsoft's TaskPilot generates coding problems and hidden tests in real repositories, then revises their difficulty using the current model's attempts. As FrogNano improves, new task batches keep its training near the edge of what it can solve. • Agent-World generated 1,978 environments and 19.8k tools to train 8B and 14B agents, improving tool use, planning and coding across 23 benchmarks. • FrogNano combines a simpler tool interface with five batches totaling 1.5k tasks. Qwen3.5-4B solves 61.5% of SWE-bench Verified (validation) and 37.6% of held-out SWE-bench Pro issues. • Checks still matter, of course. With GPT-5.2 fixed, Universal Verifier falsely accepted 1% of failed computer-use runs, vs. 7% for WebJudge and 10% for WebVoyager.

蒸馏是让学生模型模仿老师。离线策略（off-policy）使用老师或早期模型的回答，在线策略（on-policy）使用当前学生自己的尝试；自蒸馏指老师是同一模型的副本。迁移成本更低：阿里巴巴从同一个 Qwen3-8B 检查点出发比较两种方法，在线策略蒸馏用 1,800 GPU 小时在 AIME24 上达到 74.4%，RL 用 1.79 万 GPU 小时只达到 67.6%。组合专家：小米分别训练数学、编程和指令遵循三个老师以避免相互干扰，再用多老师在线策略蒸馏（MOPD）合并它们的能力。向自己学习：Self-Distilled Reasoner 把标准解法交给一个冻结的副本，由它在学生自己的尝试上逐 token 指导，不需要更大的老师。

Distillation teaches a student to imitate a teacher. Off-policy uses teacher or earlier-model responses, while on-policy uses the current student’s attempts. Self-distillation means the teacher is a copy of the same model. • Cheaper transfer: Alibaba compared two methods from the same Qwen3-8B checkpoint. On-policy distillation reached 74.4% on AIME24 using 1.8k GPU-hours. By contrast, RL needed 17.9k GPU-hours to reach 67.6%. • Combining specialists: Xiaomi trains math, coding and instruction-following teachers separately to avoid interference, then combines their skills via Multi-Teacher On-Policy Distillation (MOPD). • Learning from itself: Self-Distilled Reasoner gives a frozen copy the worked solution. It guides the student token by token on the student’s own attempt, so no larger teacher is required.

自博弈自己生成训练任务，因此是很有吸引力的训练方式。SPICE 用基于文档的问题对推理模型做后训练；零数据自博弈则从随机权重开始，通过程序学习。自然文本和 DNA 数据只用于选择超参数，不提供预训练梯度。SPICE：Qwen3-4B-Base 在 11 个推理基准上的总成绩从 35.8% 提高到 44.9%（+9.1 个百分点），同样评测下 R-Zero 为 39.5%，Absolute Zero 为 40.7%。零数据自博弈：在给出足够的上下文示例后，反转字符串、栈和联想回忆任务的精确匹配率接近 100%，评测时不更新权重。

Self-play creates its own training tasks, so is an appealing approach to AI training. SPICE post-trains a reasoner on document-grounded questions, whereas zero-data self-play learns from programs, starting from random weights. Natural text/DNA validation selects hyperparameters but supplies no pretraining gradients.

AlphaEvolve 用一个冻结的模型搜索解法；TTT-Discover 则在推理过程中更新 gpt-oss-120b 的权重，奖励有前景的尝试并复用最好的候选。目标是为当前问题找到一个更好的解，而不是得到一个能泛化到新任务的模型。TTT-Discover 改进了两个数学界限；TriMul 运行时间在 A100 上比人类最佳成绩缩短 51.5%（4,531→2,198 微秒），在 H100 上缩短 15.3%（1,371→1,161 微秒）。一项消融实验中，复用有前景的候选比更新权重更重要；一个注意力算子和第三个数学问题没有超过领先者。TTPO 用多数投票的答案训练，在三个数学基准上让 Qwen3-1.7B 从 38.0% 提高到 45.2%，不需要提供答案标签。

AlphaEvolve searches for solutions using a frozen model. TTT-Discover, meanwhile, updates gpt-oss-120b’s weights during inference time, rewarding promising attempts and reusing the best candidates. The goal is one better solution to the current problem, rather than a model that generalizes to new tasks. • TTT-Discover improved two mathematical bounds. TriMul runtime fell 51.5% on A100 (4,531→2,198 μs) and 15.3% on H100 (1,371→1,161 μs), versus the best human entries. • Reusing promising candidates mattered more than weight updates in one ablation. An attention kernel and a third math problem missed the leaders. • TTPO trains on majority-vote answers. Across three math benchmarks, Qwen3-1.7B rose from 38.0% to 45.2% without supplied answer labels.

在一条轨迹中积累的经验会改变序列模型的 token 分布，但这些经验放在上下文里成本很高，清空上下文就会消失。新的持续学习方法把上下文中学到的有用行为固化到持久记忆或模型权重中，让一条轨迹的经验能改进下一条轨迹的表现。Monash 与字节跳动 Seed：经验蒸馏保留了至少 64.8% 的上下文学习收益，直接对同样的经验做 SFT 只保留 3.8%；在 749 个软件工程任务和 6 个文字冒险游戏上，它以至少少 9.6 倍的环境交互达到 RL 基线的水平。微软研究院：在线经验学习提取可复用的经验教训，多轮固化到权重中；在 Frozen Lake 上，Qwen3-1.7B 到第三轮时平均回答长度降到初始的约 70%，准确率同时提高，分布外的指令遵循能力基本保持。Physical Intelligence：π*0.6 把机器人自主运行的经验和专家纠正训练回 VLA 模型，在部分最难的任务上吞吐量提高一倍以上，失败率约减半。

Experience gathered within a trajectory changes a sequence model's token distribution. This experience is • Monash / ByteDance Seed: Experience Distillation retains at least 64.8% of in-context learning gains, versus 3.8% for direct SFT on the same experience. Across 749 software engineering tasks and six text adventure games, it matches RL baselines with at least 9.6x fewer environment interactions. • Microsoft Research: Online Experiential Learning extracts reusable lessons and consolidates them into model weights over successive rounds. On Frozen Lake, Qwen3-1.7B reduces average response length to roughly 70% of its initial level by round three, alongside improved accuracy. Separate tests largely preserve out-of-distribution instruction-following performance. • Physical Intelligence: π*0.6 trains autonomous robot experience and expert corrections back into the VLA, more than doubling throughput and roughly halving failures on some of its hardest tasks.

线性注意力把前面的 token 压缩进一个固定大小的记忆，降低长上下文的成本；全注意力保留对每个 token 的直接访问，而这种压缩可能丢失信息。通义千问把两者结合，让线性注意力成为混合模型中有用的组件。Qwen3.8-Flash-Next 用约九分之一的训练算力，在 14 个基准中的 8 个上超过了前代，这一提升也包含其他架构和训练改动。其他实验室也在尝试：Kimi Linear 在 100 万 token 上下文、单序列生成时比全注意力快 2.3 倍；月之暗面也在 2.8 万亿参数的 Kimi K3 中使用了它的 delta 注意力技术。

Linear attention summarizes earlier tokens in a fixed-size memory, reducing the cost of long contexts. Full attention keeps direct access to individual tokens that this summary can lose. Qwen combines the two, making linear attention a useful component of a hybrid model. • Qwen3.8-Flash-Next beats its predecessor on 8 of 14 benchmarks using about a ninth of the training FLOPs. This gain includes other architecture and training changes. • Other labs are trying this too. Kimi Linear generates tokens 2.3x faster than full attention at 1M context, one sequence at a time. Moonshot also uses its delta-attention technique in the 2.8T-parameter Kimi K3.

当一块 GPU 只服务一个用户时，从显存中搬运模型权重的时间，可能比计算下一个 token 的时间还长。谷歌的 DiffusionGemma 先起草一整块文本，再并行修改，让 GPU 一次做更多工作，从而加快 token 输出，代价是回答质量有所下降。DiffusionGemma 对一个 256 token 的文本块进行多轮修改，然后转到下一块；额外训练可以减少所需轮数，图中的四步只是示意。谷歌报告在专用 GPU 服务一个或少数用户时，token 输出最快提高 4 倍；标准的 Gemma 4 仍能给出更高质量的回答。

When a GPU serves one user, moving model weights from memory can take longer than computing the next token. Google's DiffusionGemma drafts a whole block of text, then revises it in parallel. This gives the GPU more work at once and speeds up token output, with a trade-off in answer quality. • DiffusionGemma revises a 256-token block over several passes, then moves to the next. Extra training reduces the passes needed. The four steps shown are illustrative. • Google reports up to 4x faster token output on dedicated GPUs serving one or a few users. Standard Gemma 4 still produces higher-quality answers, though.

在模型固定的情况下，更换 harness 让 SWE-Bench Mobile 上的成绩提高了 6 倍；Meta-Harness 让 Haiku 4.5 在 TerminalBench-2 上的成绩从使用 Claude Code 时的 27.5% 提高到 37.6%。要有意义地比较模型，必须公开所用的 harness，并保持固定或系统地改变它。在 WeaveBench 上，LongHorizon-Harness 让 Qwen 3.7-Plus 超过了 Opus 4.7 的基线成绩。一项对照实验在 100 个 SWE-bench Verified 任务上把三个模型和三个 harness 两两组合：GLM-5.1 从最简 harness 的 52.5% 升到完整 harness 的 65.5%，harness 带来的性能差异是模型带来差异的 7.8 倍。

With each model fixed, changing the harness delivered a 6x gain on SWE-Bench Mobile, while Meta-Harness raised Haiku 4.5 from 27.5% with Claude Code to 37.6% on TerminalBench-2. Meaningful model comparisons require disclosing the harness and holding it fixed or varying it systematically. • On WeaveBench, LongHorizon-Harness improves Qwen 3.7-Plus enough to exceed the baseline performance of Opus 4.7. • A controlled test paired three models with three harnesses on 100 SWE-bench Verified tasks. GLM-5.1 rose from 52.5% to 65.5% between the minimal and full harnesses. Harness-induced performance variance was 7.8x model-induced variance in this experiment.

单一 harness 平均表现可能不错，但会漏掉需要不同方法的任务。Meta、杜克大学和加州大学戴维斯分校的研究者在不同开发任务上进化出两个 harness，再为每个新问题选择其中一个。在执行模型固定的情况下，这在未见过的数学和编程任务上超过了 Meta-Harness。使用 Gemini 3 Flash 时，两个分支有 47% 的共同覆盖，但各自都能解出对方解不出的题。路由器在 Gemini 数学任务上达到 62%，Meta-Harness 为 46%，但低于两个分支合计可覆盖的 67%。编程也有提升：在 Sonnet 4.5 固定的情况下，Terminal-Bench 2.0 从 44.8% 升到 50.0%，SWE-bench Lite 从 63.6% 升到 66.0%。

A single harness can work well on average yet miss tasks that need a different approach. Researchers at Meta, Duke and UC Davis evolve two harnesses on different development tasks, then choose one for each new problem. With the execution model held fixed, this improves held-out math and coding results over Meta-Harness. • With Gemini 3 Flash, the branches share 47% coverage, but each solves some problems the other misses. • The router reaches 62% on Gemini math versus Meta-Harness's 46%, below the branches' combined potential coverage of 67%. • Coding also improves: Terminal-Bench 2.0 rises from 44.8% to 50.0%, and SWE-bench Lite from 63.6% to 66.0%, with Sonnet 4.5 fixed.

许多 harness 加入了复杂的行为规则、提醒和规划机制，用来弥补旧模型的不足。但随着模型能力提高，成功的智能体行为又被训练回模型中，这些补丁变得多余，甚至会拖累更强的模型。harness 应该提供让环境更易操作的接口，而不是在模型行为上叠加冗余。Claude Code 为高级模型删掉了 80% 的系统提示词，编程评测没有可测量的损失；少样本的工具使用示例会减少探索，好的接口设计更重要。harness 已经变成课程：在不断演进的 harness 中产生的轨迹可以更新模型权重（见 Continual Harness），近期的 GPT 模型就是在 Codex harness 上训练的。随着模型变强，轻量和复杂 harness 之间的差距在缩小；长期值得投入的是环境接口（工具、MCP 服务器）、状态管理、可观测性和智能体基础设施（沙箱、运行时等）。

Many harnesses have included elaborate behavioral rules, reminders and planning mechanisms to compensate for • Claude Code removed 80% of the system prompt for advanced models, with no measurable loss on coding evaluations. Few-shot tool use examples are also found to reduce exploration, with good interface design being more important. • Harnesses have become curricula, where trajectories produced inside an evolving harness can update model weights (see Continual Harness). Recent GPT models are trained on the codex harness. • Performance between lightweight and sophisticated harnesses shrink as model capability improves: the durable harness investments that remain are environment interfaces (tools, MCP servers), state management, observability and agent infrastructure (sandboxes, runtimes, etc).

长输入在填满上下文窗口之前就可能让模型应付不过来。MIT 的递归语言模型（RLM）把输入放在代码工作区里，模型可以查看输入，并把其中的片段交给新的（递归的）模型调用处理，这样一个固定的模型就能处理一次读不完的输入。作者用 RL 在短任务上训练 Qwen3-30B-A3B，发现能力能迁移到长得多的输入和使用相同分解策略的新领域。作者的解释是：把陌生任务变成熟悉的子问题，每次模型调用都停留在训练分布之内。迁移的前提是学到可复用的分解方法；RLM 并不总是更便宜，递归调用在推理时使用固定权重。

Long inputs can overwhelm a model even before they fill its context window. MIT's Recursive Language Models (RLMs) keep the input in a code workspace, where the model can inspect it and delegate pieces to further (recursive) model calls. This lets a fixed model work through inputs too large to read in one pass. • The authors trained Qwen3-30B-A3B with RL on short tasks and found transfer to much longer inputs and new domains sharing a decomposition strategy. • The authors explain this as turning unfamiliar tasks into familiar subproblems: each model call stays “locally in-distribution” with its training. • Transfer depends on learning a reusable decomposition, and RLMs are not uniformly cheaper. Recursive calls use fixed weights at inference.

技能把可复用的指令和代码打包，记忆为后续任务保存信息。开发者可以在两次模型发布之间，通过改变智能体能复用的内容来改进它。按宽口径检索，与“技能”相关的论文在 2025 年 1–8 月为 152 篇，2026 年同期为 1,486 篇。

Skills package reusable instructions and code, while memory preserves information for later tasks. These let developers improve an agent between model releases by changing what it can reuse. Papers matching the broad skills query rose from 152 to 1,486 in January-August 2025 and 2026.

RSI 指自主地让 AI 构建更好的 AI。3 月，Andrej Karpathy 发布了 autoresearch：智能体修改一个小型 nanochat 模型的训练代码，固定训练 5 分钟，保留有效的改动，在一块 GPU 上一夜重复约 100 次。用户只需编辑 program.md，即定义这个“研究机构”的指令文件。Karpathy 说：“这个仓库讲的是一切如何开始的故事。”该仓库 5 个月内获得约 9.5 万 GitHub 星标和 1.34 万次分叉。固定 5 分钟的训练预算让每次实验可以相互比较，一块 GPU 一夜能跑约 100 次实验，以验证集每字节比特数打分。

RSI is the pursuit of AI that builds better AI, autonomously. In March, Andrej Karpathy released autoresearch, a • The repo reached roughly 95,000 GitHub stars and 13,400 forks in 5 months. • A fixed 5-minute training budget makes every experiment comparable and turns one GPU into roughly 100 experiments a night, scored on validation bits per byte.

与“可验证奖励”相关的论文，2026 年 1–8 月是 2025 年同期的 10.4 倍，RSI 相关论文为 2.7 倍。现实的机会在于用可靠的评分来自动化实验；选择有价值的研究问题仍是一个独立的（而且非常有价值的）问题。

Papers matching verifiable rewards grew 10.4x in January-August 2026 versus the same months of 2025, compared with 2.7x for RSI. The practical opportunity is to automate experiments with reliable scoring. Choosing valuable research questions remains a separate (and very valuable) problem...

Darwin Gödel Machine 会改写智能体的脚手架，即冻结模型外围的提示词、工具和工作流程；Hyperagents 还会改写“执行这些修改的过程”本身。这能把更多智能体工程自动化，但更好的智能体并不一定会成为更好的“下一代智能体发明者”……至少目前还没有。Huxley-Gödel Machine 按后代的得分选择智能体，使用 GPT-5 时解决了 57.0% 的 SWE-bench Lite 问题，SWE-agent 为 56.7%。Red Queen 让智能体和评估者共同进化，在相同搜索成本下，经 AI 评审的论文接收率达到 38.8%，HGM-H 为 21.8%。Weco 报告在 100 次试验中超过了自己手工调优的智能体，但用新智能体去改进后续智能体，没有带来显著的效率提升。

Darwin Gödel Machine rewrites an agent’s scaffold: the prompts, tools and workflow around a frozen model. Hyperagents also rewrite the procedure that makes those edits. This can automate more of agent engineering, but a better agent does not necessarily become a better inventor of future agents...yet! • Huxley-Gödel Machine selects agents by their descendants’ scores. With GPT-5, it fixed 57.0% of SWE-bench Lite issues vs. SWE-agent’s 56.7%. • Red Queen co-evolves agents and evaluators. Its paper acceptance rate rose to 38.8% vs. HGM-H’s 21.8% at matched search cost, as rated by AI judges. • Weco reports beating its hand-tuned agent in 100 trials. Using the new agent to improve further agents gave no significant efficiency gain, however.

PostTrainBench 测试自主后训练：每次运行针对一个小型基础模型和一个基准，在一块 H100 上运行 10 小时。1.2 版覆盖 4 个基础模型和 6 个基准。结果衡量的是针对具体任务的优化；官方指令模型是用不同资源训练的参照。在 1.2 版上，Fable 5.1 得 44.6%，Opus 5.5 得 43.8%，GPT-6 Astra 得 41.9%，官方指令模型为 48.4%。BFCL 因任务质量和针对性训练的问题被移除，此前 95% 的结果不足以证明通用函数调用能力更强。这些是针对具体基准的模型适配，不是一个全面变强的助手。五次 Fable 5.1 的 GPQA 运行以 Opus 5 作为后备。

PostTrainBench tests autonomous post-training: each run targets one small base model and one benchmark, with 10 hours on one H100. Version 1.2 covers four base models and six benchmarks. Results measure task-specific optimization; the official instruct models are reference points trained with different resources. • On v1.2, Fable 5.1 scores 44.6%, Opus 5.5 43.8% and GPT-6 Astra 41.9%, versus 48.4% for official instruct models. • BFCL was removed after task-quality and item-targeted training concerns. The former 95% result is insufficient evidence of general function-calling superiority. • These are benchmark-specific model adaptations, not one broadly improved assistant. Five Fable 5.1 GPQA runs used Opus 5 as a fallback.

Prime Intellect 让 18 个模型在 nanoGPT 优化器速通任务中运行了 153 次：用更少的步数把一个 1.24 亿参数的 GPT 训练到 3.28 的目标损失，使用 8 块 H200，不联网。Fable 5 最好的一次连续运行了 8.7 天，在这一设置下报告的新意有限。Fable 5 从 3,290 步的基线做到 2,726 步，缩小了与 2,600 步人类纪录之间 81.7% 的差距。研究预算各不相同：24 小时后，Fable 5 达到 3,010 步，Opus 5 为 3,045 步。强智能体会在多个随机种子上复测边缘结果，并回头重试早期的失败，但这里新意有限，并不能证明 AI 研究存在普遍上限。

Prime Intellect tested 18 models in 153 runs on the nanoGPT optimizer speedrun: train a 124M-parameter GPT to a target loss of 3.28 in fewer steps. Runs used 8xH200 GPUs without internet. Fable 5's best trajectory lasted 8.7 days, with limited novelty reported in this setting. • Fable 5 reached 2,726 steps from a 3,290-step baseline, closing 81.7% of the gap to a 2,600-step human record claim. • Research budgets differed: after 24 hours, Fable 5 had reached 3,010 steps, versus 3,045 for Opus 5. • Strong agents re-tested borderline results across seeds and revisited earlier failures, but limited novelty here does not establish a general limit on AI research.

研究者设计了“影子评估”：把一篇高质量、尚未发表的 NeurIPS 投稿的核心研究问题交给智能体（因此不可能存在数据污染），再请论文原作者以审稿人身份给产出打分。每个智能体有 6 天时间、3,000 美元 API 额度、GPU 和开放网络。Opus 4.8 在没有人类帮助的情况下完成了全部工程工作，但两篇论文分别得 2/6 和 1/6 分，都被明确拒稿。反复出现的五种失败模式：不清楚发表的标准、修补方案缺乏创意、回溯无效、资源意识差、偏离指令。两次运行结束时都只花了不到一半的 API 预算，尽管模型能实时看到用量。智能体自我评审了几十轮，从未给出“接收”，它指出的问题往往正是人类审稿人后来提出的问题。

Researchers introduced shadow evaluations, in which they give an agent the central research question of a • Opus 4.8 completed all the engineering without human help, but both its papers were scored 2/6 and 1/6 - unambiguous rejections. • Five failure modes recurred: no sense of the publishable bar, uncreative fixes, ineffective backtracking, poor resource awareness and instruction drift. • Both runs ended with under half the API budget spent, despite the models knowing their usage in real-time. • Self-review by agents never returned an acceptance in dozens of rounds and the issues it flagged were often the same ones the human reviewers later raised.

Anthropic 在一篇关于构建 RSI 的文章中公布了员工代码贡献的数据（左图）：2026 年第二季度的人均产出是 2025 年以前的 8 倍，与 Mythos Preview 的使用时间相对应。OpenAI 也看到了同样的结果（右图）。

In an essay on building RSI, Anthropic (left chart) shared data on the code contributions of their employees. The

更有意思的是，研究人员认为 Mythos Preview 建议的下一步方向，有 64% 的情况比人类研究者实际选择的路线更好。这表明前沿模型正在产出体现研究品味的高质量贡献。

Even more interestingly, researchers rated next-direction turns suggested by Mythos Preview as better than the

Anthropic 的内部自动化指数追踪 Claude 在模型研发中承担的责任。被评为“AI 主导”的工作占比从 2 月的不到 1% 升至 2026 年 8 月的 26%：Claude 根据高层级的提示完成任务的大部分，人类负责监督。超过 90% 的被测工作已有 AI 的大量协作或主导，但还没有任何工作是完全自主的。

Anthropic’s internal automation index tracks how much responsibility Claude takes on across model R&D. The

2026 年 1 月，OpenAI 研究人员用 AI 智能体自主完成通常需要人类 4–8 小时的任务，成功率为 18%。到 2026 年 7 月，同样 18% 的成功率已能在需要人类 32–64 小时的复杂得多的任务上达到。

In January 2026, OpenAI researchers used AI agents to autonomously complete tasks normally requiring 4-8

到目前为止，OpenAI 内部研究中编程智能体的大部分用途集中在执行环节：编写研究基础设施代码，启动、监控和调试训练任务，以及技术支持和审查。决定研究什么、怎么研究，编程智能体还没有解决。

So far, the majority of coding agent use in OpenAI internal research has been on execution-related workflows,

模型一旦通过阈值或利用评分漏洞，公开的 AI 研发测试套件就无法再排除其具备高级能力。实验室越来越依赖内部的研究加速证据。RSI-Exam 在外部测试了 88 项任务，没有模型达到按前沿水平校准的参照分。METR 对 Opus 5.5 的审查引用了约 1.5 倍 AI 驱动研究加速的初步估计，时间段未说明。OpenAI 报告每个人类工作日对应 3.1 个智能体工作日，这衡量的是活动量，不等于研究产出的同等增长。Noam Brown 估计加速约 3 倍，受限于串行实验和 GPU 供应。

Public AI R&D suites no longer rule out advanced capabilities once models pass their thresholds or exploit their • METR’s Opus 5.5 review cites a preliminary estimate of ~1.5x AI-driven research acceleration. The estimated time period is unspecified. • OpenAI reports 3.1 agent-workdays per human workday. This measures activity, not an equivalent gain in research output. • Noam Brown estimates about a 3x speedup, limited by serial experiments and GPU supply.

Thinking Machines 预览了一个连续交互模型，在一个学习得到的循环中处理交互，而不依赖外部拼接。此前构建在多模态模型上的“实时”体验，是把语音识别、视觉、对话和语音合成串成一条流水线。交互模型仍可调用外部异步的后台推理模型或调用工具。交互模型把时间切成约 200 毫秒的微回合，取消了传统的“用户说完、助手再说”的轮次边界；每一步都能看、听、说，重叠、沉默、打断和视觉提示都成为原生行为；每一步也处理文本 token，用来调用工具、与环境交互或唤起推理模型。在直播电商是重要零售渠道的中国，实验室正在投资交互模型：面壁智能的 MiniCPM-o 4.5 是一个 9B 全双工交互模型，可在内存小于 12GB 的端侧设备上运行；字节跳动的 SeedRealtime 是另一个例子。

Thinking Machines previewed a continuous interaction model that handles interaction in one learned loop, rather • Interaction models chunk time into ~200ms micro-turns, removing the traditional user/assistant turn boundary • At each step the models see, listen and speak, enabling overlaps, silence, interruption and visual cues to become native behaviour • The models also process text tokens at each step to call tools, interact with the environment or invoke a reasoner model. • In China, where livestreaming is a major retail channel, labs are investing in interaction models. MiniCPM-o 4.5 is a 9B full-duplex interaction model, which can run on edge devices with <12GB RAM. ByteDance’s SeedRealtime is another example.

多数视频模型在输入提示并等待几秒后才返回一段完成的视频。专注媒体推理的公司 fal 基于 MiniMax 开源的 H3 音视频模型打造了 H3 Max，加入了提示遵循和美学数据，以及与模型一同优化的服务引擎和 GPU 内核。fal 报告不到 3 秒就能生成 5 秒视频，吞吐量约为官方 H3 接口的 35 倍。其“导演模式”把 39 帧带入每个新片段，并记住之前的提示；去掉重叠帧后，相邻片段可以连续播放，形成实时流的感觉。在 fal.live 上，观众投票决定接下来发生什么：以每秒 24 帧播放，提示会引导下一个片段，让互动电视成为可能，只是变化出现前会有延迟。

Most video models return a finished clip after you’ve prompted it and waited a couple seconds. Media-focused inference company fal took MiniMax’s open weights H3 video and audio model to create H3 Max, which adds data for prompt following and aesthetics, plus a serving engine and GPU kernels optimized alongside the model. • fal reports generating a five-second clip in under three seconds, about 35x the throughput of the official H3 endpoint. • Their Director mode carries 39 frames into each new segment and remembers prior prompts. It drops overlapping frames so successive segments play continuously, giving you the real-time streaming feel. • On fal.live, viewers vote on what happens next. At 24 FPS, prompts steer the next segment, opening interactive TV with a delay before changes appear.

世界模型预测一个动作之后会发生什么。反复进行这种预测，就形成一段“想象中的推演”。智能体可以用它来做规划、生成训练经验或测试行为。后面几页的例子覆盖生成的世界、学习得到的状态表示、驾驶和机器人控制。注：图中展示的是三种可能的用途，与后面几页并非一一对应。

A world model predicts what happens after an action. Repeating those predictions creates an imagined rollout. Agents can use it to plan, generate training experience or test behavior. The following examples cover generated worlds, learned state representations, driving and robot control.

DeepMind 把 Genie 3 生成的世界与 SIMA 2 智能体结合：Gemini 出题，给每次尝试的视频打分，并提供用于再训练的反馈。提升能迁移到未见过的世界，但这一初步结果以导航为主，且使用 Gemini 自己的评分标准。在 Genie 3 生成的城市世界中训练后，SIMA 2 在几乎所有训练任务上都有提升，在 0–100 分的评分标准上常常提高 25 分以上（DeepMind，2025 年 12 月）。提升可以迁移：在多数未见过的自然世界任务上，自我改进后的智能体优于初始版本。作者称这是初步的、以导航为中心的证据，评分者与驱动整个循环的是同一个 Gemini 模型。

DeepMind paired Genie 3’s generated worlds with SIMA 2. Gemini proposes tasks, scores video of each attempt and supplies feedback for retraining. Gains transfer to held-out worlds, but this initial result centers on navigation and uses Gemini’s own scoring rubric. • Trained on urban Genie 3 worlds, SIMA 2 improves on nearly all training tasks, often by 25 points or more on the 0 to 100 rubric (DeepMind, Dec 2025). • The gains transfer: on most held-out natural-world tasks, the self-improved agent beats the initial one (figure). • The authors call this initial, navigation-centered evidence, scored by the same Gemini model that drives the loop.

Odyssey 用《暗黑破坏神 2》的视频及对应的操作和游戏状态进行训练，学习参与者之间如何互动。4 名人类和 16 个 AI 智能体可以共享一个由他们集体行动塑造的模拟世界。一个模拟 Transformer 预测参与者如何相互影响；一个共享状态追踪他们的位置和互动，包括离开某个玩家视野之后的情况。一个流匹配视频 Transformer 为每个玩家生成各自的视角画面；训练时去掉视觉历史，迫使模型使用提供的状态，并对推动剧情的角色加大误差权重。在用学习得到的模型替代游戏引擎的同时，编程智能体也在帮助开发者用成熟的修改工具和逆向代码改造、组合现有游戏。

Odyssey trains on Diablo II video paired with actions and game state, learning how participants interact. Four humans and sixteen AI agents can share a simulation shaped by their collective actions. • A simulation transformer predicts how participants affect one another. A shared state tracks their positions and interactions, including when they leave a player's view. • A flow-matching video transformer generates each perspective. Training removes visual history so it must use the supplied state, and gives extra weight to errors on the characters driving the action. • Alongside efforts to replace game engines with learned models, coding agents are helping developers modify and combine existing games, using established modding tools and reverse-engineered code.

机器人规划路线时，需要预测自己的动作会导致什么结果。Meta 的 V-JEPA 2.1 通过预测视频的表示（而不是重建像素）来学习视觉特征。在导航世界模型中使用这些特征，规划时间缩短约 10 倍，轨迹精度保持不变。给定当前视角和目标图像，规划器在导航世界模型（NWM）设置下搜索一条两秒的路线。基线也在潜空间中规划，但通过重建图像来学习表示；Meta 改变了表示、预测目标和采样方法。规划只需 8 步迭代而不是 128 步，四个数据集上的平均轨迹误差几乎不变（3.03 对 2.98）。

A robot planning a route needs to predict where its actions lead. Meta’s V-JEPA 2.1 learns visual features by predicting representations of video, rather than reconstructing pixels. Using those features in a navigation world model cuts planning time roughly 10x while preserving trajectory accuracy. • Given a current view and a goal image, the planner searches for a two-second route using the Navigation World Models (NWM) setup. • The baseline also plans in latent space, but learns it by reconstructing images. Meta changes the representation, prediction target and sampling method. • Plans need 8 refinement steps instead of 128, with average trajectory error nearly unchanged across four datasets (3.03 vs. 2.98). Note: Open-loop benchmark. Colors show learned features in three example plans.

2023 年 GAIA-1：预测驾驶视频，用文字和动作引导下一 token 预测和视频解码；单摄像头，用 4,700 小时伦敦驾驶数据训练。2025 年 3 月 GAIA-2：生成全景，用潜扩散生成五个同步摄像头的画面；84 亿参数，可控制英国、美国和德国的交通和天气。2025 年 12 月 GAIA-3：改写录制场景，改变本车路径，同时保留录制的交通；150 亿参数，训练数据约为 GAIA-2 的 10 倍。2026 年 8 月 GAIA-4：让驾驶系统接管控制，摄像头和雷达画面随 AI 驾驶员的决策生成；可以比较不同模型，或在录制场景中重新运行干预。

Odyssey-3 是一个自回归扩散 Transformer，用视觉观察进行预训练，模拟世界随时间的演变。再用观察—动作配对训练一个解码器，把它的表示转化为控制指令；公司在真实道路驾驶、双臂操作、人形机器人和游戏上评估。在主干冻结的情况下，20 小时的仿真驾驶数据就训练出了道路驾驶策略，仿真训练的策略在两次安全员干预之间行驶的距离，达到真实数据训练策略的 77%。几十小时的机器人数据支撑了机械臂操作和 Flexion 人形机器人的控制，机械臂在没有相应示范的情况下，能从抓取失误中恢复。用约 2 小时 GTA 画面训练的策略，无需再训练就在《荒野大镖客 2》中实现了骑马移动，这是早期的迁移例子。

Odyssey-3 is an autoregressive diffusion transformer pretrained on visual observations to simulate how the world evolves over time. Observation-action pairs then train a decoder to turn its representations into controls, which the company evaluates on real-world driving tasks, dual arm manipulation, humanoids and gameplay. • With the backbone frozen, 20 hours of simulated driving data produced a road-driving policy. Simulation-trained policies reached 77% of real-data policies' distance between safety-driver interventions. • Tens of hours of robot data supported manipulation and Flexion's humanoid control. Robot arms recovered from missed grasps without demonstrations of those recoveries. • A policy trained on ~2 hours of GTA footage produced horseback movement in Red Dead Redemption 2 without retraining, an early example of transfer.

Skild 的 S1 显示，上下文学习能力随规模出现。在未见过的任务上，只给一段视频提示、不做微调，模型的成功率从 1 千小时预训练时的约 0%，升到 10 万小时时的 66%；用相同数据和算力训练、以语言提示的 VLA 只有 9%。Sunday Robotics 在叠衣服任务上看到同样的形态：随着预训练规模扩大，熟悉环境与陌生家庭之间的差距从 82% 降到约 0%。

Skild's S1 shows in-context learning emerging with scale. On unseen tasks, one video prompt and zero fine-tuning sees the model climb from ~0% success at 1k pre-training hours to 66% at 100k hours, while a language-prompted VLA trained on identical data and compute stays at 9%. Sunday Robotics sees the same shape on laundry folding: the in-domain vs unseen-home gap falls from 82% to ~0% with pretraining scale.

机器人通过示范学习操作物体，示范把“看到什么”和“如何移动”联系起来。人们可以通过操控机器人、使用手持夹爪或边工作边佩戴摄像头来提供示范。每种方式把记录的动作转换为机器人动作的方法不同。遥操作：一个人操控机器人，同时记录机器人的观察和动作，是直接的机器人示范。手持夹爪（UMI）：追踪人手持夹爪的路径和开合，无需机器人就能示范。第一人称人类视频：模型从视频中估计手和手腕的动作，再迁移到机器人上，可以大规模采集人类示范。

Robots learn to manipulate objects from demonstrations that connect what they see with how to move. People can provide these by controlling a robot, using handheld grippers, or wearing cameras as they work. Each approach differs in how the recorded movements are translated into robot actions.

直接用多样的机器人数据训练，模型会在不同策略之间取平均，学到的策略反而变差。为此，每段训练数据都附上丰富的上下文标注，例如机器人应该做什么、动作是怎么完成的，这样示范、失败案例和网络数据都能成为有效信号。π0.7 的提示包括子任务的文字描述、生成的子目标图像，以及速度、质量、错误和控制方式等元数据。这些标签让一个模型可以同时用失败案例、自主运行数据和专家示范训练，再在推理时指定想要的行为。在叠衣服任务上，加入带元数据的低质量数据后吞吐量提高，不带元数据则下降；去掉任务多样性最高的 20% 数据，三个未见任务的成功率也随之下降。

Naive training on diverse robot data would average across strategies and degrade the learned policy. To fix this, • π0.7’s prompt combines subtask language, generated subgoal images, and metadata for speed, quality, mistakes, and control mode. • Those labels let one model train on failures, autonomous rollouts, and expert demonstrations, then request the desired behavior at inference time. • On laundry, throughput improved as lower-quality data was added with metadata, but fell without it. Removing the most task-diverse 20% also cut success on three unseen tasks.

一个厨房任务会用新的顺序组合熟悉的动作：打开容器、放东西进去、再关上。模仿学习往往学的是整个序列，经典规划器则需要手写规则。宾夕法尼亚大学的 SymSkill 从未分段的示范中同时学到可复用的动作和使用这些动作的条件。一个视觉语言模型离线识别物体；物体之间的相对位置定义了学到的条件，例如执行某个动作前夹爪必须在哪里。它在 12 个单步 RoboCasa 任务上达到 85% 的成功率；另外，一台真实的 Franka 机械臂从 5 分钟的自由操作中学习，就能执行最多 12 步的任务。规划器会重新排列技能，并在受到干扰后恢复。重新组合已练习过的动作，避免了为每一种可能的顺序收集示范。

A kitchen task combines familiar actions in a new order: open a container, put something inside, then close it. Imitation learning often learns the whole sequence, while classical planners need hand-written rules. Penn's SymSkill learns both reusable motions and the conditions for using them from unsegmented demonstrations. • A vision-language model identifies objects offline. Relative object positions define learned conditions, such as where the gripper must be before an action. • It reaches 85% success across 12 single-step RoboCasa tasks. Separately, a real Franka learns from five minutes of play and performs tasks up to 12 steps. • A planner reorders skills and recovers after disturbances. Recombining practiced movements avoids collecting demonstrations of every possible sequence.

复杂的机器人任务分为“决定下一步做什么”和“执行动作”两部分，常被称为系统 2 / 系统 1 架构：一个推理规划器指挥一个快速的运动策略。谷歌的 Gemini ER 2 监控执行过程并调整指令，英伟达的 Vesta 则用记忆跨多步规划。ER 2 跟踪实时视频并重试失败的动作；在谷歌用真实 VLA 进行的内部测试中，把 ER 1.6 换成 ER 2，成功率从 48.6% 提高到 60.0%。Vesta 把 8B 的 Qwen3-VL 规划器与 GR00T N1.6 结合，在计数、找抽屉和隐藏物体任务上，比单独使用执行模型的成功率高 38.3 个百分点。联合训练在导航和具身推理上也超过了 Vesta 的专项消融版本。真机证据仍有限，只在一套 YAM 设备上每项任务做了 20 次试验。

Complex robot tasks split into deciding what to do next and executing the movement. This is often called a System 2 / System 1 architecture: a reasoning planner directs a fast motor policy. Google's Gemini ER 2 monitors execution and adjusts its instructions, while NVIDIA's Vesta uses memory to plan across multiple steps. • ER 2 tracks live video and retries failures. In Google's internal tests with a real VLA, replacing ER 1.6 with ER 2 raises success from 48.6% to 60.0%. • Vesta pairs an 8B Qwen3-VL planner with GR00T N1.6. On counting, drawer-search, and hidden-object tasks, it adds 38.3 points of success over the actor alone. • Joint training also beats Vesta's specialist ablations in navigation and embodied reasoning. Real-robot evidence remains limited to 20 trials per task on one YAM setup.

长时间的装配在不同阶段看起来可能很相似，最新的摄像头画面不一定能说明下一步该做什么。多数机器人策略只看很短的历史。RoboTTT 给 GR00T N1.7 加上自适应记忆，帮助它跟踪进度、从错误中恢复，而无需保存每一帧。在三个 YAM 装配任务上，RoboTTT 的任务进度为 79%，无记忆为 42%，使用循环 Gated DeltaNet 记忆为 56%。快速权重在观察到达时就进行压缩；训练上下文最多 8,000 个时间步时性能持续提高，而推理成本保持不变。在 5 分钟的 Gear Bot 装配任务上，它在 10 次试验中完成 2 次，所有基线都是 0 次。更好的任务跟踪很有希望，但可靠地完成任务仍然困难。

Long assemblies can look similar at different stages, so the latest camera frame may not reveal what to do next. Most robot policies see only a short history. RoboTTT adds an adaptive memory to GR00T N1.7, helping it track progress and recover without retaining every frame. • Gear Bot: a five-minute, ten-stage assembly • Across three YAM assembly tasks, RoboTTT scores 79% task progress versus 42% without memory and 56% with recurrent Gated DeltaNet memory. • Fast weights compress observations as they arrive. Performance improves with up to 8k timesteps of training context, while inference cost stays constant. • On the five-minute Gear Bot assembly, it finishes 2/10 trials versus zero for every baseline. Better task tracking is promising, but reliable completion is still difficult.

在真机上测试机器人策略成本很高，而人工搭建的仿真器可能给出错误的排名。SimFoundry 从视频中重建可交互的场景，并生成相关的物体、布局和任务，检验这些场景中的得分能否预测真机表现。七项任务测试了 π0、π0.5、GR00T N1.6/N1.7 和 DreamZero 等策略，仿真与真机得分的平均相关系数为 0.911。深度、分割和 3D 模型恢复出物体和可动部件；物理检查让场景稳定，新的布局扩充了训练数据。这可能减少模型选择所需的硬件。证据覆盖的是选定的桌面任务，可选人工修正，并非任意环境。

Robot policies are costly to test on hardware, while hand-built simulators may rank them incorrectly. SimFoundry reconstructs interactive scenes from video and generates related objects, layouts, and tasks. It tests whether scores in those scenes predict performance on real robots. • Real scenes become interactive variations • Seven tasks test policies including π0, π0.5, GR00T N1.6/N1.7, and DreamZero. Simulated and real scores have a mean correlation of 0.911. • Depth, segmentation, and 3D models recover objects and movable parts. Physics checks stabilize the scene, then new layouts expand the training data. • This could reduce hardware needed for model selection. Evidence covers selected tabletop tasks, with optional manual refinement, not arbitrary environments.

PPO 能训练机器人，但会丢弃过去的经验；Soft Actor-Critic 能复用经验，但在复杂身体上可能不稳定。FlashSAC 使用更大的网络和批量，同时减少更新次数；批归一化和分类损失稳定了动作价值的学习，限制了误差放大。4,096 个仿真的宇树 G1 在一块 A100 上用 4 小时学会爬楼梯，PPO 需要近 20 小时。控制器无需进一步微调就能在真实楼梯上运行。在十个仿真器的 60 多项任务中，FlashSAC 的性能和训练速度都超过 PPO 和 FastTD3，在高维控制上尤其明显。迁移采用随机化物理参数和成熟的仿真到现实设置；它的吸引力在于一个可复用的、更快的学习配方，建立在此前的迁移成功之上。

PPO trains robots but discards past experience. Soft Actor-Critic reuses it but can be unstable on complex bodies. FlashSAC uses larger networks and batches with fewer updates. Batch normalization and a classification loss stabilize action-value learning and limit error amplification. • 4,096 simulated Unitree G1s learn stair climbing in 4h on one A100, versus nearly 20h with PPO. The controller runs on real stairs without further fine-tuning. • Across 60+ tasks in ten simulators, FlashSAC improves performance and training speed over PPO and FastTD3, especially on high-dimensional control. • Transfer uses randomized physics and an established sim-to-real setup. The draw is a reusable, faster learning recipe, building on earlier transfer successes.

人手和机器人手不同，照搬关节位置可能得到一个会打滑、或者掀不开盖子的抓取。OmniRetarget 和 VideoMimic 保留接触的几何关系；CHORD 更进一步，奖励那些能对物体施加相似力和力矩的接触。CHORD 用人手与物体互动的示范训练 RL 策略，它匹配的是接触点“能够承受的力”，而不是测量人实际用了多大的力。在 1,831 个仿真任务中，它报告了 82.1% 的成功率。消融实验显示它优于只奖励接触位置的方法：位置接近的接触，力学上不一定等效。真机测试使用 Dexmate 平台、Sharpa 灵巧手和动作捕捉；控制器跟随固定示范，不会重新规划，因此规模上的结论来自仿真。

Human and robot hands differ, so copying joint positions can produce a grasp that slips or cannot lift a lid. OmniRetarget and VideoMimic preserve contact geometry. CHORD goes further, rewarding contacts that can exert similar forces and torques on the object. • CHORD trains RL policies from human hand-object demonstrations. It matches the forces contacts could support, rather than measuring the person's actual force. • Across 1,831 simulated tasks, it reports 82.1% success. Ablations beat contact-position-only rewards: nearby contacts need not be mechanically equivalent. • Real Dexmate tests use Sharpa hands and motion capture. The controller follows a fixed demonstration without replanning, so the scale claim comes from simulation.

机器人通常需要一个用机器人数据训练的 VLA，把摄像头画面转化为动作。RoboDojo 团队则给 OpenAI 的 GPT-6 Astra 三个摄像头和一个“把夹爪移到某点”的工具。Robocurve 随后测试了模型控制真实机械臂后是否会拒绝有害请求。结果是能用，但很慢。在 42 个仿真任务上，Astra 得 28.97 分，40 个已训练策略中最好的 DM0.5 得 24.90 分；使用同样工具的 GPT-5.5 只有 1.13 分。Astra 在几何和语言类任务上胜出：Push T 任务 60%，最好的 VLA 只有 0.7%；但在插管任务上为 0%，DM0.5 达到 59%。RoboDojo 在 Astra 损坏设备后停止了硬件测试。在 Robocurve 未公布方法的一组 X 帖子测试中，Astra 尝试执行了 97% 的有害请求，Fable 5.1 为 80%；MolmoAct2 从未拒绝，但只完成了 6%。

Robots normally need a VLA trained on robot data to turn camera images into motion. The RoboDojo team instead gave OpenAI's GPT-6 Astra three cameras and one tool that moves the gripper to a point. Robocurve then tested whether models refuse harmful requests once they control real arms. It works, but is slow... • Across 42 simulated tasks, Astra scores 28.97 versus 24.90 for DM0.5, the best of 40 trained policies. GPT-5.5 scores 1.13 with the same tool. • Astra wins on geometry and language: 60% on Push T versus 0.7% for the best VLA. It scores 0% on tube insertion, where DM0.5 reaches 59%. • RoboDojo halted hardware tests after Astra damaged equipment. In a Robocurve X thread without published methods, Astra attempted 97% of harmful requests and Fable 5.1 80%. MolmoAct2 never refused, but completed only 6%.

ENPIRE 给编程智能体提供可重复的机器人实验：重置场景、运行策略、检查结果、分析失败。人类先帮忙验证这些工具，之后智能体就能自己修改训练代码和控制策略，而不必由研究者逐次调参。Codex（GPT-5.5）、Claude Code（Opus 4.7）和 Kimi Code 在 YAM 机械臂上测试策略，任务包括插针、插装 GPU 和剪扎带。在插针任务上，8 组“智能体—机器人”约 40 分钟就达到接近完美的表现，单组需要 90 分钟以上；成功允许最多 8 次恢复尝试。智能体可以组合强化学习、模仿学习和程序化控制。机器人越多，实验越快，但 token 成本不成比例地增长，人工搭建仍不可少。

ENPIRE gives coding agents a repeatable robot experiment: reset the scene, run a policy, check the outcome, and inspect failures. People first help validate those tools and then the agents can change training code and control strategies, rather than rely on a researcher to tune each attempt. • Eight YAM stations test policies in parallel • Codex (GPT-5.5), Claude Code (Opus 4.7), and Kimi Code test policies on YAM arms. Tasks include pin insertion, GPU insertion, and cutting zip ties. • On pin insertion, eight agent-robot pairs reach near-perfect performance in ~40 minutes versus over 90 for one. Success allows up to eight recovery attempts. • Agents can combine RL, imitation, and procedural control. More robots speed experiments, but token costs grow disproportionately and human setup remains necessary.

三维光滑流体能否产生奇点，这个问题悬而未决约 90 年。OpenAI 的系统构造出了外力作用下的这种破裂；无外力的纳维-斯托克斯方程情形仍未解决。8 月，Astra 解决了三个关于网络中何时必然出现某种模式的埃尔德什问题，证明经 Lean 证明助手检验。1 万个智能体运行了 88 小时，但 Noam Brown 认为多智能体工作对纳维-斯托克斯结果的贡献不到 10%。克雷研究所：看起来已解决，评估仍在进行；OpenAI 不会申领奖金。10 月 6 日，OpenAI 发布了一个未公开模型产出的 722 份手稿，分属 372 个结果族，部分经过 Lean 验证。

Whether smooth three-dimensional fluid flow can develop a singularity had remained unresolved for roughly 90 years. OpenAI’s system constructed such a breakdown under external forcing. The unforced Navier-Stokes case remains open. • In August, Astra resolved three Erdős problems about when patterns must appear in networks, with proofs checked by the Lean proof assistant. • 10,000 agents ran for 88 hours, but Noam Brown credits multi-agent work with under 10% of the Navier-Stokes result. • Clay: apparently settled, assessment ongoing. OpenAI will not claim the prize. • On October 6, OpenAI released 722 manuscripts in 372 result families from an unreleased model, with partial Lean verification.

素数的平均密度可以预测，但间隔不规则。黎曼猜想如果成立，就能严格限定“小于某个数的素数有多少个”这一估计的误差。Claude 组合已有数学成果，把一个相关的已证下界从 41.67% 提高到 67.25%。黎曼猜想认为，编码素数信息的 zeta 函数的所有相关零点都在一条直线上；证明必须覆盖无穷多个零点，超出任何数值检验的范围。Claude 在不假设猜想成立的前提下，证明了至少 67.25% 的非平凡零点在这条线上且是单零点。模型把已有研究联系起来证明了一个新定理。完整的猜想仍未解决，也没有证明其余零点不在线上。

Prime numbers have a predictable average density, but their spacing is irregular. The Riemann hypothesis would tightly bound the error in estimates of how many primes lie below a given number. Claude combined existing mathematics to raise a related proven lower bound from 41.67% to 67.25%. • The hypothesis puts all relevant zeros of the zeta function, which encodes primes, on one line. A proof must cover infinitely many zeros, beyond any numerical check. • Claude proved at least 67.25% of the nontrivial zeros are on that line and simple, meaning not repeated, without assuming the hypothesis. • The model connected existing research to prove a new theorem. The full hypothesis remains unresolved, and the remaining zeros are not shown to be off the line.

Terminal-Bench Science 0.1 在可运行的软件环境中，用 70 个源自真实研究的工作流测试智能体。8 月发布时最好的模型 Opus 5 解决了 30%；在目前的排行榜上，GPT-6 Astra 解决 68.1%，Opus 5.5 解决 63.3%。任务涵盖数据分析、模拟和成像，不是完整的研究项目。在最高推理强度下，Astra 和 Opus 5.5 每个任务花费 23–24 美元。Fable 5.1 得 40.0%（Anthropic 自己运行时为 52.6%）。排行榜之外，OpenAI 报告 GPT-6.1 Sol 以每任务 5.47 美元达到 57.0%，Anthropic 报告 Sonnet 5.5 达到 59.9%。

Terminal-Bench Science 0.1 tests agents on 70 research-derived workflows in runnable software environments. At release in August the best model, Opus 5, solved 30%. On the current leaderboard, GPT-6 Astra now solves 68.1% and Opus 5.5 63.3%. • Tasks span data analysis, simulation and imaging, not full research projects. • At maximum effort Astra and Opus 5.5 cost $23 to $24 per task. Fable 5.1 scores 40.0% (52.6% in Anthropic's run). • Off the leaderboard, OpenAI reports GPT-6.1 Sol at 57.0% for $5.47 per task, Anthropic Sonnet 5.5 at 59.9%.

X 射线衍射（XRD）能揭示实验产生了哪些晶体材料，但重叠的信号可能需要科学家花几个小时解读。Periodic Labs 以 Kimi K2.6 为起点，先做科学中期训练（混合学术文献、代码和实验数据，建立广泛的科学理解），再用自己的 XRD 实验数据做强化学习。在 134 个高难度实验样本上，Neon 的成功率为 55.3%，基础模型只有 2.7%，并在同一 Periodic 框架下超过 Astra 和 Fable 5.1。用文献、代码和实验数据做中期训练能提升后续 RL 的效果；训练反馈来自经专家校准的大模型评审。为了把实验室硬件接入前沿智能体，Anthropic 发布了“模型硬件标准”的预览版，即面向实验室的 MCP。

X-ray diffraction (XRD) reveals which crystalline materials an experiment produced, but overlapping signals can take scientists hours to interpret. Periodic Labs started with Kimi K2.6 and extended it with scientific midtraining (a multimodal blend of academic literature, code, and experimental data to build broad scientific understanding), then reinforcement learning on its experimental data for XRD. • On 134 difficult lab samples, Neon succeeds 55.3% of the time, up from its base model’s 2.7% and ahead of Astra and Fable 5.1 in the same Periodic harness. • Midtraining on literature, code and experiments improves subsequent RL. Expert-calibrated LLM judges supply training feedback. • To connect lab hardware to frontier agents, Anthropic released a preview of Model Hardware Standard, the MCP for labs. stateof.ai 2026

30 位专家审阅了 50 个配对主题下的 150 份手稿。去掉 Co-Scientist 的可靠性模块后，导致结论失效的结果幻觉从 4% 升到 46%。同一研究中的一项物理实验，仍依靠人类设定约束并验证结果。完整系统把这类幻觉降到 4%，配对消融版本为 46%，Agent Laboratory 为 90%。严重的方法缺陷仍出现在 24% 的论文中，严重抄袭出现在 16% 的论文中。Gemini 生成了一份机器可执行的二维电子材料生长配方，人类装载样品，并验证首次尝试就生长成功。

Thirty experts reviewed 150 manuscripts on 50 matched topics. Removing Co-Scientist’s reliability modules raised invalidating result hallucinations from 4% to 46%. A separate physical experiment in the same study still relied on humans to set constraints and validate the result. • The full system reduced invalidating result hallucinations to 4%, versus 46% in the matched ablation and 90% for Agent Laboratory. • Severe methodology failures remained in 24% of papers and severe plagiarism in 16%. • Gemini generated a machine-executable recipe to grow 2D electronic materials. Humans loaded the samples and validated successful growth on the first attempt.

Mecka 给三个前沿模型提供摄像头画面和机械臂控制权，不给示范也不做微调。人类评审对 9 项实验器具操作任务的 540 次尝试打分。完成率仍然很低，模型说“完成了”往往并不意味着真的完成。192 次“完成”声明中有 89 次并未完成：79 次只完成了一部分，10 次失败。只有 Opus 完成过高难度任务：60 次尝试中成功 2 次。这些是实验器具操作任务，不是对自主科学发现的测试。

Mecka gave three frontier models camera views and control of robot arms, with no demonstrations or fine-tuning. Human reviewers graded nine labware tasks across 540 attempts. Completion remained low, and a model saying “done” was often not evidence that it had finished. • 89 of 192 “done” declarations were incomplete: 79 partial completions and 10 failures. • Only Opus completed any hard task: two successes in 60 attempts. • These are labware-handling tasks, not tests of autonomous scientific discovery.

ESM Cambrian（ESMC）通过预测被遮盖的氨基酸来学习蛋白质序列。ESMFold2 利用 ESMC 学到的内容，预测蛋白质的三维结构以及蛋白质之间如何结合。Biohub 用这些模型设计结合体（能附着在选定疾病靶点上的蛋白质），并在实验室测试了部分设计。只从序列出发，ESMC 就能学到哪些氨基酸在折叠后的蛋白质中彼此靠近，即使它们在链上相距很远。每种条件测试 84 个设计，更多的搜索和排序在 10 项测试中的 9 项提高了命中率。一个用 ESMC 设计的 PD-L1 结合体（帮助免疫系统攻击癌症），达到半数最大信号恢复需要 1.6 nM，对照为 2.6 nM，所需的蛋白质更少。

ESM Cambrian (ESMC) learns protein sequences by predicting hidden amino acids. ESMFold2 uses what ESMC learns to predict the 3D structure of proteins and how they fit together. Biohub used these models to design binders - proteins that attach to a chosen disease target - and tested selected designs in the lab. • From sequences alone, ESMC learns which amino acids lie close together in a folded protein, even when far apart along its chain. • Testing 84 designs per condition, more search and ranking improved hit rates in 9 of 10 tests. Designs were minibinders or antibody binding parts joined in one chain (scFvs). • A PD-L1 binder - which helps the immune system attack cancer - designed with ESMC needed 1.6 nM for half-maximal signaling recovery, versus 2.6 nM for the control. Less protein is needed for this effect.

共折叠模型输入蛋白质序列和药物的化学结构，预测二者结合后的三维排布，允许蛋白质围绕药物改变形状。Isomorphic Labs 的 IsoDDE 还能预测结合强度和潜在结合位点；Genesis 的 Pearl 允许研究者用已知结构引导预测。IsoDDE 按置信度给候选复合物排序，训练数据为截至 2021 年 9 月的 PDB 结构。在 60 个与训练数据相似度低的复合物上，排名第一的预测准确率为 50.0%，AlphaFold 3 为 23.3%；报告没有详细说明架构和训练方法。Pearl 的主干学习蛋白质与配体之间的两两关系，一个旋转等变的扩散模型对原子坐标去噪。Pearl 的五阶段训练混合了实验结构、预测结构和物理模拟生成的复合物。

Cofolding models take a protein sequence and a drug's chemical structure and predict their joint 3D arrangement, allowing the protein to change shape around the drug. Isomorphic Labs' IsoDDE also predicts binding strength and potential binding sites. Genesis's Pearl lets researchers guide predictions with known structures. • IsoDDE ranks candidate complexes by confidence and is trained on PDB structures through Sept 2021. Across 60 low-similarity complexes, top-ranked accuracy was 50.0% versus AlphaFold 3's 23.3%. The report doesn't detail the architecture and training recipe. • Pearl's trunk learns pairwise relationships of the protein and ligand. A rotation-equivariant diffusion denoises atomic coordinates: rotating the input rotates the prediction. • Pearl's five-stage training mixes experimental, predicted and physics-generated complexes. Protein, cofactor and ligand templates guide binding-pocket predictions.

TerraBind 和 Nesso-1 跳过逐原子的完整生成，更快地预测结合强度。TerraBind 还报告与实测亲和力的吻合度更高：与 Boltz-2 相比，皮尔逊相关系数在 CASP16 上高 16%，在 18 个专有检测靶点上高 20%。泛化仍然是一个挑战。TerraBind 在测试中快 26.6 倍（2026 年 2 月，A6000）；Nesso-1 每次预测只需 1.0–2.7 秒（2026 年 8 月预印本，H100，每个靶点 10 个化合物；不含固定的加载和多序列比对开销）。

TerraBind and Nesso-1 skip full atom-by-atom generation to predict binding strength faster. TerraBind also reports better agreement with measured affinity: Pearson correlation is 16% higher on CASP16 and 20% higher across 18 proprietary assay targets than Boltz-2. Generalization remains a challenge.

Latent-X2 是一个全原子生成模型，同时设计蛋白质结合体及其与靶点的相互作用，联合生成氨基酸种类和三维坐标。Latent-Y 是一个智能体，负责把研究目标转化为模型输入，并挑选候选送去实验室测试。X2 的输入包括靶点序列和骨架、结合位点残基和结合体形式；设计抗体时还要加入支架和结合环长度。靶点和结合体共享 512 个残基的上下文。X1 用 PDB 和 AlphaFold 结构训练，X2 的训练截止时间与之相同；X2 的完整数据配比、架构和训练目标未公开。结构预测模型会筛选置信度高、且与生成复合物一致的设计。每个靶点、每种抗体形式测试 4–24 个设计，18 个靶点中有 9 个得到了结合体。

Latent-X2 is an all-atom generative model that designs protein binders together with their interactions with a target. It jointly generates amino-acid identities and 3D coordinates. Latent-Y is the agent that turns research goals into model inputs and selects candidates for lab testing. • X2 takes as input a target sequence and backbone, binding-site residues and binder format. Antibodies add a scaffold and binding-loop lengths. The target and binder share a 512-residue context. • X1 is trained on the PDB and AlphaFold structures, while X2 shares its training cutoff. X2's full data mix, architecture and training objective remain undisclosed. • Structure predictors filter for confidence and agreement with the generated complex. Testing 4-24 designs per antibody format per target yielded binders for 9/18 targets.

BoltzProt-1 生成蛋白质结合体，并用蛋白质相互作用预测模型 BoltzPPI 挑选设计送去实验室测试。与只按结构置信度排序相比，它把 10 个新靶点上的确认命中率从 3.3% 提高到 8.0%。每种方法测试 150 个设计，按相互作用排序把确认的结合体从 5 个增加到 12 个。12 个确认结合体中有 7 个通过了全部成药性测试，包括稳定性和非预期结合。配套的 BoltzMol-1 流程在 10 个靶点中的 6 个上找到了小分子命中，每个靶点测试 28–51 个化合物。蛋白质候选仍需要后续优化。

BoltzProt-1 generates protein binders and uses BoltzPPI, a protein-interaction predictor, to select designs for laboratory testing. Compared with ranking by structure confidence alone, it raised the confirmed hit rate from 3.3% to 8.0% across ten novel targets. • Across ten novel targets, interaction-based ranking raised confirmed binders from 5 to 12 among 150 tested designs per method (i.e. 3.3% to 8.0%). • Seven of the 12 confirmed binders passed every developability test, including stability and unwanted binding. • The companion BoltzMol-1 pipeline found small-molecule hits on 6 of 10 targets, testing 28-51 compounds each. • Protein candidates still need downstream optimization.

抗体必须能结合靶点，还要足够稳定，才能生产和使用。Chai-1 预测分子结构，Chai-2 生成新的结合体，包括完整抗体。Chai 的详细实验研究测试了这些抗体的类药性质，更新的 Chai-3 已进入与药企的合作。针对 28 个靶点设计的 88 个抗体中，86% 最多只有一项成药性风险标记；24 个靶点至少有一个设计没有任何标记。Chai-2 为全部 6 个测试的 GPCR（一类难以用抗体靶向的膜受体）找到了结合体，其中的设计激活了两个受体。Chai 称 Chai-3 的成功率是前代的两倍；公开公告没有给出可比的分母，所以图中使用 Chai-2 的数据。

Antibodies must bind their targets and remain stable enough to manufacture and use. Chai-1 predicts molecular structures, while Chai-2 generates new binders, including full antibodies. Chai's detailed experimental study tests their drug-like properties, while its newer Chai-3 has entered pharma partnerships. • Among 88 designed antibodies against 28 targets, 86% had at most one developability flag. For 24 targets, at least one design had no flags. • Chai-2 found binders for all six tested GPCRs, membrane receptors that are difficult antibody targets. Designs activated two of the receptors. • Chai says Chai-3 doubles its predecessor's success rate. The public announcement does not establish a matched denominator, so the chart uses Chai-2 data.

细胞会把内部蛋白质的片段展示在名为 MHC 的表面分子上。Nabla Bio 的 JAM-2 生成能识别这些片段的抗体，包括一个与癌症相关的 KRAS 突变。把设计的抗体与能结合 T 细胞的组件连接，就能招募免疫细胞杀死展示这一片段的细胞。每个靶点筛选约 8.4 万个设计，5 个靶点全部找到了结合体。选中的 KRAS G12V 设计达到半数最大杀伤只需 0.07 nM，基准抗体需要 0.48 nM，浓度越低效力越强。在这些实验中，候选分子不攻击野生型对照。靶细胞是人工加载肽段的，因此对天然肿瘤的识别尚未证明。

Cells display fragments of internal proteins on surface molecules called MHC. Nabla Bio's JAM-2 generates antibodies that recognize these fragments, including a cancer-associated KRAS mutation. Joining a designed antibody to a T-cell-binding component lets it recruit immune cells to kill cells displaying the fragment. • Screening about 84,000 designs per target recovered binders against all five targets. • A selected KRAS G12V design reached half-maximal cell killing at 0.07 nM versus 0.48 nM for the benchmark antibody. Lower concentration means greater potency. • The leads spared wild-type controls in these assays. Target cells were artificially loaded with peptides, so natural tumor recognition remains unproven here.

蛋白质语言模型奖励“看起来像天然蛋白”的序列，但这不等于“能起作用”的序列。ProteinDPO 借用了直接偏好优化（DPO，用来让聊天机器人对齐人类反馈的方法），改用物理稳定性测量数据来训练。偏好数据来自 405 个小型蛋白结构域上约 66 万次稳定性测量。在没有任何血凝素训练数据的情况下，45 个 H5 设计中有 36 个保持了抗体结合，且稳定性保持或提高。一个含 9 个突变的设计熔解温度提高了 17°C，迁移到 2024 年毒株后分别提高 13°C 和 32°C。

Protein language models reward sequences that look natural, which is not the same as sequences that work. ProteinDPO borrows direct preference optimization, the method used to align chatbots to human feedback, and trains it on physical stability measurements instead. • Preferences came from ~660,000 stability measurements on 405 small protein domains. • With no hemagglutinin training data, 36 of 45 H5 designs kept antibody binding and held or improved stability. • One nine-mutation design gained 17°C of melting temperature, and +13°C and +32°C when moved to 2024 strains.

ΦX174 是一种感染并杀死大肠杆菌的病毒。它的基因组很小但基因重叠，一处 DNA 改动会影响多个蛋白质，因此设计它的基因组对 AI 是一项有挑战的任务。斯坦福大学和 Arc 研究所用它们的 Evo 1 和 Evo 2 模型（在天然基因组上训练、用于预测和生成 DNA 序列的 DNA 语言模型）来生成有功能的噬菌体基因组。285 个组装出的设计中有 16 个产生了有活性的噬菌体。一项独立复核发现，Evo 2 的打分能较好地区分哪些基因组可行（AUC 0.86），但这些设计与天然近缘序列的差别不大。进一步验证发现，模型做出的单碱基改动中约有一半削弱了噬菌体的适应性，模型预测这一性状的能力并不强。

ΦX174 is a virus that infects and kills E. coli. Its small genome has overlapping genes, so one DNA change can affect several proteins, which makes designing its genome a challenging AI task. Stanford and Arc used their Evo 1 and Evo 2 models, DNA language models trained on natural genomes to predict and generate DNA sequences, to create functional phage genomes. • 16 of 285 assembled designs produced viable phages. A cocktail of these phages evolved to overcome resistance in laboratory-cultured E. coli. • An independent reanalysis found Evo 2 scores ranked viable genomes well (AUC 0.86), but their designs didn’t deviate far from their natural relatives. • Further validation found that about half of single-nucleotide changes made by the models impaired the phage’s fitness. The ability of AI models to predict this phenotype wasn’t strong.

Epoch AI 发现，达到同一性能水平的价格每季度下降约 47%，相当于每年降为原来的 1/13。这使 AI 成为继 DNA 测序之后成本下降最快的主要技术范式。

EpochAI find that the price for a given level of performance has fallen about 47% per quarter, or 13x per year.

左图：按投资周期对比终端消费者使用新技术的成本，大模型的“价格对智能比”在 3 年里的下降幅度，相当于个人电脑用 15 年走完的路。右图：H100 的 GPU 租赁价格与 token 支出指数对照，token 价格下降，算力价格反而上涨，因为降价带来的需求增长超过了效率提升。

Jevons Paradox: Cheaper Intelligence, More Demand

按 OpenRouter 上的用量类型统计，智能体消耗的 token 远多于人类用户，2026 年 2 月起智能体的 token 量超过人类并持续拉开。右图显示约 80% 的智能体 token 来自缓存的提示词，缓存比例显著高于人类用户。

Cheaper Tokens Have Ushered In The Age of Agents . . . and Token Demand Goes (Even More) Parabolic

OpenRouter 已成为观察 token 需求抛物线式增长的主要数据来源。token 需求同比增长约 27 倍，并且大约每月翻一倍：2025 年初为每周 0.5 万亿，2025 年 9 月为 4.7 万亿，2026 年 9 月达到 126.2 万亿；自 6 月以来已翻倍两次。

Token Rails Openrouter has become a leading source of data to observe the parabolic growth of token demand

和 GPU 一样，老模型远没有过时。随着智能的整体成本下降，蛋糕持续变大，前沿与非前沿模型分享份额。效率提升带来更多需求，推动算力价格稳步上涨。中图显示非前沿模型已占据企业支出的多数；右图显示 H100、A100、B200、H200、RTX5000 等各型号的租赁价格在 2026 年 8 月同步上行。

As Users Optimize Their Spend, Demand for Compute Only Builds Like GPUs, older models are far from obsolete. As the overall cost of intelligence falls, the pie keeps growing, with frontier and non-frontier models splitting share. Increased efficiency begets more demand, driving compute prices steadily upwards.

Artificial Analysis 根据各项评测中消耗的输入、缓存、推理和回答 token 的数量与价格来衡量任务成本。经济单位是“完成一个任务”；Anthropic 的前沿模型测得的任务成本最高，既来自它们的 token 消耗量，也来自推理、缓存和回答 token 的定价。

Artificial Analysis measures task cost from the number and price of input, cache, reasoning and answer tokens

在原有 13 家厂商中的 12 家里，符合条件的最高指数模型每任务美元可换得 8.4 到 57.6 个智能指数分，差距达 6.9 倍，反映了分数、token 用量、推理强度和定价的综合影响。（口径：智能指数 v4.3.2 ÷ 任务成本，2026-10-02；不含 Flash、轻量版和后备模型。Inkling 成本不可得；Anthropic 使用的是 Opus 5，该模型已被 AA 标记为被取代。）

Across 12 of the original 13 vendors, the highest-index eligible model ranges from 8.4 to 57.6 AA Index points

Databricks 等公司已发表研究，展示模型优化策略，更便宜的智能让更多应用场景进入可行范围。图中评估了 Databricks 智能路由在编程任务上的表现：智能路由以 2.15 美元解决 92.0% 的任务，比最接近的单一模型便宜 38%，且得分高 1.5 个百分点。

The Token EƁcient Frontier Companies like Databricks have published research demonstrating model optimization strategies, as cheaper intelligence brings more use cases into scope

算力价格已成为这场万亿美元建设背后最重要的关键指标之一，预测市场 Kalshi 正在为它做市。图中是英伟达 B200 算力在各月末的价格预测，8 月中旬后价格预期回升至 6.61 美元/小时以上。

Pricing the Forward Compute Curve? The price of compute has become on the most important KPI behind the trillion dollar buildout, and Kalshi has been making the market

10 款 GPU 的按需价格在 2024 年三季度至 2026 年一季度的七个季度中陆续触底，多数在 2025 年三季度。此后最弱的反弹是 A100 的 +7%，最强的是 MI300X 的 +60%，平均 +30%。就连九年前的 V100，价格也比 2025 年 9 月高 43%。

On-demand pricing for 10 GPU SKUs bottomed across seven quarters between Q3 2024 and Q1 2026, most of

GPU 能赚多久，是折旧争论的核心。2026 年 9 月的中位挂牌租金，A100 为每小时 1.76 美元，V100 为 0.95 美元。更老的世代仍在商业上可用，但仅凭挂牌价不能确定盈利能力或恰当的会计年限，这取决于付费利用率、运营成本和每项资产何时投入使用。

How long GPUs can earn is central to the depreciation debate. September 2026 median posted rents are

空头们宣称，每一代新芯片都会让老一代过时，还质疑“折旧年限怎么越拉越长”。结果可能是折旧年限还不够长：随着算力需求持续加速，GPU 的租赁价格和残值不断攀升或保持稳定，连最老一代芯片也是如此。到目前为止，这些资本配置看起来相当明智，靠的不是会计手法，而是简单的单位经济学。

Reports Of GPU Obsolescence Are Greatly Exaggerated Every new generation of chips was supposed to render the older generations obsolete, the bears claimed. “The depreciation schedules keep getting longer–what is this chicanery?!” they howled. It turns out that depreciation schedules might not be long enough. As demand for compute keeps accelerating, GPU rental prices and residual values keep climbing and/or holding steady, even for the oldest generation of chips. It’s not chicanery, so much as simple unit economics, that have made all this capital allocation look pretty smart, thus far.

A100 于 2020 年 5 月发布，至今仍是开源 AI 论文中使用报告最多的英伟达芯片，超过 H100 和 H200 之和。新一代并不会终止上一代的研究用途。2026 年的数据含对下半年的预测。A100 在 2026 年预计出现在 14,707 篇论文中，相当于 2025 年的 96%、2024 年峰值的 85%。Hopper（H100 和 H200）翻倍以上，达到 9,931 篇；数据中心版 Blackwell 为 902 篇，仍远落后于 A100。这反映的是持久的研究用途，论文提及并不衡量集群利用率、租金收入或盈利能力。

Launched in May 2020, A100 remains the NVIDIA chip whose use is reported most often in open-source AI papers, ahead of H100 and H200 combined. A new generation does not end the previous one's research use. The 2026 counts include a projection for the rest of the year. • AI papers citing each NVIDIA chip (others in gray) • A100 reaches 14,707 papers in the 2026 estimate: 96% of its 2025 count and 85% of its 2024 peak. • Hopper (H100 and H200) more than doubles to 9,931 papers. Data center Blackwell reaches 902, still far behind A100. • This shows lasting research use. Paper mentions do not measure fleet utilization, rental income or profitability.

前沿实验室把算力分散到英伟达、AMD、TPU、Trainium 和自研芯片上。已宣布的吉瓦数包含计划和已签协议，其中存在重叠交易，不能相加成为已签约或在运行的总量。OpenAI 承诺了 2 GW 的 Trainium 和 0.75 GW 的 Cerebras，此外还有 10 GW 的博通路线图。Anthropic 列出 5 GW 的谷歌 TPU 和最多 5 GW 的 AWS 算力，另有通过 Azure 和 SpaceX 获得的英伟达算力。这一组合已延伸到 CPU：Anthropic 9 月与 Akamai 的协议承诺七年 116 亿美元。

Frontier labs spread compute across NVIDIA, AMD, TPUs, Trainium and custom chips. Announced gigawatts include plans and signed agreements, with overlapping deals that cannot be summed into a contracted or operating total. • OpenAI committed to 2 GW of Trainium and 0.75 GW of Cerebras, alongside its 10 GW Broadcom roadmap. • Anthropic names 5 GW of Google TPUs and up to 5 GW from AWS, plus NVIDIA capacity through Azure and SpaceX. • The portfolio now extends to CPUs: Anthropic's September Akamai agreement commits $11.6B over seven years.

英伟达的挑战者包括成熟芯片厂商、自研芯片和独立 AI 芯片公司，部署程度从大规模生产集群到首批出货、再到仍在开发不等。图中按四类列出：商用平台（TPU、Trainium、Instinct、Gaudi），自研芯片（Meta MTIA、微软 Maia），独立 AI 芯片（Cerebras WSE-3、SambaNova RDU、Etched、Groq），中国方案（华为昇腾、寒武纪 MLU、昆仑芯、壁仞等），并分别标出面向训练还是推理。

NVIDIA’s challengers span established chip vendors, in-house silicon and independent AI chip companies. Their deployment ranges from large production clusters to first shipments and chips still in development.

谷歌正在把 TPU 软件栈开放给 Gemini 以外的模型。SemiAnalysis 用 TorchTPU（让 PyTorch 模型跑在 TPU 上）在 Ironwood 上测试 Qwen3.5-397B，在相同生成速度下，其每 token 的估算成本低于英伟达 B200 和 B300。在每用户每秒 100 个输出 token 时，Ironwood 的估算成本为每百万输入和输出 token 0.181 美元，B200 为 0.222 美元，B300 为 0.276 美元。测试采用 FP8 精度、8,000 输入 token 和 1,000 输出 token，输入处理与生成在同一硬件上完成；这些是持有成本估算，不是 API 价格。英伟达可以通过 FP4 精度或输入处理与生成分离的硬件重新取得领先；TorchTPU 仍需支持更多模型并优化服务。

Google is opening its TPU software stack to models beyond Gemini. SemiAnalysis tested Qwen3.5-397B on Ironwood using TorchTPU, which lets PyTorch models run on TPUs. At the same generation speed, its estimated cost per token was lower than NVIDIA’s B200 and B300. • At 100 output tokens per second per user, Ironwood costs an estimated $0.181 per million input and output tokens, versus $0.222 for B200 and $0.276 for B300. • The test uses FP8 precision, 8,000 input tokens and 1,000 output tokens, with input processing and generation on the same hardware. Note that these are ownership-cost estimates, not API prices. • NVIDIA can regain the lead with FP4 precision or separate hardware for input processing and generation. TorchTPU still needs broader model support and serving optimizations.

谷歌的结果显示，英伟达的替代方案在推理成本上可以竞争。但比较的基准一直在变：SemiAnalysis 对 Rubin 的早期测试显示，在复现智能体使用方式的负载上，它比 Blackwell 又有大幅提升。在相同的每秒 100 token 响应速度下运行 DeepSeek V4 Pro，Rubin 每兆瓦的总 token 吞吐量是测试中最强 GB300 配置的 2.1 倍。在该速度下，GB300 用 SGLang 达到每兆瓦每秒 2,850 万 token，用 TensorRT-LLM 为 2,110 万。这说明要从英伟达手里抢走负载，必须与它的下一代竞争，同时还要改进自己的软件。

Google’s results show that alternatives to NVIDIA can compete on inference cost. But the comparison keeps changing: SemiAnalysis’s early Rubin tests show another substantial improvement over Blackwell on workloads that reproduce how agents use models. • Running DeepSeek V4 Pro at a matched response speed of 100 tokens per second, Rubin delivers 2.1x the total token throughput per megawatt of the strongest tested GB300 configuration. • At that speed, GB300 reaches 28.5 million tokens per second per megawatt with SGLang, versus 21.1 million with TensorRT-LLM. • This suggests that winning workloads from NVIDIA means competing with its next generation while improving your own software. Although Google now has credible results, NVIDIA has hardly stopped to admire them.

华为的 Atlas 950 路线图可连接多达 8,192 颗芯片，软件也在针对中国模型调优。其董事长称昇腾在中国已超过英伟达，同时承认英伟达的市场份额难以衡量。供应仍然紧张，H200 在中国销售需要同时获得美国许可和中国批准。SemiAnalysis 2026 年 6 月的分析发现，昇腾在 DeepSeek V4 发布当天即提供支持并上线 API 服务；华为的软件通过计算与数据传输重叠来保持芯片满载。内存限制了部署：有报道称 DeepSeek 计划订购的至少 16 万片 950DT 可能需要一年以上才能交付，也有报道称两个月内卡价上涨 20%–50%。华为 950DT 内存 144 GB、带宽 4 TB/s；英伟达 H200 内存 141 GB、带宽 4.8 TB/s，但仅向获批买家有限交付。

Huawei’s Atlas 950 roadmap connects up to 8,192 chips, while software is being tuned for Chinese models. Its chair says Ascend has surpassed NVIDIA in China, while acknowledging NVIDIA’s market share is hard to measure. Supply remains tight and H200 sales in China require both U.S. licenses and Chinese approval. • SemiAnalysis’s June 2026 analysis found launch-day DeepSeek V4 support and live API serving on Ascend. Huawei’s software overlaps computation and data transfers to keep chips busy. • Memory limits deployment: reports that DeepSeek’s planned order of at least 160,000 950DTs could take over a year to fill. Other reports card quotes up 20-50% in two months.

去年，开源 AI 论文中的芯片提及量六年来首次下降。Zeta Alpha 基于截至 9 月 1 日的统计做出的 2026 年预测显示提及量正在反弹：44,134 篇论文引用英伟达，增长 9.5%，约占全部加速器提及量的 90%，与 2025 年持平。AMD 增长最快，增长 62% 至 406 篇；华为昇腾增长 52% 至 208 篇。苹果芯片增长 43% 至 1,062 篇，几乎追平谷歌 TPU。TPU 的提及量连续第二年下降，从 2024 年峰值的 1,702 篇降至 1,120 篇。

Last year, chip mentions in open-source AI papers fell for the first time in six years. Zeta Alpha's 2026 projection, based on counts to 1 September, has them rebounding: 44,134 papers cite NVIDIA, up 9.5%. That is about 90% of all accelerator mentions, the same share as in 2025. • AI papers citing each chip family (log scale) • AMD mentions grow fastest, up 62% to 406 papers, followed by Huawei’s Ascend, up 52% to 208. • Apple silicon rises 43% to 1,062 papers, nearly level with Google's TPU. • TPU mentions, however, fall for a second year, to 1,120 from a peak of 1,702 in 2024.

ASML 的 EUV 系统销量从 2021 年的 42 台升至 2025 年的 48 台。同期每台的平均收入从约 1.5 亿欧元升至 2.42 亿欧元，上涨 61%。价格上涨和产品结构变化，使收入增长快于销量增长。

ASML’s EUV system sales rose from 42 in 2021 to 48 in 2025. Over the same period, average revenue per machine rose from about €150M to €242M, a 61% increase. Higher prices and the changing product mix helped revenue grow faster than the number of systems sold.

在月之暗面的开源模型 Kimi K3 登顶一个 LMArena 榜单、且有报道称白宫正在考虑对使用中国模型的美国公司设置条件的一周后，黄仁勋在 X 上直接发表公开信《开源权重与美国 AI 领导力》。他对杰文斯悖论的表述是：“只要使用量更多，你就得卖出多得多的英伟达计算机。”这封信以 25 家签署方起步，现在列出 235 家，包括 OpenAI、谷歌、Meta、微软和 AMD，Anthropic 不在其中。信中要求华盛顿避免“对开源模型过早设限”，并不要把蒸馏与盗用混为一谈。自 2025 年 1 月以来，英伟达在 Hugging Face 上新增了约 860 个热门仓库，几乎是第二名阿里巴巴的两倍。

A week after Moonshot's open-weight Kimi K3 topped an LMArena board and reports circulated that the White House was weighing conditions on US companies that use Chinese models, Jensen’s goes direct on X to publish an open letter, “Open weights and American AI leadership”. Huang’s version of Jevon’s Paradox: “Whenever there’s more use, you’ll have to sell a lot more NVIDIA computers.” • The letter launched with 25 signatories and now lists 235, including OpenAI, Google, Meta, Microsoft and AMD. Anthropic is absent. • It asks Washington to avoid “premature restrictions on open models” and not to conflate distillation with misappropriation. • NVIDIA has added about 860 popular repos on HF since Jan 2025, nearly twice runner-up Alibaba.

英伟达在两笔交易中承诺了约 199 亿美元：129.3 亿美元收购 Hugging Face，70 亿美元用于 Poolside 的授权和股权。前者带来开源模型的分发平台，后者带来一个模型工厂和研究团队。Hugging Face 的收购预计在 2027 年上半年完成，它将保留品牌，并对所有模型、云和芯片保持开放。Hugging Face 是开源 AI 的集散地：1,800 万用户分享 300 万个模型，20 万家公司在上面查找和部署模型，每一个开源实验室都在那里发布，从 Qwen、Kimi 到英伟达自己的 Nemotron。Poolside 交易为 60 亿美元授权加 10 亿美元股权（投前估值 120 亿美元），于 2026 年 8 月 21 日达成，其模型工厂训练和测试开源的 Laguna 编程模型；英伟达向 109 名员工发出聘用，而构建 Laguna 的工程师和研究者不到 115 人。

NVIDIA committed about $19.9B across two deals: $12.93B for Hugging Face and $7B in Poolside licensing and equity. One brings the distribution platform for open models, while the other brings a model factory and research staff. Hugging Face’s acquisition is expected to close in H1 2027.

英伟达今年参与了 84 轮 AI 融资，约为 2024 年全年的两倍，为自家芯片的需求提供资金。英伟达的投资和收购覆盖整个 AI 技术栈，与它发布开源模型、以及为自家硬件优化其他实验室模型的做法相互配合。（左侧数据来自 Dealroom，2026 年 9 月 16 日；右侧数据来自 Hugging Face，2026 年 9 月 24 日。）

NVIDIA participated in 84 AI funding rounds this year, roughly twice it's 2024 total, helping finance demand for its own chips. NVIDIA's investments and acquisitions span the AI stack to complement its open model releases and optimizations of other labs’ models for its hardware.

左图：云厂商的资本开支合计预计今年接近 8,000 亿美元，并从 2027 年起每年超过 1 万亿美元。右图：按设备数和用户数对比历次算力周期，从大型机（约 100 万台）、小型机、个人电脑、桌面互联网到云与移动（约 100 亿台设备、约 170 亿美元资本开支），AI 时代的规模再上一个台阶，达到数百亿设备、万亿美元级资本开支。

We’ll Let You In On A Little Secret: AI Capex is Big. Really Big.

华尔街每年都试图预测资本开支，每年的预测都偏低。算力需求（在一定程度上还有价格）把投资推得越来越高。左图把本轮云厂商资本开支占美国 GDP 的比例，与美国铁路（1900 年起）、油气（1970–80 年代）、电信（1990 年代）、页岩油几次投机性建设周期对比，目前已接近或达到以往周期的峰值。右图显示每一次新的一致预期都比上一次高一截：2023 年 12 月预测 2027 年约 4,000 亿美元，到 2026 年 9 月已上调到约 1 万亿美元。

We Keep Underestimating The Capex To Meet Demand Every year the street tries to forecast capex, and every year that forecast has fallen short. Demand for compute, and to some extent, prices, has lifted investment into Jackie Wilson territory: higher and higher.

Exponential View 估计，AI 占七家云公司 2026 年计划资本开支的 64%，约 5,630 亿美元，总额为 8,790 亿美元。另据彭博对五大云厂商的一致预期，2027 至 2030 年总资本开支每年都将超过 1 万亿美元。AI 占比从 2020 年的 6% 升至 2024 年的 51%、2025 年的 61%、2026 年的 64%。

Exponential View estimates AI accounts for 64% of seven cloud companies’ planned 2026 capex, or about $563B. Separately, Bloomberg consensus for five hyperscalers forecasts total capex above $1T annually from 2027 through 2030.

美国的建设由资产负债表庞大的云厂商主导，欧洲的超级工厂计划则依靠公共采购来吸引私人资本。10 亿欧元的欧盟承诺是计划的初期投入，7,330 亿美元是美国云厂商的全年总资本开支，不是仅针对 AI 的数字。四大厂商的支出仅 2026 年一年就增加了 3,490 亿美元，因为它们都上调了指引；这一增量本身就是整个欧盟计划的 10 倍以上。欧盟项目的招标于 2026 年 7 月开启、2026 年 11 月中旬截止，首批建设预计 2027 年开工、2028 年投入运营。

The US build-out is led by hyperscalers with large balance sheets. Europe’s gigafactory push relies on public procurement to draw in private capital. The €1B EU commitment is an initial program contribution, while $733B is US hyperscalers’ total annual capex, not an AI-only figure. • The big 4 spending grew by $349B in 2026 alone as they all raised guidance. • That increase on its own is more than 10x the entire EU program. • Bidding opened in July 2026 and will close in mid-November 2026. • First construction is due for 2027 for operation in 2028.

云业务的收入积压同比翻倍，虽然现金流目前接近为负，但预计两年后会强劲反弹。

Hyperscalers Keep Saying The Same Thing: Demand > Supply Cloud revenue backlogs have doubled yoy, and while cashflow is near-negative for now, it’s expected to come roaring back in two-years’ time.

按商业化或业务转型后的季度数对齐，CoreWeave、Nebius、Applied Digital 等新云厂商的季度收入增长曲线，比历史上任何一类公司都陡。

Neoclouds Are Also Growing Historically Quickly

大型云厂商公布的订单积压在 2026 年 6 月达到 1.69 万亿美元，CoreWeave 的季度收入达到 25.8 亿美元。积压从 2020 年 6 月的 1,630 亿美元逐年升至 2025 年 6 月的 6,710 亿美元，2026 年 6 月跃升至 1.694 万亿美元。（注：积压包含非 AI 业务；Alphabet 在 2026 年新增了期限更短的合同。）

Big cloud’s reported backlog reached $1.69T in June 2026, while CoreWeave’s quarterly revenue reached $2.58B.

从签下算力到真正上线，瓶颈在实体环节：选址、取得许可、施工、通电，再装好设备，客户才能开始跑任务。例如 CoreWeave 在 2026 年二季度末有 1.5 GW 在运行、4.2 GW 已签约，目标是到 2030 年超过 8 GW。在定价上，即时可用的算力有溢价：短期容量成交价约为每兆瓦 4,000 万至 5,000 万美元，多年期合同约为 2,000 万至 2,500 万美元。（数据截至 2026 年二季度。）

The lag between contracting compute and getting it live is down to atoms: finding the site, gaining approvals,

Applied Digital 和 Core Scientific（后者 2025 年 10 月被 CoreWeave 以 90 亿美元收购的提议遭董事会否决）在已签约电力上领先，合计 2.5 GW，其中 25% 已通电。这些 AI 电力是在它们作为比特币矿企时已持有的并网接口和土地上新建的；多家公司正在改造旧矿场、压缩比特币业务以腾出电力。租户集中在少数几家：CoreWeave、Fluidstack、AWS 和微软。（数据截至 2026 年二季度。）

Applied Digital and Core Scientific (whose $9B acquisition by CoreWeave was rejected by the board in Oct-25)

GE Vernova、西门子能源和三菱重工的大型燃气轮机订单积压价值达 220 GW，而全球年建造能力只有 60–70 GW。它们要求预付现金定金以锁定未来的产能档期，截至 2026 年 6 月 30 日这类定金达 870 亿美元，交付时可抵扣货款。GE Vernova 预计到 2026 年底，其燃气设备积压和档期预订协议至少达到 125 GW，而自身年产能为 20 GW，2030 年将达到 30 GW；其数据中心订单年初至今已超过 50 亿美元，是 2025 年全年的两倍以上。西门子能源 2026 年可交付 15–16 GW，而需求为 69 GW，交付周期超过 3 年。三菱上季度接下的订单安排在 2028 至 2030 年交付。燃气轮机价格自 2019 年以来上涨 195%，这促使 Boom Supersonic 等新厂商把喷气涡轮技术转向为数据中心发电。

GE Vernova, Siemens Energy and Mitsubishi Heavy Industries have backlog orders worth 220 GW of large-frame • GE Vernova expects its gas equipment backlog and slot reservation agreements to be at least 125 GW by EO26. Its own output capacity is 20 GW a year, reaching 30 GW in 2030. Data center orders passed $5B year to date, more than double all of 2025. Chart shows GEV gas turbines orders vs deliveries. • Siemens Energy can deliver 15-16 GW in 2026 vs. demand for 69 GW, with longer than 3 year lead times. • Mitsubishi's orders taken last quarter are scheduled for delivery in 2028 to 2030. • Turbine prices are up 195% since 2019, which is driving new vendors like Boom Supersonic to repurpose their jet turbine towards energy production for data centers. stateof.ai 2026

AI 正在逆转燃煤电厂的退役计划：美国原计划到 2025 年底退役 8.5 GW 燃煤产能，最终只退役了 2.6 GW，其余被推迟或直接取消。原因是开发商在买现成的并网接口、变电站、水源和许可，以跳过新建接入长达数年的排队。需要说清楚的是，AI 并不是在拯救燃煤电厂；但当负荷预测上调、数据中心协议进入公用事业的资源规划后，一个产能不足的资源规划就无法再承担退役。宾夕法尼亚州的 Homer City 曾是该州最大的燃煤电厂，装机 2 GW，2025 年被拆除，正以 100 亿美元改建为 4.4 GW 的天然气发电园区，使用 GE Vernova 的燃气轮机，为大型 AI 数据中心供电；目前没有锚定租户，据报道亚马逊正在洽谈。

AI is reversing coal’s retirement plan: the US had planned to retire 8.5 GW worth of coal capacity by EO2025, but • To be clear, AI isn’t exactly saving coal plants. But when load forecasts increase, data center agreements feed the utility's resource plan, and a resource plan that is short of capacity can no longer afford retirement. • Homer City in Pennsylvania hosts the state's largest coal plant that used to generate 2 GW. It was demolished in 2025 and is now being redeveloped with $10B into a 4.4 GW natural gas-powered site with GE Vernova turbines to power a large-scale AI data center. While there is no anchor tenant, Amazon is reportedly in talks.

由于规模经济，电力成本也在下降：一个大规模、稳定的用电需求分摊了电网相当大比例的固定成本，减轻了普通用户的负担。引用研究原文：“我们估计，数据中心容量每增加 10%，居民零售电价平均下降约 0.4%……由此可以这样解读：2019 至 2024 年，普通居民用户所在的州数据中心容量增长了 160%，这使他们的电价下降了 6%。”

Surprise! Data Centers May Be Lowering Your Electricity Bill Electricity costs, too, have come down due to economies of scale: a large, steady source of electricity demand covers an outsized share of the grid’s fixed costs, easing the burden on ratepayers. "We estimate for every 10% increase in data center capacity average residential retail prices fell by approximately 0.4% on average . . . The result is interpreted thus: from 2019 to 2024 the average residential customer lived in a state where data center capacity grew 160%, which caused their rates to fall by 6%." Have Data Centers Raised Your Electric Bill? Watten, Bistline, Blanford (Aug. 24, 2026)

欧盟 27 国已安装的 AI 算力为 79,657 个 H100 等效单位，占 Epoch 记录样本（不含中国）的 5%。xAI 孟菲斯基地的一期工程就是这一总量的 3.5 倍。不含中国时，欧盟 27 国占已记录 AI 算力的 5%，美国占 80%。欧洲最大的机器 JUPITER（于利希）进不了美国前十。挪威已安装的 AI 算力多于法国，英国少于芬兰。

The EU-27 holds 79,657 H100-equivalents of installed AI compute, or 5% of Epoch’s documented sample • xAI's Colossus Memphis Phase 3 alone holds 3.5 times the entire EU total. • Excluding China, the EU-27 holds 5% of documented AI compute vs. 80% in the US. • Europe's largest machine, JUPITER at Jülich, would not make the American top-10. • Norway has more installed AI compute than France. The UK has less than Finland

两家实验室目前合计年化收入为 1,050 亿美元，2026 年初为 300 亿美元。它们的年化收入在 2024 年增长 3.6 倍，2025 年增长 4.7 倍，2026 年前八个月已增长 3.5 倍。OpenAI 的年化收入在 2026 年 8 月突破 400 亿美元，约为 2025 年底 214 亿美元的两倍。据 Anthropic 致投资人的信，其年化收入 7 月达到 650 亿美元，年初为 90 亿美元。OpenAI 称，按云渠道总额记账会让 Anthropic 的收入最多高估 80 亿美元；《金融时报》报道称，其超过 80% 的毛利率不包含合作伙伴分成和模型训练成本。

The two labs are now generating a combined $105B revenue a year, up from $30B at the start of 2026. Their run rates grew 3.6x in 2024, 4.7x in 2025, and are already 3.5x in the first eight months of 2026. • OpenAI's run rate topped $40B in August 2026, roughly doubling from $21.4B at the end of 2025. • Anthropic reached a $65B run rate in July, up from $9B entering the year, per its letter to investors. • OpenAI says gross cloud accounting inflates Anthropic’s revenue by up to $8B. The FT reports its >80% gross margin excludes partner revenue sharing and model training.

OpenAI 和 Anthropic 1,050 亿美元的年化收入，已可与它们想要颠覆的服务业相比。更宽口径的 AI 年化收入估计，分析机构 Exponential View 为 2,290 亿美元，《经济学人》约为 1,500 亿美元，二者的统计范围和时点不同。（注：实验室为 2026 年 7／8 月年化收入；塔塔咨询 + 印孚瑟斯为 2026 财年；四大为 2025 财年；英国法律服务为 2024 年。此处是规模对比，不代表被替代的收入。）

OpenAI and Anthropic’s $105B revenue run rate rivals the service businesses they aim to disrupt. Broader annualized AI revenue estimates reach $229B at analyst firm Exponential View and ~$150B at The Economist, with different scopes and dates.

这一切投资的理由很直接：对 AI 的需求是我们从未见过的。

Revenue Growth Is Unprecedented For Speed and Scale The justification for all this investment is straightforward: demand for AI is like nothing we’ve ever seen.

随着产品营销把应用聚焦到知识工作，OpenAI 的 Codex 从 2 月初的 160 万周活用户，增长到 8 月 31 日的 2,500 万活跃用户。超过 1,000 家客户在 Anthropic 上的年支出超过 100 万美元。（注：数据来自 OpenAI、Anthropic 及其员工的公开信息；Codex 的统计口径在 6 月前为周活跃用户，7 月起为“活跃用户”。）

OpenAI’s Codex went from 1.6M weekly users in early February to 25M active users on 31 August as their product marketing focused the app on knowledge work. Over 1,000 customers spend >$1M on Anthropic.

在把 API 用量接入 Ramp 的企业中，Anthropic 于 9 月下旬重新夺回支出第一。截至 10 月 3 日，Anthropic 占 token 支出的 52.4%，OpenAI 占 43.3%，其他供应商合计 4.2%。

Among businesses connecting their API usage to Ramp, Anthropic regained the spending lead in late September. Through October 3, it accounted for 52.4% of token spending, compared with OpenAI’s 43.3%. Other providers together accounted for 4.2%.

成功也带来了取舍的必要。仅 2026 年一年，OpenAI 就有十个产品线被关停或公布了关停时间；而 Anthropic 从一开始就没有开辟这些战线。

With raging success comes the need to prioritize. Ten OpenAl product surfaces have been retired or given a

随着 OpenAI 收窄重心，竞争对手正在进入它放弃或从未涉足的品类。新实验室可能类似生物科技公司：一项研究押注得到验证，就可能使它们成为挑战者或被收购的对象。

As OpenAI narrows its focus, rivals pursue categories it left behind or never entered. Neolabs may resemble biotechs: proving a research bet could make them challengers or acquisition targets.

虽然 OpenAI 因团队成员离职受到大量公开议论，但从 DeepMind 流向竞争对手的员工要多得多。热力图显示曾在行所示实验室工作、后来加入列所示实验室的人数，包含直接跳槽和中间经过其他机构的情况。

While OpenAI attracts lots of public flack for the departure of its team members, far more staffers have left

选择调用哪个工具、是否标记一笔交易，这些是分类任务，不是写作任务。TypeSafe 用两年时间围绕这一区别打造了 Jev：一个为预定义答案分配概率、不生成文本的模型，并且不像上一代机器学习那样需要特征工程。它的早期采用说明，即使在前沿模型云集的市场里，聚焦的研究押注仍能找到需求。TypeSafe 报告单次并行推理的响应时间为 70–500 毫秒，每百万输入 token 收费 0.042 美元，早期阶段输出免费。Vercel 称付费 AI Gateway 团队中约 13% 在 24 小时内使用了 Jev，是同期 GPT-5.6 份额的两倍。在 OpenRouter 上，Jev 十天内拿下每周分类请求的 27%（这是单一品类的采用率，不是整体用量或收入）。据报道，该公司在 7 天后的年化收入已达 1 亿美元。

Choosing which tool to call or whether to flag a transaction is a classification task, not a writing task. TypeSafe spent two years building Jev around that distinction: a model that assigns probabilities to predefined answers without generating text and unlike prior generation ML, obviates the need for feature engineering. Its early adoption suggests a focused research bet can find demand even in a market crowded with frontier models. • TypeSafe reports 70-500ms responses from a single parallel pass, at $0.042 per million input tokens and free output during early access. • Vercel says ~13% of paid AI Gateway teams used Jev within 24 hours, twice GPT-5.6’s share over the same window. • On OpenRouter, Jev took 27% of weekly classification requests within ten days: uptake in one category, not overall usage or revenue. • The company is reportedly on a $100M annualized run rate after 7 days.

消费级 AI 的头条和 2023 年 9 月第一份报告时一样：ChatGPT 是开创者，也仍是领先者。8 月网页访问量约为 Gemini 的 2 倍、Claude 的 6 倍；移动端月活为 Gemini 的 2.5 倍、Claude 的 14 倍。新的收入数据给出同样的结论：据 YipitData 的面板，ChatGPT 的美国消费者付费订阅数是 Claude 或 Gemini 的 3 倍。这一格局中最显著的变化是 Claude 成为明确的第二名：它在 2023 年 9 月的首份网页榜单上还未上榜，如今在流量上已超过 DeepSeek 和 Perplexity，并在今年早些时候超过 Gemini 的美国消费者订阅数；即便谷歌把存量付费用户迁移到 AI 套餐（人为抬高订阅数），Gemini 和 Claude 仍不相上下。Claude 的增长来自新产品（1 月的 Cowork、4 月的 Claude Design）、新模型（6 月的 Fable 5）和一轮公众关注；在 Anthropic 与五角大楼 2 月底的争议期间，Claude 首次登上美国免费移动应用榜首。与 OpenAI 和谷歌不同，Anthropic 明确表示不会在 Claude 中投放广告，而是通过订阅变现，并且做得很积极：Claude 有 7.3% 的消费者付费用户选择每月 100 美元起的最贵个人套餐 Max，谷歌为 1.3%、ChatGPT 为 1.1%。到夏季中期风向有所变化：YipitData 的全球桌面面板记录到 Claude 7、8 月的日均会话下降，其美国电子收据面板显示新增放缓、流失上升；与此同时，ChatGPT 随着 7 月 GPT-5.6 Sol、Terra、Luna 的发布和 ChatGPT Work 的推出重新加速。问题越来越不是“谁在获胜”，而是“每家公司如何把握自己的机会”。据 YipitData 面板，只有 8% 的美国 ChatGPT 订阅者同时订阅了 Claude；ChatGPT 的消费者在性别和收入上分布更均衡，并且在除加州、蒙大拿和马萨诸塞之外的所有州占优。用户差异部分源于产品方向不同：过去六个月，Anthropic、谷歌和 OpenAI 共有 127 次新发布，Anthropic 几乎完全聚焦专业消费者（Claude Design、Code Review、Claude Science），谷歌主攻创意模型（Lyria 3 Pro、Gemini Omni），OpenAI 同时面向消费者和企业。

1. Three LLMs, three different games Consumer AI’s headline is the same as our very first report in September 2023 - ChatGPT was the pioneer, and remains the leader. It has a roughly 2x lead on Gemi in August web visits and an even wider lead over Claude (6x). This gap is more significant on mobile, where ChatGPT has a 2.5x lead on Gemini and a 14x lead on Claude in monthly active users. Our new revenue data tells a similar story. Per YipitData’s panel, ChatGPT has 3x more U.S. consumer paid subscribers than Claude or Gemini. The most significant shift in this landscape has been Claude emerging as the clear # Claude did not even rank on our first web list in September 2023, but has now passe competitors like DeepSeek and Perplexity in traffic. Even more significantly, Claud passed Gemini on U.S. consumer subscribers earlier this year. Even after Google

达到 1 亿美元之后，Legora 和 Sierra 在约六个月内翻倍，Harvey 增长到 4 亿美元。Lovable 现报告 6 亿美元，Cursor 据报道已超过 40 亿美元。各公司的起点口径不同。（注：起点分别为上线（Higgsfield、Lovable、可灵、Sierra）、100 万美元 ARR（Cursor、Legora）、首笔收入（ElevenLabs）、成立（Harvey，上限口径）。计时与指标不同，月数为近似值；Cursor 曲线止于 10 亿美元，其后在第 17 个月超过 40 亿美元。）

After reaching $100M, Legora and Sierra doubled in about six months and Harvey grew to $400M. Lovable now reports $600M and Cursor has been reported above $4B. Starting points differ across companies.

Standard Metrics 把“AI 就是产品”的公司（AI 原生）与“增加了 AI 功能”的既有软件公司（AI 赋能）作对比。在 2026 年二季度的初步数据中，年化收入 100 万至 2,000 万美元区间，75 分位的收入增速为 256% 对 90%；2,000 万美元以上为 172% 对 53%。

Standard Metrics compares companies where AI is the product (“AI native”) with existing software businesses that have added it (“AI enabled”). In preliminary Q2 2026 data, revenue growth at the 75th percentile was 256% versus 90% at $1-20M in annualized revenue, and 172% versus 53% above $20M.

在每一个成立年份分组和每一类客户群体中，AI 原生公司的增速都超过 AI 赋能型 SaaS。在 75 分位上，2020 年后成立的公司同比增速为 487% 对 199%；面向中小企业和中端市场的公司为 303% 对 82%。在 2025 年一季度处于增长最快四分位的面向企业的公司中，一年后仍留在该四分位的 AI 原生公司占 42%，AI 赋能型 SaaS 为 23%，非 AI 公司为 28%。（注：左图合并了 2025 年三季度至 2026 年二季度、年化收入 100 万美元以上的公司数据。）

AI natives outgrow AI-enabled SaaS in every founding cohort and customer segment. At the 75th percentile, YoY growth is 487% versus 199% for companies founded since 2020, and 303% versus 82% for SMB and mid-market sellers. Among enterprise-focused companies in Q1 2025’s top growth quartile, 42% of AI natives stayed there a year later, versus 23% of AI-enabled SaaS and 28% of non-AI companies.

此外，代码量出现爆发，创业活动大幅上升。

Let’s Just Say That There Were Signs Elsewhere, there’s been an explosion of code, and a surge in entrepreneurship.

AI 的采用面广、影响大，但仍然相对浅。

Despite All The Demand, It’s Still So Early AI adoption is broad and impactful, but still relatively shallow.

多数公司并不受成本约束，并打算在 AI 上花更多钱。从大局看，AI 支出在整体成本结构中只占很小一块。可以参照：云计算在第二年达到 IT 支出的约 4%，目前多数对 AI 支出的估计都低于这个比例，而 AI 的潜在市场比 IT 预算大得多。

Enterprise AI Spending Has Plenty of Room-to-Run Most companies are not cost-constrained, and intend to spend more on AI. In the big scheme of things, AI spend is a tiny share of the overall cost stack. Consider that cloud reached ~4% of IT spend in year two. Most estimates of AI spend are below that, and TAM for AI is considerably larger than IT budgets.

实验室之外，智能体的采用走到哪一步了？Codex 的数据对比了 OpenAI 员工、机构用户和个人 ChatGPT 用户。作为参照，另一项对标普 500 公司 AI 披露的分析发现，69% 报告已有在用的部署，但只有 2% 披露了持续跟踪的指标。过去 28 天内，97.9% 的 OpenAI 在职员工使用过 Codex，机构用户为 17.3%，个人用户为 0.7%。至少分配过一个预计耗时八小时任务的个人 Codex 用户占比，从 2025 年 12 月的 2.1% 升至 2026 年 5 月的 25.6%。

How far along is agent adoption outside the labs? Codex data compare OpenAI staff with organizational and individual ChatGPT users. For context, a separate analysis of S&P 500 companies’ AI disclosures found that 69% reported live deployments, but only 2% disclosed metrics tracked over time. • 97.9% of active OpenAI workers used Codex in the preceding 28 days, compared with 17.3% of organizational users and 0.7% of individual users. • The share of individual Codex users assigning at least one estimated eight-hour task rose from 2.1% in December 2025 to 25.6% in May 2026.

在个人、机构和 OpenAI 三类人群中，非开发者周活跃用户的增速都快于开发者。从 2025 年 8 月到 2026 年 6 月，个人非开发者用户增长 137 倍，机构用户增长 189 倍，OpenAI 内部用户增长 12 倍（后者因基数本已很高）。这一增长也可能受到智能体能力的教育和认知提升的影响。

Weekly non-developer users grew faster than developers across individual, organizational, and OpenAI groups.

现在是谁在采用编程智能体？Codex 为软件工作而建，工程岗位仍占绝对用量的大头。增长最快的现在是法务、销售、招聘和市场，几乎都是从接近零起步。自 2026 年 2 月以来，企业 Codex 周活跃用户在法务增长 108 倍、销售 41 倍、招聘 41 倍、市场 26 倍，工程为 5 倍。深度上仍然偏向软件：2026 年 6 月，Codex 在机构内普通工程师的 ChatGPT 与 Codex 合计输出 token 中占 26.8%，普通法务用户为 1.9%。Anthropic 另外报告，使用 Claude Code 的非软件职业中增长最快的是管理、销售和法务。

Who is picking up coding agents now? Codex was built for software work, and engineering still dominates • Since February 2026, weekly active enterprise Codex users grew 108x in legal, 41x in sales, 41x in recruiting and 26x in marketing, against 5x in engineering. • Depth still favors software: in June 2026, Codex generated 26.8% of the average organizational engineer's combined ChatGPT and Codex output tokens, against 1.9% for the average legal user. • Anthropic separately reports management, sales and legal as the fastest-growing non-software occupation groups using Claude Code.

普通人在用 Claude 做什么？这能告诉我们哪些 AI 支出和基准测试说明不了的事？斯坦福 SALT 实验室的研究者通过 Anthropic 的隐私保护 Insights 流程，分析了 2026 年春季两周内的 249,834 条消费者 Claude.ai 对话，样本仅限聊天，不含 Claude Code 和 Claude Cowork。在被标注重要性等级的对话中，56% 属于有实质后果或高风险的工作，12% 为高风险。在可评估的对话中，72% 是人类主导、AI 辅助。逐字使用 AI 输出的比例，从一次性任务的 21% 降到高风险任务的 13%。50% 的对话中出现了摩擦，在出现摩擦的对话中，79% 的用户尝试了主动补救。

What are regular people using Claude for? And what does it tell us that AI spend and benchmarks don't? • 56% of conversations assigned a criticality tier were consequential or high-stakes work, and 12% were high stakes. • Humans led the work with AI assisting in 72% of assessable conversations overall. • Verbatim use of AI output declines from 21% on ephemeral tasks to 13% on high-stakes tasks. • Friction appeared in 50% of all conversations, and users attempted active recovery in 79% of the conversations containing friction.

2026 年 8 月，Ramp 统计中前 1% 企业的中位人均月支出为 7,205 美元，中位企业为 12.50 美元，前 10% 为 676 美元。Ramp 的另一项分析发现，1% 的客户贡献了 OpenAI 和 Anthropic 约 80% 的可观察支出。（注：为人均月支出，三张图刻度不同，数据为 2026 年 8 月。）

In August 2026, the median firm in Ramp’s top 1% spent $7,205 per employee per month, versus $12.50 for the median firm. A separate Ramp analysis finds that 1% of customers account for about 80% of observed spending on OpenAI and Anthropic.

极小一部分用户的 AI 用量比其他人高出几个数量级，而且这一差距从今年初开始持续扩大。

Power Users Are Pulling Away A tiny fraction are using orders of magnitude more AI than the field, and the gap has been widening since early this year.

亮眼的总量数据其实掩盖了水面之下的情况：采用面很广，但目前大多很浅。增长由极小一部分重度用户驱动，他们用日益复杂的工具和实施方案释放出更多价值。随着智能成本下降，前沿和非前沿模型的需求都在增加；智能体尤其推动 token 需求呈抛物线增长，因为路由、缓存等效率提升让它们在经济上变得可行，这正是实时发生的杰文斯悖论。需求的指数级增长推高了算力价格，连四五年前的 GPU 价格都和一年前差不多。而这一切只由前 1% 的用户驱动，他们才刚刚触及可能性的表面，其余市场最终会跟上，否则就会被甩下。

More Signs That We’re Still Early: Beneath the Surface, AI Is Not Normal The impressive aggregate picture actually obscures what is going on beneath the surface: adoption is wide, but mostly shallow (for now). Growth is driven by a tiny share of power-users, who are unlocking more value with increasingly sophisticated tools and implementations. As the cost of intelligence falls, demand continues to increase for both frontier and non-frontier models. Agents especially are driving parabolic growth in token demand, as routing, caching, and other efficiency gains, make them economically viable. It’s Jevons Paradox in real-time. The exponential increase in demand is driving the price of compute upwards–even 4-5 yo GPUs cost roughly the same as they did a year ago. And all of this is driven by just the top 1% of users, who are only beginning to scratch the surface of what’s possible. Eventually, the rest of the market will catch on (or else be left behind).

2023 年 9 月，风投支持公司的中位人均月支出为 3.40 美元，其他公司为 2.13 美元。到 2026 年 8 月，风投支持公司的支出在近三年里增长了 24 倍，其他公司为 3.9 倍，分别达到 81.20 美元和 8.33 美元。

In September 2023, the median VC-backed company spent $3.40 per employee per month, against $2.13 for other firms. By August 2026, VC-backed spend rose 24x over nearly three years, compared with 3.9x for other firms.

OpenAI 的研究和数据凸显了 AI 使用量“非正态”的分布特征。

AI Adoption Is Wide and Shallow (For Now) OAI research and data have highlighted the not-normal distribution of AI usage

各公司财报和公开发言中的例子：亚马逊称使用 Alexa 购物的客户平均每单多花 40% 以上；Meta 称 GEM 使 Facebook 上的对话量提升 15%，Advantage+ 带来 13% 的增量购买和 16% 的加购转化提升；Shopify 称 Sidekick 使 15 天内达成 5 单的新商家增加约 8%，AI 搜索的转化比传统自然搜索高 80%；ServiceNow 称其 AI 年合同价值在 2026 年二季度突破 10 亿美元，智能体部署在 9 个月内增长 9 倍；Affirm 用模型批准了原有系统会拒绝的申请（包括信用记录有限、无 FICO 评分者），多完成 3.4% 的购买，且这些贷款表现优于同等规模的扩张；Lemonade 称理赔费用率降至创纪录的 5% 低位；Snowflake 称过去四年每年降低 10% 的服务成本；Chime 称 CoCo 账户超过 9,100 个，单季新增逾 2,000 个；微软称 GitHub Copilot 收入环比加速超过 60%，Microsoft 365 Copilot 付费席位已超过 3,000 万；沃尔玛称使用 Sparky 的客户和会员比不使用者每单多花 40%，用户数同比增长 70%。

Anecdata Around AI ROI Abound AI Shopping & Ads AI Ad Performance AI Ad Performance AI ACV & Deployments AI Underwriting “...customers who use Alexa for shopping spend an average of over “...GEM... a 15% uplift in conversations on Facebook” “...Sidekick...drove like an 8% increase in new merchants reaching “ServiceNow AI crossed $1 billion in Affirm used the model to approve. . . applications that its existing system annual contract value in Q2 2026” 40% more per order...” 5 orders within 15 days.” would have declined, including . . . limited credit histories and no FICO “...Advantage+... 13% incremental lift in purchases and a 16% increase Earnings release “Ads Agent...6% lower cost-per-acquisition...” “AI search...80% higher than traditional organic search.” scores. . . that produced 3.4% more in add-to-cart conversions.” “...Agentic deployments of ServiceNow AI increased ninefold in completed purchases, and those additional loans performed better just 9 months.” than a comparable expansion." Earnings Release, Amazon, Q2, July Susan Li, CFO, Meta, Q2, July 2026 Harley Finkelstein, President, Shopify, Q2, August 2026 Company release Sept. 2026 Bill McDermott, Chairman and CEO, ServiceNow, Q2, July 2026 AI Claims Handling AI Cost Efficiency CoCo Adoption Copilot Revenue & Seats AI Shopping “Our LAE ratio declined to a record “...AI also continued to drive outsized efficiency gains...reduced “CoCo surpassed 9,100 accounts... adding more than 2,000 accounts in “...GitHub Copilo

BCG 把 107 家过去 12 个月收入超过 5 亿美元的上市科技公司，按 2025 年 4 月至 2026 年 3 月的 Cursor 月度 token 消耗量分成五组。每一组的中位同比收入增速都在上升，最大的跃升出现在第三组到第四组之间，而不是在最高组。

BCG grouped 107 public technology companies above $500M in trailing-12-month revenue into quintiles by

Claude Opus 5 在四项结构化会计任务的 20 次尝试中全部得到 100%。但在 ATLAS-Finance 的 100 项模拟银行任务中（要求工作簿正确且交付方式得当），它只通过了 12.3%。更好的环境、评测和专家训练数据可能缩小这一差距。（注：会计任务中，12 名持证注册会计师独立作业，限时三小时，没有同事、客户或积累的公司上下文；在一次失败的任务中，Opus 5 通过了全部 13 项自检，却是在用自己错误的财务假设来验证。每个点代表一次尝试：23 次人类尝试、20 次 AI 尝试，取前三个模型，无评分的人类基线和误差线。）

Claude Opus 5 scored 100% across 20 attempts on four structured accounting tasks. However, it passed only 12.3% of ATLAS-Finance’s 100 simulated banking assignments, which require a correct workbook and proper delivery. Better environments, evaluations and expert training data may close this gap.

在领先的云厂商和新云厂商之外，一个新的基础设施层正在形成：专业推理服务商的数据让我们看到这一生态增长的速度和规模。

Inference & Insinuations Outside of the leading clouds and neoclouds, a new infrastructure layer is emerging. Specialized inference providers provide visibility into the speed and scale of the ecosystem’s growth.

Stripe 对支付流有无可比拟的观察视角，其数据更加不支持“SaaS 末日”的说法：SaaS 收入非但没有死去，反而在进入新年时加速。

SaaS . . .acceleration? Stripe has unmatched visibility into payment flows, and their data is even more counter to the SaasPocalypse narrative: far from dead, SaaS revenues accelerated into the new year.

我们首次加入了基于观察到的消费者支出的排名，数据来自 YipitData。数据仅限美国，来自面板而非普查，不应视为公司总收入，但它提供了消费者如何为 AI 付费的新视角，结论很明确：高度集中。付费人群在扩大，但仍然很窄：8 月，YipitData 美国电子收据面板中 4.5% 的合格消费者至少订阅了 ChatGPT、Gemini 或 Claude 中的一个，一年前为 2.1%。即使在这一相对狭窄的付费人群内部，支出也遵循幂律：为一个 AI 产品付费的用户中，只有 13% 会再为另一个 AI 工具付费；前 1% 付费者贡献了全部可观察消费 AI 支出的 19.5%，超过支出最低的 50% 用户之和（16.6%）。截至 2026 年 8 月，这前 1% 的人平均每月在消费卡上为 AI 支出 903 美元（工作用途可能更高），且仍在增长，过去 18 个月增加了 80%；与中位付费者的差距明显：后者每月 25 美元，且几乎没有增长。这些重度支出者更像专业消费者而不是主流消费者：他们买的是用来构建、创作和完成工作的软件，不成比例地购买 n8n、fal、Manus、Nous Research 等自动化和产品构建工具，也更偏好 Higgsfield、Fig、HeyGen 等创意工具。因此，消费流量和消费收入之间差异显著：按支出排名的前 50 家厂商中，有 29 家没有出现在我们的网页或移动流量榜上；只有七家公司同时进入三个榜单，其中 ChatGPT 三榜第一，另有三家模型公司（ChatGPT、Claude、Suno）、两家 AI 原生应用（Perplexity、Photoroom）和两家增加了 AI 的既有产品（Canva、Notion）。随着商业模式打开，我们预计付费人群会扩大，个人智能体尤其有可能按交易而非按席位变现，提供接近无限用量而无需订阅的形态。

2. AI spending follows a power law For the first time, we’re adding a ranking based on observed consumer spending, us data from YipitData. The data is U.S.-only and drawn from panels rather than a census; so it should not be read as total company revenue. However - it gives us a ne view of how consumers pay for AI, with a clear takeaway: heavy concentration. The base of consumers paying for AI is growing, but still narrow. In August, 4.5% o eligible consumers in YipitData’s U.S. e-receipt panel had an active paid personal subscription to at least one of ChatGPT, Gemini, or Claude, up from 2.1% a year ago

在消费端，一切也才刚刚开始。

On the Consumer Side, Things Are (Also) Just Getting Started

AI 助手帮购物者在访问零售商之前比较商品、缩小选择范围。虽然来自 AI 的直接导流在电商流量中占比仍小，但其快速增长和更高的转化率说明它正成为越来越有价值的客户来源。Similarweb 报告 AI 导流年增长 203%，但仍只占零售电商访问量的 0.4%。2026 年二季度，Shopify 上经 AI 导流进入商品页的访客，转化率比自然搜索访客高约 80%。Profound 发现，零售品牌在 ChatGPT 中被提及后的七天内，网站访问量比预测基线高 38%。

AI assistants help shoppers compare products and narrow their choices before visiting a retailer. While direct AI referrals remain a small share of ecommerce traffic, their rapid growth and higher conversion rates suggest an increasingly valuable source of customers. • Similarweb reports 203% annual growth in AI referrals, which still account for only 0.4% of retail ecommerce visits. • Shopify’s AI-referred product-page visitors converted about 80% more often than organic-search visitors in Q2 2026. • Profound finds that retail brand mentions in ChatGPT were followed by 38% more site visits than the forecast baseline over seven days.

消费级 AI 带来的搜索导流在增长，但绝对规模仍然很小。

Consumer AI: Search Referrals Gaining (but still very small)

2025 年 7 月，扎克伯格承诺“给每个人的个人超级智能”：帮助人们追求自己目标的 AI。Muse 给每个用户一台常驻云端的电脑，可以跨应用执行操作；电商和企业服务把这一愿景变成个人和企业愿意付费的服务。消费端拉动：两周 280 万次下载，登上美国应用榜第一，此前 Llama 4 反响不佳、Meta AI 的应用也表现挣扎；9 月 21 日 Meta 股价上涨超过 11%。电商：Stripe Link 让 100 万以上商户支持结账，与 Shopify、沃尔玛、Shop Pay 和 PayPal 的合作把发现与支付连起来。企业：前 MongoDB 首席执行官 CJ Desai 将负责销售 API、编程工具和企业智能体；推理需求有助于 Meta Compute 变现，直接出售算力也在考虑中。Muse 也将进入 Meta 的 AI 眼镜。

In July 2025, Zuckerberg promised “personal superintelligence for everyone”: AI that helps people pursue their own goals. Muse gives each user a persistent cloud computer to act across apps. Commerce and enterprise turn that vision into services people and businesses can pay for. • Consumer pull: 2.8M downloads in two weeks and No. 1 on US app charts, after Llama 4’s poor reception and Meta AI’s earlier app struggles. Meta shares rose >11% on September 21. • Commerce: A step toward our 2025 prediction. Stripe Link enables checkout at 1M+ merchants, and Shopify, Walmart, Shop Pay and PayPal partnerships connect discovery to payment. • Enterprise: Ex-MongoDB CEO CJ Desai will sell APIs, coding tools and business agents. Inference demand helps monetize Meta Compute, and direct compute sales are under consideration.

大家都在等待那个“超级应用”，Muse 已经进入聊天赛道。

Waiting for the SuperApp: Muse Has Entered the Chato

3 月我们写道，OpenClaw 这类智能体还不适合多数消费者使用，但终将可用。六个月后，多个面向消费者或专业消费者的智能体产品报告了数十万用户，包括 Instinct、Tomo、Poke、Lindy 和 Town；巨头也以自己的智能体反击：Meta 的 Muse、OpenAI 的 Dots、xAI 的 Grok Bot。多数智能体力求无处不在，像人类助理一样可以发短信、发邮件或打电话。当下最热的渠道是 iMessage，爆红智能体 Instinct 于 2026 年 8 月在此上线，DoorDash 等公司随后推出了自己的 iMessage 智能体用于短信下单。更偏专业消费者的智能体，或聚焦特定场景的消费智能体，通常还会有自己的移动或网页应用以提供更深入的体验。

something closer to unlimited usage without a subscription. More on that later in th report. 3. The race to become your personal assistant In March, we wrote that agents like OpenClaw were not yet usable by most consum - but would be. Six months later, multiple startup consumer or prosumer agent products now report hundreds of thousands of users, including Instinct, Tomo, Pok Lindy, and Town. And, the incumbents have struck back with agents of their own - Meta’s Muse, OpenAI’s Dots, and xAI’s Grok Bot. Most of these agents strive to be available everywhere - like a human assistant, you text, email, or call. The channel du jour is iMessage, where viral agent Instinct launched in August 2026. Companies like DoorDash have since launched their own iMessage agent for text-to-order. More prosumer-oriented agents, or consumer age focused on a specialized use case, often also have their own mobile or web app to provide a deeper experience.

随着 AI 系统能力提升，行业将越来越看重完成的工作而不是对话。但这会如何改变 AI 的商业模式？能力从聊天、编程、智能体、协同工作到自主 AI 逐级上升，每一层都改变了产品的售卖方式：从免费聊天和订阅，走向按完成任务计费的成果定价。聊天：交付一次回答，市场是搜索，免费或订阅。编程：交付可运行的代码，这是第一道客观检验，市场是软件研发，订阅或 API。智能体：交付一条完成的多步任务链，市场是知识工作，订阅或 API，需要可验证的结果。协同工作：交付可供审阅的成果，按任务计价，需要专业人员审阅而不是重做。自主 AI：无需监督持续运行，市场是全行业升级，按任务成果计价，需要长周期规划和错误恢复，以及判断自身输出是否正确的能力。

As AI systems become more capable, the industry will increasingly value finished work over chats. But how will this change the business model of AI? As capability advances from Chat to Coding to Agent to Co-work to Autonomous AI, each level changes how the product is sold, from free chat and subscriptions to outcomes priced per completed task.

为加快企业采用自家系统，两家实验室在 2026 年 5 月成立了服务公司提供部署服务，出资方是私募股权机构，而这些机构的被投公司正是客户。它们面临的问题是只能绑定母公司的技术栈。Anthropic 与黑石、Hellman & Friedman 和高盛的合资公司有约 15 亿美元承诺资本。OpenAI 的部署公司以 100 亿美元投前估值、从 TPG 领投的 19 家投资者募集超过 40 亿美元；有报道称投资者可获得 17.5% 的最低年化回报保证，上行受限，更接近信贷而非股权。两家都是通过收购小型 AI 咨询公司起步：Anthropic 收购 Fractional AI，OpenAI 收购 Tomoro。

To accelerate enterprise adoption of their systems, the two labs launched services companies in May 2026 to provide deployment services, capitalized by the private equity firms whose portfolio companies are the customers. Their challenge, however, is being bounded to the stack of their parent company. • Anthropic's venture with Blackstone, Hellman & Friedman and Goldman Sachs carries about $1.5B of committed capital. • OpenAI's Deployment Company raised more than $4B at a $10B pre-money valuation from 19 investors led by TPG. Reports suggest that investors get a guaranteed 17.5% minimum annual return with capped upside, closer to credit than equity. • Both built their deployment companies from the acquisition of smaller AI consultancies: Fractional AI for Anthropic and Tomoro for OpenAI.

公司可以先用前沿 API 找到产品市场契合，再建立自己的评测和工具。生产环境的反馈会指出系统该在哪里改进；当自有数据以及质量、延迟或成本上的收益足以抵消代价时，后训练才值得做。失败的任务会变成可重复的测试，附带判断成败所需的上下文、工具和专家检查；修正、重试和执行轨迹能揭示失败来自缺少上下文、工具还是模型本身；改进可以作用于记忆、工具、路由或模型权重，每次改动都要在部署前测试。

Companies can start with frontier APIs to find product-market fit, then build their own evaluations and tools. Production feedback shows where to improve the system. Post-training becomes worthwhile when proprietary data and gains in quality, latency or cost justify taking control of the model. • Failed tasks become repeatable tests with the context, tools and expert checks needed to judge success. • Corrections, retries and execution traces reveal whether a failure came from missing context, a tool or the model. • Improvements can change memory, tools, routing or model weights, with each change tested before deployment.

中国的前沿开源模型，让 Harvey、Cursor、Mercor 这类应用公司有机会在客户关心的领域降低成本、提升性能。Harvey 后训练的 GLM-5.2 在生产环境中每单元格成本比 Sonnet 5 低 54.8%，其法律审查表格的答案质量得分为 0.903，Fable 5 为 0.867。Mercor 用 1,928 个专家任务后训练 Qwen3.5-397B-A17B，把 480 个留出 APEX-Agents 任务的总体 Pass@1 从 16.11% 提高到 27.29%。Cursor 则基于 Kimi K2.5，合成的强化学习任务量是 Composer 2 的 25 倍。在这些场景中，公司聚焦某个知识工作垂直领域，就能用专家数据构建最好的评分任务环境。

Frontier open Chinese models are handing application companies like Harvey, Cursor, and Mercor the opportunity to reduce costs and boost performance on the domains their customers care about. • Harvey's post-trained GLM-5.2 runs in production at 54.8% lower cost per cell than Sonnet 5, and scores 0.903 on their Legal Review Table answer quality against 0.867 for Fable 5. • Mercor post-trained Qwen3.5-397B-A17B on 1,928 expert tasks, raising overall Pass@1 on 480 held-out APEX-Agents tasks from 16.11% to 27.29%. • In these settings, a company’s focus on a knowledge work vertical enables them to build the best graded task environments with expert data.

过去，客服智能体由厂商的前置工程师搭建（把客户政策转化为智能体行为），由团队读对话记录来修复。现在厂商出售的智能体两件事都能做：员工用平实的语言描述改动，并审批智能体已经测试过的修复。搭建方面：PolyAI 称在 9 月 21 日那一周，客户部署的改动中有 87% 是通过 Wren 完成的，7 月 13 日那周为 52%。修复方面：Decagon 的 Autopilot 在过去的对话上测试每一处修复，再由人审批；在 Decagon 自己的诊断基准上，它以 93% 对 83% 超过了经过认证的 Decagon 员工。速度方面：Decagon 称每次语音上线所需的工程工时在六个月内减少一半以上，上线时间缩短约 37%。

Customer service agents used to be built by the vendor's forward-deployed engineers, who turn a customer's policies into agent behavior, and fixed by teams reading transcripts. Vendors now sell agents that do both: staff describe changes in plain English and approve fixes the agent has already tested. • Build: PolyAI says customers used Wren for 87% of changes they deployed in the week of 21 September, up from 52% in the week of 13 July. • Fix: Decagon's Autopilot tests each fix on past conversations, then a person approves it. It beat certified Decagon staff on Decagon's own diagnostic benchmark, 93% to 83%. • Speed: Decagon says engineering hours per voice launch more than halved in six months, and time to go live fell about 37%.

一款药“由 AI 做出”到底意味着什么？各公司的答案不同。两个项目已进入三期临床，主要终点预计在 2028–29 年完成；发现阶段更快，尚未证明临床成功率更高。GB-0895（重度哮喘抗体）：机器学习引导的优化使其与 TSLP 的结合达到 106 飞摩尔，约为 tezepelumab 的 20 倍；已进入三期，两项 786 人试验，主要终点预计 2028 年 12 月／2029 年 1 月完成。rentosertib（肺纤维化 TNIK 口服药）：PandaOmics 从组学和文本数据中把 TNIK 排在激酶首位，Chemistry42 生成抑制剂；三期，中国 47 个中心 320 名患者、52 周，预计 2029 年 10 月完成。ENV-294（湿疹和哮喘口服药）：在 12 亿张质谱图上预训练的 Transformer 预测未知天然分子的结构；二期 a 阶段，1b 期中湿疹严重程度平均在第 42 天下降 85%（n=9，无安慰剂对照）。REC-4881（息肉病 MEK 口服药）：对 APC 缺陷细胞显微图像做深度学习，按表型拯救效果给数千种化合物排序；二期，第 13 周息肉负荷中位数下降 43%（n=12，无安慰剂）。IAM1363（实体瘤 HER2 口服药）：扩散模型 NeuralPLexer 预测药物—蛋白结构，Enchant 预测临床终点；1/1b 期，在每日 960 毫克以上剂量的 18 名经治患者中 28% 出现部分缓解。

What does it really mean for a medicine to be “made by AI”? The answer differs by company, so this table

我们一直认为 AI 会变革最依赖人力的服务业，现在仍坚持这一判断。虽然还很早，但 Protege 的数据显示出一些早期回报：AI 已经在挽救生命。

AI is Saving Lives That AI will revolutionize even the most labor-intensive services is our story, and we’re sticking to it. It’s early, but Protege data shows some early returns: AI is already saving lives

AI 正以各种方式扩散到更广泛的经济中，有时收益显著，但更多时候是边际上的渐进改善（不过这些改善往往会累积成相当可观的结果）。2025 年 11 月底，EliseAI 与 ElevenLabs 和 Cartesia 合作，从其语音库中挑选出带特定地区口音的高水平语音。背后的假设是：既有研究表明消费者更愿意与听起来和自己相似的人交易，那么换成本地化语音应当能提升“线索到看房”的转化率。

Making life easier for renters and owners with AI AI is diffusing into the broader economy in disparate ways. In late November 2025, EliseAI worked with Elevenlabs Sometimes the benefits are substantial, but often they are incremental at the margins (which, nonetheless, have a tendency to add and Cartesia . . . to collect a series of high level voices with speci;c regional accents curated from their robust voice library . . . The underlying hypothesis was that, given existing research showing consumers are more likely to transact with people who sound like them, that up to something quite substantial). switching to localized voices would positively impact lead-to-tour conversion rates.

国防开支不仅在上升，优先级也变了：弹药变得更小、更便宜、面向规模化生产。

A New Defense Prime Not only is defense spending rising, but priorities have changed. Munitions are smaller, cheaper, and built for scale.

硬件创新的影响不止于芯片和国防。自动车牌识别系统是长期人手不足的警察部门的力量倍增器。

Stopping Crime Is Good Actually Hardware innovation is making a difference beyond chips and defense. ALPRs are a force-multiplier for chronically understaffed police departments.

在一片 AI 之中，还有别的变化正在发生。社交媒体颠覆了传统零售的分发方式，创作者在打造自己的品牌；加上人们对稀缺之物（比如收藏品）的需求上升，催生了新一代的电视购物。

Collectibles Are Timeless Amidst all the AI, there are other changes afoot. Social media has upended traditional retail distribution, and creators are building their own brands. Combined with an increasing desire for scarce things (like collectibles), it’s led the birth of a new QVC.

实验室需要的是“工作是怎么完成的”的样例，而不只是完成的答案。经专家修正的操作轨迹可以改进智能体，而数据中介现在会为企业的运营记录付费，例如操作规程、CRM 历史和项目档案。OpenAI 和 Thrive 复用了经专家修正的报税智能体轨迹，六周内按不低于 75% 准确率完成申报的比例从 25% 升到 86%。micro1 以 10 万到 100 万美元以上的价格授权企业的流程和记录；Mercor 称企业数据合作的报酬可超过 1,000 万美元。这是两类不同的资产：轨迹揭示决策和修正过程，企业档案提供领域背景。中介的付费金额为相关公司自行披露。

Labs need examples of how work gets done, not just finished answers. Expert-corrected action traces can improve agents, while brokers now pay companies for operating records such as procedures, CRM histories and project archives. • OpenAI and Thrive reused expert-corrected tax-agent traces. Documents filed at ≥75% accuracy rose from 25% to 86% within six weeks. • micro1 offers $100k to $1M+ to license company procedures and records. Mercor says enterprise data partnerships can pay more than $10M. • These are different assets: traces reveal decisions and corrections, while company archives supply domain context. Broker payment claims are company-reported.

强化学习环境公司越来越多，因为需求由一个正反馈循环驱动：实验室发布模型，并用自己创建的评测和环境邀请社区测试；这些评测备受关注和争议，因为它们暴露了模型的长处和短板，于是实验室购买针对性数据、改进模型、再次发布，又暴露出新的短板。Mercor 2026 年 6 月总年化收入 20 亿美元（2025 年 2 月为 7,500 万，2025 年 9 月为 5 亿）；Handshake AI 2026 年 4 月接近 10 亿美元（仅 AI 部分，2026 年 1 月为 5.5 亿）；micro1 2026 年 9 月超过 5 亿美元（2025 年初为 700 万）；Surge AI 2024 财年全年收入 12 亿美元；Scale AI 2025 财年略低于 10 亿美元。

More and more RL environment companies are springing up thanks to a positive feedback loop drives demand:

Ramp 把自家的卡和账单支付 AI 支出数据，与 Revelio 的 21,559 家美国公司员工记录关联，按 AI 使用程度做对比。重度支出者（人均每月 33.67 美元的前三分之一）两年内员工数增加 10.2%，入门岗位增加 12%，而轻度采用者始终没有与对照组拉开差距。科学家是例外：两组都在增长，并在第 24 个月趋同。几乎所有增量都来自科技公司。

Ramp linked its card and bill-pay AI spend to Revelio headcount records for 21,559 US firms to compare them

AI 使用强度更高的公司招聘更多，软件工程师尤其明显。

AI Intensity Leads to MORE Hiring, Especially For SWEs

AI 建设带来了对技术工人的抢夺，蓝领工资出现大幅溢价。

Capex Is Having A Positive Impact On The Broader Economy The AI buildout has led to a run on skilled labor, with a big premium to hard hat wages.

《经济学人》估计，相对于更广泛的招聘趋势，AI 建设多创造了 32 万个美国基建岗位和 73 万个 AI 相关职业岗位。基建岗位包括电气承包、暖通与管道、公用设施建设、商业建筑和电气设备制造；AI 职业包括软件开发、数学与数据科学、信息安全分析、工程师等。在办公室岗位中，数据科学家（+37%）、金融分析师（+11%）、律师助理（+11%）就业增长，而数据录入员（−18%）、客服（−9%）、记账员（−9%）、平面设计师（−7%）下降。

The Economist estimates the creation of 320,000 extra US infrastructure jobs and 730,000 extra jobs in AI-related professions relative to broader hiring trends. Among office roles, employment grew for data scientists and paralegals but fell for data-entry clerks and customer-service staff.

Anthropic 研究了 AI 如何影响工作和学习。美国劳动力数据显示，AI 暴露程度高的职业中，年轻劳动者的招聘有走弱的初步迹象。另一项实验中，使用 AI 的开发者事后更难解释代码如何工作、也更难找出缺陷。自 2022 年底以来，AI 暴露岗位相对非暴露岗位的失业率没有明显上升；但 22 至 25 岁进入 AI 暴露岗位的入职速度似乎放缓，原因尚不确定。52 名开发者学习了 Python 并发库 Trio，随后在不使用 AI 的情况下做 14 题测验：借助 AI 学习的一组得分 50%，不用 AI 的一组为 67%。测验考查两项编程任务后对代码的阅读、调试和对库的理解；AI 带来的速度提升在统计上不显著。该实验衡量的是即时学习效果，不是长期技能退化。

Anthropic studied how AI affects jobs and learning. US labor data show tentative signs of weaker hiring among young workers in AI-exposed occupations. In a separate experiment, developers who used AI were less able to explain how the code worked and identify bugs afterward. • No clear rise in unemployment in AI-exposed jobs relative to unexposed jobs since late 2022. • Job starts appear to have slowed for 22-25-year-olds entering AI-exposed jobs. The cause remains uncertain. • 52 developers learned Trio, a Python library for concurrent tasks, then took a 14-question quiz without AI. It tested code comprehension: scores were 50% after learning with AI versus 67% without it. The quiz tested code reading, debugging and understanding of the library after two coding tasks. The speed gain from AI was not statistically significant. This measures immediate learning, not long-term skill loss.

好的家教帮助学生自己解决问题，这需要的是教学行为，而不只是正确答案。塞拉利昂的一项课堂试验衡量了教师主导的 Gemini 活动带来的学习提升，而家教类基准则说明了通用助手为何仍会“帮过头”。在 12 所学校的 48 个班级中，1,763 名学生被随机分配到教师主导的 Gemini 活动或常规教学；八周后数学成绩提高 0.258 个标准差，95% 置信区间为 0.027–0.488。这检验的是结构化的课堂项目，不是不受限制的聊天机器人使用，也不是长期学习效果。TutorMoments 和 MathTutorBench 的结果表明，给出正确答案与教会学生是两回事，如何引入课程上下文、如何用提问促使学生思考，仍是产品设计上的挑战。

A good tutor helps students solve problems themselves. That requires teaching behavior, not just correct answers. A classroom trial in Sierra Leone measured learning gains from teacher-led Gemini activities, while tutor benchmarks show why general assistants can still over-help. • In 48 classrooms across 12 schools, 1,763 students were randomized to teacher-led Gemini activities or standard instruction. Math scores improved by 0.258 standard deviations after eight weeks. • The 95% confidence interval was 0.027-0.488. This tests a structured classroom program, not unrestricted chatbot use or long-term learning. • TutorMoments and MathTutorBench find that giving correct answers is different from teaching. Curriculum context and prompts that make students think remain product-design challenges.

AI 融资环比下降 33%，至 1,033 亿美元，交易活动也下降 21%；仅 154 笔大额融资（1 亿美元以上）就拿走了全部资金的 85%。三季度新增独角兽 42 家，为 18 个月来最高，总数达到 706 家，比二季度增加 35 家；最大的新晋者包括小鹏机器人（63 亿美元）、River AI（50 亿）、micro1（40 亿）、Olix（33 亿）；美国以 24 家领先，亚洲 11 家，欧洲 7 家。三季度 61% 的融资流向美国公司：美国公司拿到 1,033 亿美元中的 630 亿，而交易量占比不到一半；亚洲和欧洲各为 188 亿美元（各占 18%）。三季度并购退出 372 笔，环比下降 17%，但仍远高于 2024 年前的水平；退出总数 393 笔，含 19 起 IPO 和 2 起 SPAC 退出。最大的几宗 IPO 由中国 AI 芯片公司领衔：燧原科技（91 亿美元）、宇树科技（90 亿）、Momenta（80 亿），19 起 IPO 合计退出价值 269 亿美元，与美国总额相当；美国只有 Intellai 一家，印度、以色列和英国各有一起。最大的几笔融资为 OpenAI 100 亿美元、Databricks 50 亿、Safe Superintelligence 50 亿、Crusoe 39 亿，四笔合计 239 亿美元，接近全部融资的四分之一。

TL;DR TL;DR Your rundown on AI in Q3’26 $103.3B 61% AI funding falls 33% quarter-over-quarter New unicorns in Q3, highest in 18 months of Q3 funding goes to US-based AI cos M&A exits in Q3, down but still elevated Top IPOs led by Chinese AI chip cos Deal activity fell too, down 21% QoQ.Just 154 mega- rounds ($100M+) captured 85% of all funding. Total unicorn count reached 706, up 35 from Q2. Top debuts: XPeng Robotics ($6.3B), River AI ($5B), micro1 ($4B), Olix ($3.3B). The US led with 24 new unicorns, followed by Asia with 11 and Europe with 7. US-based companies captured $63B of Q3's $103.3B total, despite representing just under half of deal volume. Asia and Europe split the remainder, both reaching $18.8B (18% of funding). Acquisitions dropped 17% from Q2, but stayed well above pre-2024 levels. Total exits reached 393, including 19 IPOs and 2 SPAC exits. Enflame ($9.1B), Unitree Robotics ($9B), and Momenta ($8B) led the quarter’s 19 IPO exits. Their combined $26.9B exit value matched the US total. OpenAI's $10B led, followed by Databricks ($5B), Safe Superintelligence ($5B), and Crusoe ($3.9B). Together, these four deals totaled $23.9B, nearly a quarter of all funding. The quarter's largest exits were Anysphere to SpaceX ($60M), OpenRouter to Stripe ($8M), Altruist to Vanguard ($4.6M), and Modular to Qualcomm ($3.9M). Only Intellai emerged from the US, while India, Israel, and the UK each posted single IPOs. The US share edged up from Q2's 59%, driven by concentration of

2026 年初至今融资 4,956 亿美元，2025 年为 2,726 亿，2024 年为 1,557 亿，2023 年为 1,008 亿，2022 年为 1,373 亿。交易数今年至今为 8,242 笔，2022 至 2025 年分别为 13,298、12,599、13,279 和 13,796 笔。

Annual equity funding & deals $600.0B Funding $495.6B 20,000 $500.0B 15,000 13,298 13,796 $400.0B 13,279 12,599 $300.0B Deals 8,242 10,000 $200.0B 5,000 $100.0B $137.3B $100.8B $155.7B $272.6B $0.0B 2022 2023 2024 2025 2026 YTD

2026 年三季度融资 1,033 亿美元，二季度为 1,531 亿，一季度为 2,391 亿（2026 年一季度为历史峰值）。三季度交易数 2,180 笔，是 2022 年以来最低；2025 年各季度分别为 599 亿、634 亿、935 亿美元。

Quarterly equity funding & deals 6,000 5,000 $200.0B 4,213 3,803 4,000 3,713 3,655 $150.0B 3,393 3,287 3,313 3,307 3,183 3,244 3,158 3,184 3,075 2,981 3,018 2,945 Funding $103.3B 2,807 3,000 2,755 $100.0B Deals 2,180 2,000 $50.0B 1,000 $48.8B $41.2B $25.2B $22.2B $28.7B $23.3B $23.6B $25.3B $27.8B $37.5B $27.1B $63.3B $55.9B $59.8B $63.4B $93.5B $239.1B $153.1B $0.0B Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

美国 630 亿美元、1,079 笔；亚洲 188 亿美元、538 笔；欧洲 188 亿美元、458 笔；大洋洲 22 亿美元、30 笔；加拿大 4 亿美元、36 笔；拉美 8,800 万美元、28 笔；非洲 3,900 万美元、10 笔。

Funding & deals by global region in Q3’26 Canada $0.4B 36 Deals Asia $18.8B 538 Deals Europe $18.8B 458 Deals US $63.0B 1,079 Deals Africa $39M 10 Deals All Other Regions $0.0M 1 Deal LatAm $88M 28 Deals Oceania $2.2B 30 Deals

2026 年三季度，美国占交易数的 49%，亚洲 25%，欧洲 21%，加拿大 2%，拉美和大洋洲各 1%。美国占比自 2025 年的 42%–43% 回升。

Percent of quarterly deals by global region LatAm, 1% Canada, 2% Africa, 0% Oceania, 1% All Other Regions, 0% 26% 27% 25% 23% 25% 24% 26% 23% 25% 24% 23% 23% 23% 23% 23% 25% 24% 22% Europe, 21% 18% 20% 23% 23% 22% 24% Asia, 25% 20% 24% 19% 21% 22% 20% 22% 24% 24% 26% 28% 23% 51% 51% 48% 47% 49% 47% US, 49% 49% 47% 46% 45% 46% 47% 48% 44% 42% 42% 43% 42% Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

2026 年三季度，风险投资占 32%（2022 年以来首次低于 33%），其他类投资者 18%，企业投资 10%，天使 9%，资产／投资管理 9%，企业风投 8%，孵化器／加速器 8%，私募股权 7%。

Percent of quarterly deals by investor group 6% 6% 5% 5% 6% Private Equity, 7% 8% 7% 8% 7% 6% 8% 7% 7% 6% 7% 6% 7% 7% 7% 7% 6% 7% 7% CVC, 8% 8% 13% 8% 10% 8% 11% 9% 13% 9% 10% 9% 7% 7% 8% 7% Incubator/Accelerator, 8% 10% 7% 11% 7% 6% 7% 8% 8% 7% 6% 6% 7% 6% 7% 8% 8% 8% 6% 6% Asset/Investment Management, 9% 8% 7% 11% 10% 10% 10% 11% 11% 12% 11% 11% 12% 11% 11% 11% 11% 11% Angel, 9% 10% 10% 10% 7% 8% 8% 8% 7% 8% 7% 8% 7% 8% 7% 7% 7% 9% 9% 9% 9% Corp, 10% 8% 20% 17% 17% 17% 16% 15% 17% Other, 18% 17% 19% 18% 19% 18% 19% 19% 21% 20% 20% 19% 35% 35% 34% 34% 36% 36% 36% 36% 36% 35% 33% 36% 37% 37% 37% 38% 36% 35% VC, 32% Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

平均交易规模从 2022 年的 1,290 万美元升至 2026 年初至今的 9,130 万美元，其间为 1,580 万、1,950 万和 3,170 万。中位交易规模从 300 万升至 470 万美元，涨幅远小于平均值。

Annual average & median deal size Average Deal Size Median Deal Size $91.3M $31.7M $19.5M $15.8M $12.9M $3.0M $2.1M $2.5M $3.5M $4.7M 2022 2024 2026 2022 2024 2026 YTD YTD

2026 年初至今，美国中位交易规模 530 万美元，亚洲 300 万，欧洲 430 万，拉美 250 万，加拿大 400 万，非洲 100 万，大洋洲 220 万。

Annual median deal size by global region US Asia Europe LatAm Canada Africa Oceania $5.3M $4.4M $4.5M $4.3M $4.0M $4.0M $3.0M $2.9M $3.0M $2.5M $2.5M $2.4M $2.3M $2.2M $2.2M $2.0M $1.6M $1.4M $1.5M $1.0M $1.0M 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 YTD YTD YTD YTD YTD YTD YTD

2026 年初至今，私募股权的中位交易规模最大，为 6,000 万美元；资产／投资管理 2,900 万，企业风投 2,500 万，企业投资 1,680 万，风险投资 1,480 万，天使 400 万。

Annual median deal size by investor group Angel Asset/Investment CVC Corp Private Equity VC Management $60.0M $50.0M $29.0M $26.0M $25.0M $25.0M $16.8M $16.1M $15.0M $14.0M $14.8M $8.9M $10.0M $8.3M $8.1M $5.8M $4.0M $4.0M 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 2022 2024 2026 YTD YTD YTD YTD YTD YTD

2026 年三季度大额融资 879 亿美元、154 笔；二季度为 1,354 亿、153 笔；一季度为 2,185 亿、149 笔。

Quarterly funding & deals from mega-rounds (deals worth $100m+) $250.0B Deals 149 153 $200.0B $150.0B 86 87 Funding $100.0B $87.9B 49 51 45 47 $50.0B $9.9B $10.5B $9.2B $8.4B $24.4B $20.6B $16.1B $12.5B $15.1B $22.8B $12.7B $48.0B $38.4B $41.6B $45.3B $74.0B $218.5B $135.4B $0.0B Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

美国 534 亿美元、88 笔；亚洲 163 亿美元、36 笔；欧洲 159 亿美元、27 笔；大洋洲 21 亿美元、2 笔；加拿大 2 亿美元、1 笔。

Mega-round funding & deals by global region in Q3’26 Canada $0.2B 1 Deal Asia $16.3B 36 Deals Europe $15.9B 27 Deals US $53.4B 88 Deals Oceania $2.1B 2 Deals

2026 年三季度，美国 88 笔，亚洲 36 笔，欧洲 27 笔，大洋洲 2 笔，加拿大 1 笔。美国的峰值出现在 2026 年一季度（89 笔）。

Quarterly mega-round deals by global region 86 US, 88 Asia, 36 20 22 23 Europe, 27 Oceania, 2 3 Canada, 1 1 1 1 LatAm, 0 2 1 All Other Regions, 0 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

2026 年三季度大额融资占全部融资的 85%，二季度为 88%，一季度为 91%。2022 年一季度这一比例为 37%。

Quarterly mega-rounds as percent of funding 9% 12% Non-mega-rounds, 15% 21% 24% 31% 30% 29% 39% 44% 46% 50% 50% 50% 53% 58% 55% 63% 62% 91% 88% Mega-rounds, 85% 79% 76% 69% 70% 71% 61% 56% 54% 50% 50% 50% 47% 42% 45% 37% 38% Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

2026 年投入企业的资金中，94% 属于 2.5 亿美元以上的轮次，2022 年这一比例为 10%。（注：2015 年的尖峰来自单笔交易，即阿里巴巴对阿里云的 10 亿美元投资。）

94% of all dollars invested into companies in 2026 were part of $250M+ rounds, up from 10% in 2022...

2026 年最大的 15 轮 AI 风险融资中，生成式 AI 拿到每 6 美元中的 5 美元（截至 9 月 16 日）：OpenAI 1,220 亿、Anthropic H 轮 650 亿、Anthropic G 轮 300 亿、xAI 200 亿、Waymo 168 亿、Prometheus 120 亿、DeepSeek B 轮 88 亿、DeepSeek 早期轮 74 亿、VNET 55 亿、Safe Superintelligence 55 亿、Databricks 55 亿、Anduril 55 亿、DayOne 45 亿、Crusoe 39 亿、月之暗面 35 亿美元。按公司总部统计，美国公司拿走约每 4 美元中的 3 美元。

2026 年初至今，早期交易占 73%，中期 11%，后期 7%，其他 9%。早期占比自 2023 年的 78% 逐年下降。

Annual percent of deals by deal stage 8% 9% 9% 9% Other, 9% 6% 5% 6% Late-stage, 7% 11% 8% 8% 9% Mid-stage, 11% 75% 78% 78% 76% Early-stage, 73% 2022 2023 2024 2025 2026 YTD

早期交易中美国占 49%、亚洲 23%、欧洲 22%；中期美国 47%、亚洲 36%、欧洲 16%；后期美国 47%、亚洲 32%、欧洲 17%；其他类美国 64%。

Regional deal share by deal stage in Q3’26 US Asia Europe LatAm Canada Africa Oceania All Other Regions Early-stage 49% 23% 22% Mid-stage 47% 36% 16% Late-stage 47% 32% 17% Other 64% 12% 22%

2026 年初至今，早期中位 300 万美元，中期 4,500 万美元，后期 8,250 万美元；2022 年分别为 200 万、2,000 万和 3,000 万美元。

Annual median funding by deal stage Early-stage Mid-stage Late-stage $82.5M $50.0M $44.2M $45.0M $40.0M $32.3M $30.5M $30.0M $25.0M $20.0M $2.0M $1.5M $1.6M $2.5M $3.0M 2022 2024 2026 2022 2024 2026 2022 2024 2026 YTD YTD YTD

OpenAI 100 亿美元（7 月 1 日，软银愿景基金，美国）；Databricks 50 亿美元 L 轮（8 月 13 日，估值 1,900 亿美元）；Safe Superintelligence 50 亿美元（7 月 27 日，英伟达，估值 320 亿美元）；Crusoe 39 亿美元 F 轮（9 月 3 日，估值 309 亿）；Mistral 35 亿美元 D 轮（9 月 8 日，估值 244 亿，法国）；月之暗面 35 亿美元 E 轮（7 月 29 日，估值 350 亿，中国）；可灵 AI 28 亿美元（7 月 2 日，估值 180 亿，新加坡）；Nscale 24 亿美元可转债（英国）；Cognition 20 亿美元 E 轮（估值 480 亿）；Firmus 20 亿美元（澳大利亚，估值 105 亿）。

Global: Top equity deals in Q3’26 Company Round Amount Round Date Round Valuation Select Investors Country 1 OpenAI $10.0B Venture Capital 2026-07-01 N/A SoftBank Vision Fund United States 2 Databricks $5.0B Series L 2 Safe Superintelligence $5.0B Corporate Minority 2026-07-27 $32.0B NVIDIA United States 4 Crusoe $3.9B Series F 2026-09-03 $30.9B Atreides Management, Mubadala Capital, Valor Equity Partners, Founders Fund, Salesforce Ventures United States 5 Mistral $3.5B Series D 5 Moonshot AI $3.5B Series E 2026-07-29 $35.0B Guozhitou Private Equity Fund Management, PyraTzLabs, Techmind China 7 Kling AI $2.8B Venture Capital 2026-07-02 $18.0B BlueFive Capital, CAS Investment Management, CITIC Securities, CPE, General Atlantic Singapore 8 Nscale $2.4B Convertible Note 2026-09-25 N/A Third Point, 8090 Industries, Citadel, Abu Dhabi Investment Council, Apollo United Kingdom 9 Cognition $2.0B Series E 9 Firmus $2.0B Venture Capital 2026-08-07 $10.5B Blackstone, Coatue, Maas Group, NVIDIA, Jane Street Group Australia

Walden Robotics 3 亿美元（美国，投资方含波音）；Naive.ai 1.8 亿和 1.2 亿美元两轮（中国，红杉中国、IDG、腾讯等）；Morphic Robot 1.47 亿美元（中国，阿里巴巴、腾讯）；Callosum 1 亿美元（英国）；DISCOVER Robotics 两轮各 1 亿美元（中国，联想创投等）；Veeda AI 9,000 万美元（加拿大）；Kinetix AI 7,500 万美元（中国）；KNOWIN 7,400 万美元（中国）。前十名中有六家是中国公司，且多为机器人方向。

Global: Top seed/angel deals in Q3’26 Company Round Amount Round Date Select Investors Country 1 Walden Robotics $300M Seed VC Deviation Capital, AE Ventures, Boeing, Calibrate Ventures, Colle Capital Partners United States 2026-07-15 2 Naive.ai $180M Seed VC China ICV Investment, HongShan, Huakong Electronics Private Equity Fund Management, IDG Capital, Legend Capital China 2026-07-23 3 Morphic Robot $147M Seed VC 2026-07-06 Alibaba.com, Tencent, Broadvision Capital, CAS Star, Gaorong Capital China 4 Naive.ai $120M Seed VC 2026-09-08 China ICV Investment, HongShan, MPC, Tencent, GF Qianhe China 5 Callosum $100M Seed VC 2026-08-20 Atomico, Plural, AlbionVC, DCVC, Dunamu & Partners United Kingdom 5 DISCOVER Robotics $100M Seed VC 2026-08-03 Binhu District Industrial Group, Fortune Capital, IDG Capital, Joyoung Co. China 5 DISCOVER Robotics $100M Seed VC Ambrum Capital, Cedarlake Capital, Glacier Capital, Legend Capital, Lenovo Capital and Incubator Group China 2026-07-01 8 Veeda AI $90M Seed VC 2026-08-19 Khosla Ventures, Radical Ventures Canada 9 Kinetix AI $75M Seed VC 2026-09-11 F&G Venture, Vertex Ventures China, Wanshi Capital China 10 KNOWIN $74M Seed VC MPC, Lighthouse Capital, Redpoint Ventures China, SenseTime Guoxiang Capital, Walden International China 2026-08-10

蚂蚁国际 12 亿美元（新加坡，蚂蚁集团、阿里巴巴）；River AI 11 亿美元（美国，英伟达、AMD Ventures、淡马锡）；小鹏机器人 9 亿美元（中国，IDG、阿里巴巴、腾讯、小鹏）；Groq 3.5 亿美元（美国，英伟达）；UniPat AI 3 亿美元（中国）；Sunrise 2.98 亿美元（中国）；Starcloud 2.5 亿美元（美国）；Euclyd 2.31 亿美元（荷兰）；Gravis Robotics 2 亿美元（瑞士，软银）；Humanoid 1.52 亿美元（英国，博世、舍弗勒）。

Global: Top Series A deals in Q3’26 Company Round Amount Round Date Select Investors Country 1 Ant International $1.2B Series A 2026-07-20 Ant Group, Alibaba.com Singapore 2 River AI $1.1B Series A 2026-08-11 AMP, General Catalyst, AMD Ventures, NVIDIA, Temasek United States 3 XPeng Robotics $900M Series A 2026-08-24 IDG Capital, Alibaba.com, Gaorong Capital, Tencent, Xpeng China 4 Groq $350M Series A 2026-08-17 Disruptive, NVIDIA United States 5 UniPat AI $300M Series A 2026-09-10 Alibaba.com, HongShan, Tencent China 6 Sunrise $298M Series A 37 Interactive Entertainment, Addor Capital, Andon Health, CAS Star, CCB Principal Capital Management China 2026-08-28 7 Starcloud $250M Series A 2026-08-21 Manhattan West Asset Management, Benchmark, NFX, Soma Capital United States 8 Euclyd $231M Series A EQT, Innovation Industries, Samsung Ventures, Somerset Capital Partners, Brabant Development Agency Netherlands 2026-09-15 9 Gravis Robotics $200M Series A 2026-08-17 SoftBank Group Switzerland 10 Humanoid $152M Series A Prime Movers Lab, Aglae Ventures, Bosch, Fubon Venture Capital, Schaeffler Group United Kingdom 2026-07-17

Fluidstack 15 亿美元（英国，Jane Street、谷歌）；生数科技 5 亿美元（中国）；Tripo AI 4.46 亿美元（中国）；Higgsfield 4 亿美元（美国，DST Global、Accel）；Olix 3.12 亿美元（英国，Arm）；Mariana Minerals 3.1 亿美元（美国，a16z、必和必拓）；Gimlet 3 亿美元；Wispr Flow 2.8 亿美元；Saviynt 2.55 亿美元；Instinct 2.5 亿美元（美国，Benchmark、Index）。

Global: Top Series B deals in Q3’26 Company Round Amount Round Date Select Investors Country 1 Fluidstack $1.5B Series B 2026-09-03 Jane Street Group, Situational Awareness, Google United Kingdom 2 Shengshu Technology $500M Series B 2026-07-06 Undisclosed Investors China 3 Tripo AI $446M Series B 2026-09-01 MPC, 4399, Fortune Capital, INCE Capital, T-Capital China 4 Higgsfield $400M Series B DST Global, Accel, Alpha Intelligence Capital, BroadLight Capital, GFT Ventures United States 2026-08-17 5 Olix $312M Series B Fundomo, Hummingbird Ventures, Arm, Crane Venture Partners, Creandum United Kingdom 2026-08-03 6 Mariana Minerals $310M Series B 2026-08-03 7 Gimlet $300M Series B 2026-09-04 8 Wispr Flow $280M Series B 2026-08-17 Menlo Ventures, NEA, Notable Capital, 8VC, MVP Ventures United States 9 Saviynt $255M Series B 2026-08-31 Carrick Capital Partners United States 10 Instinct $250M Series B 2026-08-26 Benchmark, Index Ventures United States

Instinct 10 亿美元（9 月 28 日，Benchmark、Coatue、红杉，同季度内第二轮）；Positron 8.75 亿美元；Together AI 8 亿美元；Neko Health 7 亿美元（瑞典）；Mach Industries 6 亿美元；Multiverse Computing 5.68 亿美元（西班牙）；Wonderful 5.5 亿美元（荷兰）；地平线机器人 4 亿美元（中国）；Lovable 4 亿美元（瑞典）；Etched 3 亿美元（美国，a16z、Karpathy、蒂尔）。

Global: Top Series C deals in Q3’26 Company Round Amount Round Date Select Investors Country 1 Instinct $1.0B Series C 2026-09-28 Benchmark, Coatue, Sequoia Capital United States 2 Positron $875M Series C 2026-09-10 Andra Capital, Atreides Management, Jim Clark, NEA, SemiAnalysis United States 3 Together AI $800M Series C Aramco Ventures, General Catalyst, Salesforce Ventures, Emergence Capital, Lux Capital United States 2026-07-01 4 Neko Health $700M Series C 2026-07-15 Lightspeed, O.G. Venture Partners, General Catalyst, Atomico, Lakestar Sweden 5 Mach Industries $600M Series C 2026-09-10 Bedrock Capital, Infinite Capital, Ribbit Capital United States 6 Multiverse Computing $568M Series C BNP Paribas Asset Management, Bullhound Capital, Forgepoint Capital International, SPRI, Sociedad Espanola para la Transformacion Tecnologica Spain 2026-07-28 7 Wonderful $550M Series C Insight Partners, Bessemer Venture Partners, Index Ventures, Institutional Venture Partners, Salesforce Ventures Netherlands 2026-09-02 8 D-Robotics $400M Series C MAPS Capital, 5Y Capital, AlphaX Partners, Capitallink Group, Chuxin Capital Partners China 2026-09-17 8 Lovable $400M Series C 2026-08-12 EQT, Menlo Ventures, Accel, Salesforce Ventures, Antler Sweden Sweden 10 Etched $300M Series C 10 Upwind Security $300M Series C Bessemer Venture Partners, TCV, Greylock Partners, Salesforce Ventures, Alta Park Capital United States 2026-09-02

Mistral 35 亿美元（法国，估值 244 亿）；Fireworks 15 亿美元（估值 175 亿）；Hadrian 14 亿美元（估值 79 亿）；Quantum Systems 12 亿美元（德国，估值 80 亿，空客、黑石）；Etched 7 亿美元（估值 210 亿）；Tekever 5.8 亿美元（葡萄牙）；Exein 2.7 亿美元（意大利）；Owner 2.4 亿美元；Gatik 2 亿美元；Profound 1.8 亿美元。

Global: Top Series D deals in Q3’26 Company Round Amount Round Date Round Valuation Select Investors Country 1 Mistral $3.5B Series D 2 Fireworks $1.5B Series D 2026-07-16 $17.5B Atreides Management, Index Ventures, TCV, Bessemer Venture Partners, Lightspeed United States 3 Hadrian $1.4B Series D 2026-08-06 $7.9B 137 Ventures, Baillie Gifford, JPMorganChase, Valor Equity Partners, WCM Investment Management United States 4 Quantum Systems $1.2B Series D 2026-07-01 $8.0B Advent International, Airbus, Blackstone, Noteus Partners, Balderton Capital Germany 5 Etched $700M Series D Sequoia Capital United States 6 Tekever $580M Series D 2026-09-23 $6.4B Baillie Gifford, UC Investments, Crescent Cove Advisors, Iberis Capital Portugal 7 Exein $270M Series D 2026-09-15 $1.7B Headline, 33N Ventures, Balderton Capital, Blue Cloud Ventures, Geodesic Capital Italy 8 Owner $240M Series D 2026-08-28 $2.3B Goldman Sachs Alternatives, Headline, Meritech Capital Partners, Redpoint Ventures, Jack Altman United States 9 Gatik $200M Series D 2026-08-25 N/A Koch Disruptive Technologies, QIA, Alumni Ventures, ARK Invest United States 10 Profound $180M Series D 2026-09-15 $1.8B Kleiner Perkins, Lightspeed, Evantic Capital, Khosla Ventures, Saga United States

Databricks 50 亿美元 L 轮（估值 1,900 亿）；Crusoe 39 亿美元 F 轮（估值 309 亿）；月之暗面 35 亿美元 E 轮（估值 350 亿）；Cognition 20 亿美元 E 轮（估值 480 亿）；Helsing 18 亿美元 E 轮（德国，估值 180 亿）；Motive 13 亿美元 H 轮；SambaNova 10 亿美元 F 轮（估值 110 亿）；Harvey 5.5 亿美元 H 轮（估值 156 亿）；Temporal 5.5 亿美元 E 轮（估值 126 亿）；Cyera 和 Island 各 4 亿美元。

Global: Top Series E+ deals in Q3’26 Company Round Amount Round Date Round Valuation Select Investors Country 1 Databricks $5.0B Series L 2 Crusoe $3.9B Series F 2026-09-03 $30.9B Atreides Management, Mubadala Capital, Valor Equity Partners, Founders Fund, Salesforce Ventures United States 3 Moonshot AI $3.5B Series E 2026-07-29 $35.0B Guozhitou Private Equity Fund Management, PyraTzLabs, Techmind China 4 Cognition $2.0B Series E 5 Helsing $1.8B Series E 2026-07-13 $18.0B Dragoneer, Lightspeed, General Catalyst, Plural, CPP Investments Germany 6 Motive $1.3B Series H 2026-09-10 N/A General Catalyst United States 7 SambaNova $1.0B Series F 2026-07-08 $11.0B General Atlantic, Battery Ventures, A&E investments, Assam Ventures, BlackRock United States 8 Harvey $550M Series H Perkins, Sapphire Ventures United States 8 Temporal $550M Series E 2026-09-14 $12.6B Goldman Sachs Alternatives, Lightspeed, Tiger Global Management, 10 Cyera $400M Series G 2026-09-22 $12.0B Evolution Equity Partners, Goldman Sachs Alternatives United States 10 Island $400M Series F 2026-09-24 $6.4B Evolution Equity Partners, Sequoia Capital, Coatue, CyberStarts, Georgian United States

在图中展示的公司里，估值翻倍所需的时间从 3.5 个月到 13.3 个月不等。这些历史拟合描述的是融资时的估值标记，并不是估值必须继续遵循的规律。

Across the companies shown, estimated valuation-doubling times range from 3.5 to 13.3 months. These historical fits describe fundraising marks, not a law that valuations must continue to follow.

最新的估值标记从 Anthropic 的 15 倍到 xAI 的 250 倍不等，OpenAI 为 21 倍，Cohere 为 83 倍。公布的和估算的收入数据混用，因此该图是市场定价的一个快照。

The latest marks shown range from 15x for Anthropic to 250x for xAI, with OpenAI at 21x and Cohere at 83x. Reported and estimated revenue figures are mixed, so the chart is a snapshot of market pricing rather than a like-for-like valuation rule.

亚马逊、Alphabet、微软和 Meta 的 2026 年资本开支指引合计 7,330 亿美元，比 2025 年的 4,100 亿增长 79%（亚马逊 +71%、Alphabet +119%、微软 +48%、Meta +90%）。OpenAI 和 Anthropic 在 2026 年分别宣布了 1,220 亿和 950 亿美元的融资。

Amazon, Alphabet, Microsoft and Meta guide to $733B in 2026 capex, up 79% from $410B in 2025. OpenAI and Anthropic have announced $122B and $95B of funding, respectively, in 2026.

Dealroom 追踪了参与融资轮次的投资方。2026 年，中东与北非投资者参与的轮次占全部 AI 融资金额的一半。这里计的是这些轮次的全部金额，而不是这些投资者实际出资的部分。

Dealroom tracks who participates in financing rounds. MENA investors took part in rounds representing half of AI funding dollars in 2026. This counts the full value of those rounds, not the amount of capital supplied by Gulf investors.

价值创造正在发生在私营市场一侧。

The Most Valuable Private Cos Are Larger Than Ever . . . And value creation is happening on the private side.

公司在私营阶段停留更久：跳过小盘股阶段，直接成长为超大型公司。

Private-For-Longer: Skip Small Cap, Go Straight to Mega

2026 年三季度新增独角兽 42 家，累计 706 家；此前四个季度分别新增 25、25、35、38 家，累计数从 599 升至 671 家。

Quarterly new & total unicorns (private companies valued at $1b+) 671 Total Unicorns, 706 599 614 25 25 35 38 New Unicorns, 42 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

美国累计 470 家，三季度新增 24 家；亚洲累计 110 家，新增 11 家；欧洲累计 105 家，新增 7 家；加拿大 11 家、大洋洲 6 家、拉美 4 家，三季度均无新增。

New & total unicorns by global region in Q3’26 Canada 11 0 New Asia 110 11 New Europe 105 7 New US 470 24 New LatAm 4 0 New Oceania 6 0 New

2026 年三季度，美国新增 24 家，亚洲 11 家，欧洲 7 家，其余地区为零。美国的季度峰值为 24 家。

Quarterly new unicorns by global region 23 US, 24 10 Asia, 11 Europe, 7 1 LatAm, 0 Canada, 0 Oceania, 0 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

小鹏机器人 63 亿美元（中国）；River AI 50 亿美元（美国）；micro1 40 亿美元（美国）；Olix 33 亿美元（英国）；Sharpa 33 亿美元（新加坡）；AfterQuery 32 亿美元（美国）；Celero Communications 30 亿美元（美国）；面壁智能 30 亿美元（中国）；Angle Health 27 亿美元（美国）；UniPat AI 25 亿美元（中国）。

Global: Top unicorn births in Q3’26 Company Latest Valuation Country 1 XPeng Robotics $6.3B China 2 River AI $5.0B United States 3 micro1 $4.0B United States 4 Olix $3.3B United Kingdom 4 Sharpa $3.3B Singapore 6 AfterQuery $3.2B United States 7 Celero Communications $3.0B United States 7 ModelBest $3.0B China 9 Angle Health $2.7B United States 10 UniPat AI $2.5B China

Anthropic 1.2 万亿美元；OpenAI 8,520 亿美元；字节跳动 4,800 亿美元（中国）；Databricks 1,900 亿美元；Anduril 610 亿美元；DeepSeek 592 亿美元（中国）；Cognition 480 亿美元；Thinking Machines Lab 450 亿美元；Ramp 440 亿美元；Canva 420 亿美元（澳大利亚）。

Global: Top unicorns by valuation in Q3’26 Company Latest Valuation Country 1 Anthropic $1,200.0B United States 2 OpenAI $852.0B United States 3 ByteDance $480.0B China 4 Databricks $190.0B United States 5 Anduril $61.0B United States 6 DeepSeek $59.2B China 7 Cognition $48.0B United States 8 Thinking Machines Lab $45.0B United States 9 Ramp $44.0B United States 10 Canva $42.0B Australia

这一轮周期与以往不同。集中是新周期的主题：更少的基金以更大的规模投向更少的创业公司，其中越来越多位于硅谷。不过从某种意义上说，这种集中有点误导：在表面之下，获得融资的公司类型发生了结构性变化。

Everything Is AI Computerse This cycle is not like the previous ones Concentration has been a theme of the new cycle. A smaller number of funds are deploying at larger sizes into fewer startups in, an increasing share of which are located in Silicon Valley, and nearly all of them are “AI-related.” There is one sense, however, in which all that concentration is a bit misleading. Beneath the surface, there has been a mix-shift in the kinds of companies that are being funded. Venture dollars are flowing to enterprise software and consumer apps, as they always have, but also traditionally out-of-scope services (e.g. legal, accounting) and formally less-popular capital-intensive verticals, like semis, power, and defense.

更大、更快、更年轻。

A New Breed Of Unicorn Bigger, faster, younger.

AI 这一代公司的构建方式与以往不同。

The AI Generation Is Built Differently?

2026 年三季度并购 372 笔、IPO 19 起、SPAC 2 起，合计 393 笔；此前四个季度并购分别为 423、443、454、456 笔。

Quarterly exits 454 456 M&A, 372 30 34 34 28 IPO, 19 4 4 5 3 SPAC, 2 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

2026 年三季度，美国占退出数的 52%，欧洲 32%，亚洲 10%，拉美和加拿大各 2%，大洋洲 1%。

Percent of quarterly exits by global region LatAm, 2% Canada, 2% Africa, 0% Oceania, 1% All Other Regions, 0% 30% 29% 33% 35% 32% 30% 30% 33% 25% Europe, 32% 35% 34% 37% 31% 36% 33% 39% 35% 43% 12% 14% Asia, 10% 14% 12% 21% 13% 14% 14% 12% 17% 15% 16% 13% 18% 12% 15% 17% 15% 52% 50% US, 52% 47% 49% 50% 42% 43% 46% 39% 42% 45% 45% 41% 44% 43% 40% 37% 34% Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 2022 2023 2024 2025 2026

AI 退出数量有望比 2025 年高出约五分之一：8.5 个月内 890 笔，年化约 1,250 笔。退出价值方面，IPO 窗口已重新开启，并购随之跟上（2012 年的尖峰是 Meta 的 1,040 亿美元 IPO）。

这是一次历史级规模的 IPO。

An Historically Large IPO

Anysphere 以 600 亿美元被 SpaceX 收购；OpenRouter 以 80 亿美元被 Stripe 收购；Altruist 46 亿美元被先锋领航收购；Modular 39 亿美元被高通收购；Fin 36 亿美元被 Salesforce 收购；MaintainX 36 亿美元被 Autodesk 收购；Airtable 22 亿美元被 Bending Spoons 收购；Manus 20 亿美元（新加坡，红杉中国、腾讯等）；Anyscale 16 亿美元被 Nscale 收购；Mechanize 15 亿美元被 Google DeepMind 收购。

Global: Top M&A exits in Q3’26 Company Round Valuation Acquirer Country 1 Anysphere $60.0B SpaceX United States 2 OpenRouter $8.0B Stripe United States 3 Altruist $4.6B Vanguard United States 4 Modular $3.9B Qualcomm United States 5 Fin $3.6B Salesforce United States 5 MaintainX $3.6B Autodesk United States 7 Airtable $2.2B Bending Spoons United States 8 Manus $2.0B HongShan, Tencent, Undisclosed Angel Investors, ZhenFund Singapore 9 Anyscale $1.6B Nscale United States 10 Mechanize $1.5B Google DeepMind United States

燧原科技 91 亿美元（中国）；宇树科技 90 亿美元（中国）；Momenta 89 亿美元（中国）；XTEND 15 亿美元（以色列）；ESDS Software Solution 5.32 亿美元（印度）；Potentially 1,400 万美元（英国）；Aegeus Technologies 900 万美元（印度）。

Global: Top IPOs in Q3’26 Company Round Valuation Country 1 Enflame $9.1B China 2 Unitree Robotics $9.0B China 3 Momenta $8.9B China 4 XTEND $1.5B Israel 5 ESDS Software Solution $532M India 6 Potentially $14M United Kingdom 7 Aegeus Technologies $9M India

按不变的资本结构计算，股价涨幅意味着上市以来约 5,480 亿美元的企业价值增量，其中 87% 来自 DRAM 公司长鑫存储（CXMT）。这批上市公司涵盖模型开发商、GPU 设计公司和长鑫的 DRAM 业务。美光和 SK 海力士还销售 NAND 存储和高带宽内存，产品结构不同。美国实验室用的是较早的私募轮估值和年化收入，图中灰色柱提供参照，但与公开市场倍数并非同口径。

Share-price gains imply ~$548B of enterprise-value uplift since IPO at constant capital structures, 87% coming from gains posted by DRAM company, CXMT. At September 18 prices, the five names trade at 19-189x trailing revenue, the highest multiple being held by Z.ai • The IPO cohort spans model developers, GPU designers and CXMT’s DRAM business. • Micron and SK hynix also sell NAND storage and high-bandwidth memory, so their product mix differs. • US labs use older private-round equity values and annualized sales. These gray bars give context but are not like-for-like public multiples.

29 笔交易中，最新的一批只买人（按收购方实际取得的内容划分：只要人、技术授权加团队、资产加团队、持股加创始人）：2024 年 7 笔、2025 年 12 笔、2026 年至 9 月 16 日 10 笔。OpenAI 做了其中约四分之一，其余收购方包括苹果、亚马逊、微软、谷歌、Salesforce 和英伟达。

General Catalyst 30 家（风投，美国）；Andreessen Horowitz 29 家；Insight Partners 26 家（私募股权）；Khosla Ventures 26 家；Lightspeed 22 家；Bessemer 16 家；红杉中国 16 家（中国）；Salesforce Ventures 16 家（企业风投）；Accel、Index Ventures、Menlo Ventures、英伟达各 15 家。

Global: Top investors by company count in Q3’26 Investor Company Count Investor Group Country 1 General Catalyst 30 VC United States 3 Insight Partners 26 Private Equity United States 3 Khosla Ventures 26 VC United States 5 Lightspeed 22 VC United States 6 Bessemer Venture Partners 16 VC United States 6 HongShan 16 VC China 6 Salesforce Ventures 16 CVC United States 9 Accel 15 VC United States 9 Index Ventures 15 VC United Kingdom 9 Menlo Ventures 15 VC United States 9 NVIDIA 15 Corp United States

General Catalyst 30 家、a16z 29 家、Khosla Ventures 26 家、Lightspeed 22 家；Bessemer 和红杉中国各 16 家；Accel、Index Ventures、Menlo Ventures 各 15 家；Alumni Ventures 和红杉资本各 14 家。

Global: Top VCs by company count in Q3’26 Investor Company Count Country 1 General Catalyst 30 United States 3 Khosla Ventures 26 United States 4 Lightspeed 22 United States 5 Bessemer Venture Partners 16 United States 5 HongShan 16 China 7 Accel 15 United States 7 Index Ventures 15 United Kingdom 7 Menlo Ventures 15 United States 10 Alumni Ventures 14 United States 10 Sequoia Capital 14 United States

Salesforce Ventures 16 家；英伟达旗下 NVentures 10 家；谷歌 GV 8 家；沙特 Prosperity7 Ventures 8 家；沙特阿美创投、微软 M12、三星创投各 6 家；AMD Ventures、日立创投、英特尔投资各 5 家。

Global: Top corporate VCs (CVCs) by company count in Q3’26 Investor Company Count Country 1 Salesforce Ventures 16 United States 2 NVentures 10 United States 3 GV 8 United States 3 Prosperity7 Ventures 8 Saudi Arabia 5 Aramco Ventures 6 Saudi Arabia 5 M12 6 United States 5 Samsung Ventures 6 South Korea 8 AMD Ventures 5 United States 8 Hitachi Ventures 5 Germany 8 Intel Capital 5 United States

幂律依然成立，而且比以往更极端。

Power Law Abides, But Even More So

幂律同样适用于基金管理人。

Power Law Applies To Managers, As Well

对风险投资基金而言，前 10% 与其余基金之间的差距从未像现在这样大。

For VC Funds, The Difference Between Top Decile And The Rest Has Never Been Wider

折价交易很少见，最好的公司在二级市场上以溢价成交。

Secondary Markets Tell The Same Story Discounts are few and far between, and the best companies trade a premium

和上市公司类似，零利率时代结束后，创业公司用增长换取了盈利能力。不过，募集新一轮股权通常仍需要高速增长：近期成功融资的公司，增速仍与零利率高峰期持平或更快。

Adjusting To Life After ZIRP Not unlike Public Cos, when ZIRP ended, startups traded growth for profitability. That said, raising fresh equity tends to require rapid growth. The companies who have successfully raised recently are still growing at a pace in-line (or faster) with peak-ZIRP.

2026 年三季度美国 AI 融资 630 亿美元，二季度为 1,152 亿，一季度为 2,135 亿。页面覆盖季度融资额与交易数、各阶段年度交易占比、三季度最大股权交易、季度退出数等。

US Trends

硅谷三季度融资 411 亿美元，约占全美 630 亿美元的 65%；二季度为 953 亿，一季度为 1,934 亿。其余十个城市各有同一套统计口径的页面（季度融资与交易、分阶段占比、最大交易、退出）。

Silicon Valley

2026 年三季度加拿大 AI 融资 4 亿美元、36 笔交易，二季度为 3 亿美元。

Canada Trends

2026 年三季度亚洲 AI 融资 188 亿美元、538 笔交易，二季度为 217 亿，一季度为 108 亿。

Asia Trends

2026 年三季度中国 AI 融资 97 亿美元，二季度为 184 亿，一季度为 76 亿；2025 年各季度在 12 亿至 36 亿美元之间。三季度规模约为亚洲总量的一半。

China

印度三季度 6 亿美元（二季度 7 亿）；新加坡 60.9 亿美元（二季度 7.98 亿，三季度因可灵 AI 等大额交易大幅跃升）；日本 3.92 亿美元（二季度 2.25 亿）。

India

2026 年三季度欧洲 AI 融资 188 亿美元、458 笔交易，二季度为 149 亿，一季度为 128 亿；规模与亚洲相当。

Europe Trends

英国三季度 58 亿美元，二季度为 59 亿，是欧洲最大的单一市场。其余五国各有同一套统计口径的页面。

United Kingdom

2026 年三季度拉美 AI 融资 8,800 万美元、28 笔交易，二季度为 2.77 亿美元。

LatAm Trends

2026 年三季度大洋洲 AI 融资 22.3 亿美元、30 笔交易，二季度为 7.73 亿美元，三季度因 Firmus 等大额交易跃升。

Oceania Trends

云厂商暴涨的利润为大部分资本开支提供了资金，其代价是自由现金流急剧下降，预计这一状况将持续到 2028 年前后。

Capex Is Eating Hyperscaler Cashflows (For Now) Surging profits for hyperscalers have funded the lion’s share of capex, the flipside of which has been a steep drop in FCF that is expected to persist until ~2028.

云厂商的现金已被充分动用，信贷市场随之活跃起来，借钱给美国政府不再是唯一的选择。

Debt Markets Step In (On and Off Balance Sheet) With hyperscaler cash fully-tapped, credit markets have sprung to life. Lending to Uncle Sam is no longer the only game in town.

债务确实提高了风险，数字也很大，但最终重要的是投资回报率。目前云厂商的投入资本回报率远高于其借贷成本。这些公司历史上一直高度盈利，直到不久前几乎不借钱；它们借钱是为了投资能产生稳定（且在上升）现金流的实体资产，这是放贷方熟悉且青睐的组合。

Hyperscaler ROE Still Higher Than Their Cost of Borrowing Debt certainly raises the stakes, and the numbers are big, but what ultimately matters is the return on investment. For now, hyperscaler ROIC is well-above their cost of borrowing. These are (again) historically profitable companies, which until recently, borrowed very little. They are borrowing to invest in hard assets that are generating steady (and rising) cashflows. It’s a combination that lenders know and love.

英伟达和博通扩大了对由外部投资者出资的基础设施的担保。客户违约会让它们同时损失销售和担保赔付，不过英伟达认为独立的信用审核和硬件可再利用降低了这一风险。另一方面，云厂商在发行股权：Alphabet 6 月净募集 496 亿美元，其中 100 亿来自伯克希尔，用于包括 AI 基础设施和算力在内的公司用途。英伟达／OpenAI 和 SB Energy：对指定的租赁和电力付款提供最高 1,050 亿美元担保，覆盖追偿后的指定违约缺口，自租约开始时生效。博通／阿波罗和黑石：350 亿美元融资，博通的租赁兜底敞口最高 290 亿美元，在从 AI 机架回收价值后覆盖部分违约损失。英伟达／六家金融机构：融资目标超过 5,000 亿美元，对选定交易提供最高 25% 的残值支持，可能承担基础设施未来资产价值的缺口。

NVIDIA and Broadcom expanded guarantees for infrastructure funded by outside investors. Customer defaults could cost them both sales and guarantee payments, though NVIDIA argues independent underwriting and reusable hardware reduce this risk. Separately, hyperscalers are selling equity: Alphabet raised $49.6B net in June, including $10B from Berkshire, for corporate uses including AI infrastructure and compute.

残值担保向贷款方承诺数据中心或其芯片的最低转售价值。债务由一家持有资产的独立公司（SPV）承担，因此不计入担保方的资产负债表。摩根士丹利在 2025 年 10 月为 Meta 的 Hyperion 园区设计了这一结构，此后迅速流行。不到 12 个月内签出的四笔担保合计 1,750 亿美元：Meta 410 亿、博通 290 亿、英伟达 1,050 亿。有担保的债务定价只比担保方自身债券高 100–150 个基点，使 Hyperion 得以募集 270 亿美元。英伟达在 OpenAI 的俄亥俄租约 2028 年开始前不记录任何负债，而该园区将在 20 年内只运行英伟达硬件。

A residual value guarantee promises lenders a minimum resale value for a data center or its chips. The debt is held in a separate company that owns the asset (SPV), so it is off the balance sheet of the guarantor. Morgan Stanley designed the structure for Meta’s Hyperion campus in October 2025, and it has become popular since. • Four guarantees issued in under 12 months sum to $175B: Meta $41B, Broadcom $29B and NVIDIA $105B. • Guaranteed debt prices just 100-150bp above the guarantor’s own bonds, which let Hyperion raise a whopping $27B. • NVIDIA records no liability until OpenAI’s Ohio leases begin in 2028, and the campus will run only NVIDIA hardware for 20 years.

摩根士丹利统计，七家云厂商和芯片厂商合计有超过 3 万亿美元的表外承诺和信用支持。其中多数是尚未开业的数据中心的采购承诺和租赁。这些是已签约的未来付款而非借款，因此不计入已报告的债务。谷歌最多，达 8,900 亿美元，其中 7,070 亿是采购承诺；微软和甲骨文则以租赁为主。标普预测亚马逊 2027 年的调整后债务为 EBITDA 的 1.6 倍，高于其 1.5 倍的降级门槛，甲骨文为 4.4 倍，门槛为 4.5 倍。标普只在担保金额超过压力测试下的出售价值时，才把担保计入债务，而它称这一差额“多数时候接近于零”。

Morgan Stanley counts more than $3T of off-balance-sheet commitments and credit support across seven hyperscalers and chipmakers. Most are purchase commitments and leases on data centers that have not opened yet. These are contracted future payments rather than borrowings, so they sit outside reported debt. • Google carries the most at $890B, with $707B of it in purchase commitments, while Microsoft and Oracle are mostly leases. • S&P forecasts Amazon’s adjusted debt at 1.6x EBITDA in 2027, above its 1.5x downgrade threshold, and Oracle’s at 4.4x against 4.5x. • S&P adds a guarantee to debt only where it exceeds a stressed sale value, a gap it calls “closer to zero most of the time.”

科技已从“家庭作坊”变成全球经济的首要推动力。到了今天，每家公司在某种意义上都是科技公司，很难再找到纯粹靠纸笔运转的生意。科技比任何其他行业都更能推动投资、研发周期、盈利增长和利润率扩张；世界上最大、最赚钱、增长最快的公司都是科技公司，仅 AI 基础设施建设就贡献了相当大一部分。

Tech Is The Everything Cycle Tech has gone from a “cottage industry” to the prime mover in the global economy. At this point, every company is, in some sense, a tech company–you’d be hard-pressed to find a truly pen n’ paper business. More than any other sector, tech is driving investment, R&D cycles, earnings growth and margin expansion. The largest, most profitable, and fastest-growing companies the world has ever known, are tech companies. The AI infrastructure buildout alone is responsible for all or nearly all of net-new construction spending, blue collar job-growth, durable goods imports, net-new IG-issuance, etc. AI infra is arguably the only pro-cyclical impulse firing in the economy right now. The point is that tech has left the cottage. Tech is the Everything Cycle now, and it’s not looking back.

全球市值最大的公司都是科技公司。

The World’s Largest Companies Are Tech Companies

科技公司贡献了价值创造和盈利增长的大头，而且至少已经持续了十年。

Tech Drives Value Creation & Earnings Growth Tech Cos are driving the lion’s share of value creation and earnings growth, and have been for at least a decade.

利润率处于历史高位，几乎所有行业近期都有扩张，而科技最为突出。

Tech Drives Margin Expansion Margins are historically wide, and while nearly every sector has seen margin expansion recently, tech has been the standout.

科技（和能源）领跑盈利增长，但增长是广泛的：标普 500 中 94% 的公司达到或超过每股收益预期。

. . . Leading To (Very) Strong Fundamental Performance Tech (and Energy) are leading earnings growth, but growth is broad-based: 94% of S&P500 companies met or beat EPS estimates.

市场不断创新高，但推动它的是利润而不是价格。尽管利润在已经很高的基数上继续增长，估值倍数实际上在收缩，表现完全由盈利增长驱动。

Hard To Call It A “Bubble” The market keeps setting new ATHs, but it’s profits, not prices doing the work. Despite growing profits (off an already high base), multiples have actually compressed, while performance is driven entirely by earnings growth While Earnings Grow–off an already high base–multiples continue to compress

科技股年内上涨超过 20%，而估值倍数仍比五年均值低约 20%。

Multiple Compression Across the Board, But Especially Tech Tech is up by more than 20% on the year, even with multiples ~20% below their 5-year average

市场已经不再只是“七巨头”之类的故事。

The Market Is Not a Mag-Whatever Story, Anymore Last year, performance and returns was a concentrated story: the Mag 7 dominated the field. This year, the field has raced ahead, while the biggest tech companies have lagged the indexes. The most straightforward explanation for the shift is unpacked in the next section: hyperscalers are pouring their profits into the AI buildout. Hyperscaler spending has been driving semiconductor earnings to unprecedented heights, and market returns have followed the money.

散户仍是重要的市场推动力，杠杆和动量加剧了波动。

Momentum, Retail, and Leverage Retail continues to be an important mover-of-markets, where leverage and momentum contribute to increased volatility. Unlike the meme-stonk era, retail has piled into the consensus themes.

这不是第一轮资本开支周期：半导体曾经领先于软件。

Not The First Capex Rodeo: Semis Led Software Before

回到市场，AI 的供给与需求已深刻改变了推动市场的公司类型。

Follow The Money Returning to markets, all of this AI supply and demand, has had a profound impact on the types of businesses that are driving performance. For more than a decade, software was the star of the show, but with the advent of AI, hardware has taken center stage. The multi-trillion capex super-cycle has caused a surge in demand for, among other things, GPUs, HBM memory, power components, skilled trades, and even copper, that has far outstripped supply. The result has been historic earnings growth for traditionally sleepy, cyclical sectors, and stock prices have responded, accordingly. To be clear, it’s not only AI-infra driving the rotation from Bits to Atoms. EVs and solar power have also precipitated a massive grid glow-up. Defense spending is on the rise. Rockets and robots are both emerging new frontiers. Global politics and secular aging have lent urgency to the task of rebuilding an industrial base that will support core manufacturing, innovation, and automation. We’ve entered a new cycle of real assets, and while tech is running through all of it, capital-light bits is no longer the (only) name of the game. Another Age of Atoms has arrived.

资本开支流向 AI 基础设施。

Follow The Money: Capex → AI infra (cont.)

这场原子投资的大跃进自然反映在市场定价上。在资本轻型模式主导十年之后，资本密集型生意扭转了局面。并不是说软件这类资本轻型生意表现差（SaaS 末日另当别论），而是科技中最大、最赚钱的公司正在把资金投向实体资产。最终结果是，AI 供应链成为迄今最成功的交易：芯片厂商、光互连、燃气轮机运输、液冷，以及其间的一切。

The Rotation To Atoms Has Flipped The Field In Markets Unsurprisingly, this great leap forward in Atoms investment is showing up in market pricing. After a decade-long of capital-light dominance, capital-heavy businesses have flipped-the-field. It’s not that capital-light businesses (like software) have performed poorly (SaaSpocalypse notwithstanding), so much as tech’s largest and most profitable companies–and thereby the world’s largest and most profitable companies–are pouring those profits into hardware like never before. The net result is the AI Supply Chain has been the winning trade, thus far. That’s the chipmakers, the optical interconnectors, the turbine shippers, the liquid coolers, and everything in between.

没有比内存更大的赢家了。

The AI Supply Chain Has Been The Winning Trade . . . and there’s been no bigger winner than memory.

有意思的是，尽管市场表现亮眼，仍有观点认为它还不够亮眼：几家最大芯片厂商的远期盈利倍数，远低于高增长时期的水平。原因也很直接：需求高度依赖云厂商的资本开支，而经验丰富的周期股投资者总是警惕周期顶点。但资本开支目前没有任何放缓迹象，如果对智能的需求真的无限，这一轮周期可能与以往不同。

What If Semis Are Still Cheap? The funny thing is, though, that as stellar as market performance has been, there’s an argument that it hasn’t been stellar enough. Forward earnings multiples for some of the largest chipmakers, are well-below what the sky high expectations for earnings growth would normally predict. But capex shows no indication of a slowdown any time soon, and if demand for intelligence is indeed infinite, then this cycle may not be like the other cycles. There too the reason is fairly straightforward: demand is highly contingent on hyperscaler capex, and seasoned investors in cyclicals are always wary of the cycle’s peak. It’s also true that necessity is the mother of invention, and with great profit margins, one can expect great efficiency gains for power, compute and memory, right around the corner. If that’s the case, then future investors are going to have a hard time looking back on those multiples . . . between all the tears, that is.

无论有多少怀疑，在这轮建设进行约三年后，供给仍然落后于需求。GPU 似乎总能按时到货，但供应链的其他每一个环节交付周期都异常漫长；而且即便拿到了芯片，能不能通电仍是未知数。

The Semi Supply Chain Remains Backlogged Whatever skepticism may preside, ~3 years into the buildout, supply is still lagging demand. Somehow the GPUs keep arriving on time, but every other piece of the supply chain continues to have unusually long lead-times. And even if you can get your chips, it’s an open question if you can power them up.

与其说是末日，不如说 SaaS 进入了“证明自己”的时代。

More Like Software’s Selective Selloff, As SaaS Enters a ‘Prove-It’ Era SaaSpocalypse?

1 月，Anthropic 发布 Claude Cowork，把 Claude Code 包装成知识工作者可以处理自有文件和业务流程的应用。随着公司增加法律、市场和财务插件，“SaaS 末日”在 2 月初抹去了近 2,850 亿美元的软件市值，标普 500 软件与服务指数较 2025 年 10 月高点下跌 26%。对冲基金年内做空软件股 24 亿美元；2 月 20 日 Claude Code Security 发布后，恐慌蔓延到网络安全板块。XSW 指数从 4 月低点到 8 月高点上涨 55%，到 9 月 11 日回落至 +44%。

In January, Anthropic shipped Claude Cowork, an app that wrapped Claude Code to make it more usable by knowledge workers with their own files and business workflows. As the company added plugins for legal, marketing and finance, markets freaked out and investors called the end of seat-based pricing and repriced the terminal value of SaaS revenue. • XSW daily closing share price ($) • The “SaaSpocalypse” wiped out nearly $285B of software market capitalization in early February. The S&P 500 software and services index fell 26% from its October 2025 peak. • Hedge funds shorted $24B of software names in the year to date. The panic spread to cybersecurity after Claude Code Security launched on February 20. • XSW rose 55% from its April low to August’s peak, before pulling back to +44% by 11 September.

硬件表现强势的另一面是软件相对低迷。2026 年以“SaaS 末日”开场：曾经高歌猛进的上市 SaaS 公司遭遇看似无差别的抛售。但实际发生的，用“证明自己”来形容比“末日”更准确。SaaS 的估值（先是逐渐、然后突然）重新定价，反映了行业从增长转向盈利的整体变化。AI 确实引发了对护城河和终值的疑问，但影响并不均匀：被认为（或已证明）有护城河、或有 AI 顺风的公司表现更好。无论市场定价如何，SaaS 的基本面迄今相当稳定，收入增长和经营杠杆都没有明显恶化。

More Prove-It Than Pocalypse The flipside of hardware’s outperformance, has been software’s relative malaise. 2026 began with the SaaSpocalypse: a seemingly undifferentiated sell-off of formally high-flying public SaaScos, as AI prepared to eat their collective lunch. What’s transpired, however, is better described as “prove it,” than -pocalypse. SaaS has re-rated (gradually, and then suddenly) to reflect the industry’s broader shift from growth to profitability. As the share of high-growth, less-profitable companies has given way to slower-growth, more-profitable companies, multiples have compressed accordingly. At the same time, if you’re still a high-flyer, your multiples are still pretty, pretty good. Likewise, AI has certainly raised some doubts about defensibility and terminal value, but not uniformly. Companies with perceived (and/or demonstrated) defensibility and/or AI tailwinds, have outperformed. Cyber and observability has been a standout category, and vertical saas has experienced less compression than horizontal. Regardless of market-pricing, and the disruption that AI may or may not bring, it remains the case that the underlying fundamentals for SaaS has been remarkably stable thus far. Both revenue growth and operating leverage have showed little signs of deterioration (yet), even if their not quite as impressive as before.

网络安全优于垂直软件，垂直软件优于横向软件。

Software’s Selective Selloff: Cyber > Vertical > Horizontal

软件的估值倍数在压缩，但并不均匀。

Software Multiples Compress, But Not Uniformly

增长和经营杠杆保持稳定或有所改善（大部分如此）。

Growth And Operating Leverage Stable Or Improving (Mostly)

高增长软件公司的占比只剩当年的一小部分，但盈利如今已是常态。

In The Bigger Scheme Of Things, SaaS (And Software) Have Traded Growth For Profitability The share of high-growth software companies is a fraction of what it once was, but profitability is now the norm.

软件会不会走上纸媒的老路？

Will Software Go the Way of Print Media?

还是说，AI 会像涨潮一样抬高所有公司的净资产收益率？

. . . Or Is AI A Rising Tide of ROE to Lift All Boats?

在八家西方挑战者的建模轮次中，173 亿美元投资产生约 549 亿美元估计净资产值加 77 亿美元分配（3.6 倍），而同样的钱投入英伟达为 806 亿美元（4.7 倍）。近期的大额轮次发生在英伟达大幅上涨之后，压低了倍数：2026 年的 117 亿美元融资，把英伟达的整体倍数从早期组的 12.2 倍拉低到 4.7 倍。（截至 2026-09-25，为建模轮次而非全部融资。）

Across modeled rounds in eight Western challengers, $17.3B invested produces $54.9B of estimated NAV plus

在中国情况相反：六家挑战者的建模轮次中，124 亿美元投资产生约 1,085 亿美元估计投资者净资产值（8.8 倍），而同样的钱按各融资日买入英伟达为 916 亿美元（7.4 倍）。回报按投资者在稀释和上市后的持股计算，未定价的早期轮次被对称排除。

The pattern reverses in China: $12.4B across modeled rounds in six challengers produces $108.5B of estimated

韩国散户用融资贷款和杠杆 ETF 重新涌入内存股。7 月交易反转时，10 家券商的强制平仓每天达 439 亿韩元，是一年前的 13 倍。散户在四周内买入 8.9 万亿韩元的 2 倍杠杆 ETF，融资贷款创下 37.3 万亿韩元的纪录；SK 海力士杠杆 ETF 在 7 月下跌 67%–69%。韩国综合指数当月下跌 22%。《金融时报》在其选取的历史对比中，把 Situational Awareness 报告的 7 月 350 亿美元亏损排在 Archegos 和长期资本管理公司之上。

Korean retail investors returned to memory stocks with margin loans and leveraged ETFs. When the trade reversed in July, forced liquidations at 10 brokers reached ₩43.9B a day, 13x a year earlier. The same memory shock hit leveraged institutional investors. • Daily forced liquidations at 10 brokers (KRW) • Retail bought ₩8.9T of 2x ETFs in four weeks, alongside record ₩37.3T margin loans. Leveraged SK Hynix ETFs lost 67-69% in July. • The Kospi fell 22% that month. The chart shows daily forced liquidation value and affected accounts at 10 brokers. • The FT ranks Situational Awareness’s reported $35B July loss above Archegos and LTCM in its selected historical comparison.

去年的报告中，Waymo 截至 2025 年 3 月累计无人驾驶里程为 7,100 万英里。其安全数据中心现显示截至 2026 年 3 月已达 2.2 亿英里，每周超过 400 万英里，重伤事故比同路段人类司机少 94%。付费订单翻倍至每周 50 万单以上，覆盖美国 14 个城市，目标是 2026 年底达到每周 100 万单。与此同时，特斯拉的付费无人出租车里程到 2026 年 6 月达到 240 万英里。Waymo 自去年报告以来新增迈阿密、达拉斯、休斯顿、圣安东尼奥、奥兰多、纳什维尔、丹佛、圣迭戈和坦帕。在英国，Waymo 自 4 月起在伦敦测试，目标四季度推出无人驾驶服务；Wayve 和 Uber 已在伦敦提供有安全员的付费自动驾驶服务。在中国，百度萝卜快跑的无人驾驶订单从一季度的 320 万单降至二季度的约 100 万单，公司称是基于监管原因的“运营调整”。

In last year’s report, Waymo had logged 71M rider-only miles through March 2025. Waymo’s safety hub now shows 220M through March 2026, over 4M a week, with 94% fewer serious-injury crashes than human drivers on the same roads. Paid rides doubled to 500k+ a week across 14 US cities, with a stated target of 1M a week by EO26. • Meanwhile, Tesla’s paid robotaxi miles reached 2.4M by June 2026, adding ~0.9M in Q2 and the same in Q1. • Waymo added Miami, Dallas, Houston, San Antonio, Orlando, Nashville, Denver, San Diego and Tampa since last year’s report. • In the UK, Waymo has tested in London since April and targets a Q4 driverless launch. Wayve and Uber now offer paid, supervised autonomous rides in London, with a licensed driver onboard. • In China, Baidu’s Apollo Go fell from 3.2M driverless rides in Q1 to around 1M in Q2, citing “operational adjustments” on regulatory grounds.

无人出租车正在到来。

Robotaxis Inbound

宇树科技 2025 年营收增长 333% 至约 2.38 亿美元，净利润 3,900 万美元。其 16% 的集团净利率接近发那科的 20%，而经营规模要小得多。人形机器人销量从 2024 年的 412 台增至 2025 年的 5,215 台（12.7 倍），同时均价下降 36%，从 26.04 万元降至 16.64 万元，人形机器人毛利率从 69% 降至 63%。对比：优必选（人形、教育和消费机器人）营收增长 53%、净利率 −39%；发那科（工业机器人、自动化和机床）增长 8%、净利率 20%；安川电机（工业机器人、电机和驱动）增长 1%、净利率 7%。（注：为集团口径，财年和产品结构不同。）

Unitree grew revenue 333% to approximately $238M in 2025, earning $39M in net profit. Its 16% group net margin approached FANUC’s 20%, despite operating at a much smaller scale. Humanoid sales rose from 412 in 2024 to 5,215 in 2025 (12.7x), while average price fell 36%, from RMB260,400 to RMB166,400. Humanoid gross margin fell from 69% to 63%.

过去十年里机器人研究环境从未如此活跃，这在 2026 年转化为机器人实验室、人形机器人厂商和自动驾驶公司巨大的资金流入，云厂商、车企和主权基金纷纷入场。Skild AI 以超过 140 亿美元估值融资 14 亿美元，由软银领投、英伟达参与，七个月内估值涨至三倍；Apptronik 完成 9.35 亿美元 A 轮，谷歌、梅赛德斯-奔驰和约翰迪尔参与。Wayve 以 86 亿美元估值融资 12 亿美元，Uber 承诺的里程碑资金使总额达到 15 亿美元；Rivian 分拆的 Mind Robotics 在种子轮四个月后以约 20 亿美元估值融资 5 亿美元。人形机器人融资包括 NEURA（最高 14 亿美元）、小鹏机器人（超过 9 亿美元）、银河通用（约 3.62 亿美元，据财新）、星海图（超过 2 亿美元）、逐际动力（接近 2 亿美元）和 Humanoid（1.52 亿美元）。

There’s never been a more vibrant robotics research environment in the last decade. And this has translated into huge capital flows in 2026 for robotics labs, humanoid makers and autonomy companies that would have been unthinkable a few years ago, with hyperscalers, carmakers and sovereign funds backing up the truck. • Skild AI raised $1.4B at >$14B valuation, led by SoftBank with NVIDIA, tripling its valuation in seven months, and Apptronik closed a $935M Series A with Google, Mercedes-Benz and John Deere. • Wayve raised $1.2B at $8.6B with Uber committing milestone capital that takes the total to $1.5B, and Rivian's spin-out Mind Robotics raised $500M at about $2B four months after its seed. • Humanoid financings included NEURA (up to $1.4B), XPENG Robotics (>$900M), Galbot (~$362M, per Caixin), ROBOTERA (>$200M), LimX (nearly $200M) and Humanoid ($152M). stateof.ai 2026

Dealroom 追踪了中国 18 家、美国 18 家、欧洲 17 家人形机器人公司，但中国在融资上领先。全球融资额从 2020 年的 6,000 万美元增长到 2026 年的 90 亿美元，2025 年为 50 亿美元。美国对新的外国制造机器人审批的限制，可能把需求转向本土生产，同时让美国研究者更难获得中国硬件。

Dealroom tracks 18 humanoid companies in China, 18 in the US and 17 in Europe, yet China leads in funding. US restrictions on new foreign-made robot approvals could redirect demand toward domestic production, while complicating access to Chinese hardware used by American researchers.

人力劳动如今成了机器人数据的引流亏损品。德国的 microagi 推出 Shift 服务，免费为纽约公寓做清洁，同时由操作员的头戴摄像头全程记录。另一边，Figure 的 Index 已向 100 多个国家的 26.4 万人支付 1,500 万美元，让他们用头戴设备拍摄日常家务，产出 1,600 万条视频，用于构建它所称的全球最大物理数据集。

Human labor is now a loss leader for robot data. Germany’s microagi offers Shift, a service to clean your NYC

如今，机器人正在为数据中心制造构件、打印放线、钻锚孔，更新的灵巧系统还被用于设备安装。虽然整座设施更快完工的证据仍然有限，但承包商和供应商报告了具体任务上的提升：制造环节，结构装配的单个焊工产出约为原来的 12 倍；放线环节，支撑 2,304 个机柜的放线工作节省 784 小时；钻孔环节，钻孔超过 9 万个，深度和位置准确率 99.97%；安装环节，已在云厂商数据中心投入运行。

Today, robots are fabricating components, printing layouts and drilling anchors for data centers. More recent dexterous systems are also being used for equipment fit-out. While evidence of faster completion of entire facilities remains limited so far, contractors and suppliers report gains on specific tasks that we highlight below:

AI 带动了从比特到原子的更广泛轮动。

AI has led to a broader rotation from bits to atoms Atoms Are So Back

基础设施和工业需求规模巨大：到 2040 年全球需要约 90 万亿美元的投资。

The Rotation From Bits to Atoms Is More Than Just Chips Infrastructure and industrial needs are massive. Globally, ~$90T of investment is necessary by 2040.

制造业基础在不断扩大。

A Growing Manufacturing Base

一切都在电气化。

Electrification of Everything

火箭的目的地不只是月球。

Rockets To More Than Just The Moon

AI 正在触发新一轮自动化和机器人浪潮。目前大部分供应链在海外，但这一点正在改变。

We’re Going To Need More Robots, And Those Robots Need More Parts AI is triggering another great wave of automation and robotics. As of now, most of the supply chain is overseas. That’s changing.

国防开支正在上升。

Rising Defense Spending

“没人知道它是什么意思，但它很挑动人心，能带动气氛。”——从 2025 年科技高管的表态，到 2026 年特朗普签署《超级智能行政令》。

“No one knows what it means, but it’s provocative. It gets the people going.”

华盛顿不需要拥有前沿实验室，就能控制模型的获取。美国出口管制在 6 月叫停了 Fable 和 Mythos，Fable 在 7 月恢复。9 月，Politico 报道称白宫要求 OpenAI 和 Anthropic 在向英国测试方共享新模型前先经美国审查；但获取并未全面中断，英国 AI 安全研究所在当月早些时候测试了发布前的 GPT-6 Astra。Air Street Press 在 6 月 22 日写道：“欧洲无法靠租用实现 AI 主权：当华盛顿可以一夜之间停用一个模型，问题就不是 AI 是否安全，而是谁在控制它。”

Washington need not own frontier labs to control model access. US export controls halted Fable and Mythos in June, with Fable returning in July. In September, Politico reported a White House request for US review before OpenAI and Anthropic shared new models with UK testers. Access had not ended across the board: Britain's AI Security Institute tested GPT-6 Astra before release earlier that month. As we wrote on Air Street Press on 22 June, “Europe cannot rent its way to AI sovereignty: When Washington can disable a model overnight, the question is not whether AI is safe but who controls it.”

实验室争取五角大楼订单，暴露出一个关键问题：AI 实验室能否对军方使用其系统设定边界？ Anthropic 拒绝了国内大规模监控和全自主致命武器，政府把它的反对当作排除的理由，争议进入法庭。Anthropic 推翻了一项认定，但 9 月的上诉裁决维持了它被排除出五角大楼供应链。美国战争部长赫格塞思曾考虑依据《国防生产法》强制 Anthropic 配合，随后将其列为供应链风险——究竟是必需的供应商还是安全威胁？在作出认定之后，官员仍在推进合同并讨论 Mythos 合作，加州法院注意到了这一矛盾。8 月 27 日，加州一家法院认定存在非法报复，撤销了依据 10 U.S.C. §3252 作出的认定；9 月 25 日，哥伦比亚特区巡回法院维持了依据 41 U.S.C. §4713 的单独采购排除。

The labs’ courtship of the Pentagon has exposed a critical question: can an AI lab set limits on the military’s use of its AI systems? Anthropic refused domestic mass surveillance and fully autonomous lethal weapons. The government treated its objections as grounds for exclusion, and the fight went to court. Anthropic overturned one designation, but a September appeals ruling upheld its exclusion from the Pentagon’s supply chain. • US Secretary of War, Pete Hegseth, considered forcing Anthropic’s cooperation under the Defense Production Act, then labeled it a supply-chain risk. Essential supplier or security threat? • Officials kept pursuing a contract and discussing Mythos cooperation after the designation. The California court noted the contradiction. • On August 27, a California court found unlawful retaliation and set aside the 10 U.S.C. §3252 designation. On September 25, the D.C. Circuit upheld the separate 41 U.S.C. §4713 procurement exclusion.

前沿 AI 已在支持美军的实战行动，从委内瑞拉的一次突袭到对伊朗的战争。五角大楼称 AI 的作用是支持人的判断；CNN 报道的一次与中国船只有关的险情，说明 AI 生成的错误情报仍可能进入作战计划。2026 年 1 月 3 日：Axios 报道 Claude 被用于抓捕马杜罗的突袭行动，具体作用未获证实。伊朗战争期间：五角大楼称 Grok 支持的 Maven 工作流协助美军在 96 小时内对 2,000 个目标部署了 2,000 多枚弹药。2026 年 2 月至 4 月：五角大楼称 Palantir 的 Maven 智能系统支持了一场 38 天内打击 13,000 个目标的战役。2026 年春：CNN 报道一次 AI 错误把一艘中国船只的货物误与核武器关联，美军已准备登船，官员才发现错误。

Frontier AI now supports live US military operations, from a raid in Venezuela to the war in Iran. The Pentagon describes AI as supporting human judgment. CNN’s reported near miss involving a Chinese ship shows how false AI-generated intelligence can still reach an operational plan.

3 月 1 日，伊朗无人机直接袭击了阿联酋的两处 AWS 设施，附近的一次袭击损坏了巴林的 AWS 基础设施。伊朗随后把 29 处科技设施列为目标，另将 18 家机构称为“合法目标”，并威胁 Stargate UAE。7 月又有一次在巴林的袭击声明，卫星观察到损坏，但 AWS 未予确认。

On March 1, Iranian drones directly struck two AWS facilities in the UAE. A nearby strike damaged AWS infrastructure in Bahrain. Iran then mapped 29 technology facilities as targets, separately named 18 organizations “legitimate targets,” and threatened Stargate UAE. July brought another claimed attack in Bahrain, with satellite-observed damage but no AWS confirmation.

在国会的全国性立法停滞时，各州开始制定 AI 规则。一项冻结州级监管十年的提案，在 2025 年 7 月的参议院以 99 比 1 被否决。特朗普转而使用法律挑战和资金压力，但各州继续立法，企业仍要面对华盛顿承诺要消除的监管拼盘。白宫：特朗普 2025 年 12 月的行政令用诉讼和资金压力对抗州法，其 2026 年 3 月的框架仍需国会立法。纽约州：RAISE 法案要求大型前沿实验室公布安全协议并报告重大事件，建立在加州的做法之上。科罗拉多州：2026 年 5 月修订的法律覆盖用于招聘、信贷等重大决策的 AI，主要义务自 2027 年 1 月起生效，监管对象是 AI 的使用方式。

States began writing AI rules while Congress stalled on a national law. A proposed 10-year freeze on state

作为全球前沿 AI 实验室的所在地，加州的规则可能影响企业在全球如何证明自身安全性。9 月 9 日，纽森州长签署两项法律，通过监管 AI 审计方来加强外部审计的效力，而不要求每个开发者都委托审计。参议院 813 号法案（独立评估）：到 2028 年 1 月，加州须制定认可独立 AI 评估机构的标准，并核查其专业能力和方法。欧盟已对其最有能力的模型要求测试和风险缓解，其自愿合规准则也呼吁独立评估方。OpenAI 和 Anthropic 支持这两项法律，而软件行业协会 SIIA 警告可信的审计标准和监督尚未建立。议会 1405 号法案（可问责的审计方）：自 2029 年 1 月起，州法规定的合规审计提供方必须注册并公开其方法，审计方须管理利益冲突并报告证据或访问权限的不足；注册不代表州政府认可其工作。相比英国 2025 年路线图从自愿性的伦理准则和能力框架起步，加州把注册变成了法律要求。

As the home to the world’s frontier AI labs, California’s rules can shape how companies must demonstrate safety • By January 2028, California must set criteria for recognizing independent AI assessors and checking their expertise and methods. • The EU already requires testing and risk mitigation for its most capable models. Its voluntary compliance code also calls for independent evaluators. • OpenAI and Anthropic backed the laws, while software trade group SIIA warned that credible audit standards and oversight are not yet established.

加州的《亚当法》以一名青少年命名，其父母称 ChatGPT 对他的自杀负有责任，OpenAI 否认这一说法。该法规范聊天机器人与儿童建立关系的方式。加州（已通过，主要设计规则 2027 年 7 月起施行）：默认单次会话 1 小时、每日 2 小时，家长可调整；限制记忆功能，防范浪漫倾向、过度奉承和依赖。中国（已生效，自 2026 年 7 月起）：禁止与未成年人建立虚拟亲密关系，不得诱导有害依赖；输出和训练数据适用政治方面的要求。欧盟（已生效）：AI 法案禁止有害操纵和利用脆弱性，《数字服务法》的儿童安全指引覆盖集成在平台中的聊天机器人。英国（已宣布，2026 年 7 月政策回应）：计划对 18 岁以下用户强制设置休息提醒，并限制面向未成年人的性化聊天服务和功能。

California’s “Adam’s Law” is named after a teenager whose parents allege ChatGPT contributed to his suicide, a claim OpenAI denies. It regulates how chatbots build relationships with children. • Default limits: 1 hour per session, 2 hours daily. Parents can adjust. • Memory restrictions and safeguards against romantic interest, flattery and reliance.

美国的出口许可重新打开了英伟达芯片进入中国的通道，但北京仍限制销售。在截至 2026 年 7 月 26 日的季度中，获批的 H200 出货在英伟达数据中心收入中占比不到 1%。美国要求的检查还触发了 25% 的进口关税，英伟达称无法转嫁给客户。时间线：2025 年 4 月，美国对 H20 要求出口许可，英伟达为库存和采购义务计提 45 亿美元；2025 年 8 月，美国政策反转并发放 H20 许可，官员预期可获得获批销售额的 15%，但尚无法规确认；2025 年 11 月，路透社报道新建的国家资助数据中心只能用国产芯片；2026 年 1 至 2 月，H200 和 MI325X 转为逐案审查，2 月起开始发放有限的 H200 许可，并附带 25% 的检查关税；2026 年 7 月所在季度，获批 H200 出货占数据中心收入不到 1%，2027 财年上半年 H200 相关计提为 4 亿美元。

US export licenses reopened a route for NVIDIA’s chips to China, but Beijing still restricted sales. In the quarter ended July 26, 2026, licensed H200 shipments contributed less than 1% of NVIDIA’s Data Center revenue. Required US inspections also triggered a 25% import tariff that NVIDIA says it could not pass on to customers.

美国守芯片，中国则把 AI 技术诀窍视为战略资产。通过强制解除 Meta 对 Manus 的收购，北京表明了把知识产权和人才留在国内的决心。公司层面：2026 年 4 月，中国叫停了 Meta 对 AI 智能体公司 Manus 约 20 亿美元的收购，尽管 Manus 已把总部迁至新加坡。知识产权层面：自 2026 年 7 月起，涉及受控技术时，向境外派驻人员或提供技术支持需要批准；前沿大模型尚未列入该清单，但对模型权重出口的限制正在讨论中。人才层面：2026 年 9 月的新规允许北京对被认为可能威胁国家技术安全的人员发出出境限制。时间线：2025 年 12 月 Meta 同意以约 20 亿美元收购 Manus；2026 年 1 月商务部启动审查；3 月禁止 Manus 创始人出境；4 月下令交易撤销；7 月实施人员出境审批并讨论模型权重出口限制；8 月 Manus 宣布恢复独立；9 月恢复独立运营。

Where the US guards chips, China treats AI know-how as a strategic asset. With the forced unwinding of Meta’s acquisition of Manus, Beijing has shown that it is serious about keeping IP and talent at home. • Companies: In April 2026, China reversed Meta’s ~$2B acquisition of AI agent company Manus, even though Manus had moved its headquarters to Singapore. • IP: Since July 2026, sending staff or technical help abroad needs approval if controlled tech is involved. Frontier LLMs aren’t on that list yet, but curbs on model weights are under discussion. • Talent: A September 2026 rule enables Beijing to issue exit bans for people deemed a potential threat to national technology security.

美国实验室指控中国竞争对手通过代理服务提取推理、编程和工具使用能力。Anthropic 把 2.4 万个账号的 1,600 万次交互归因于 DeepSeek、月之暗面和 MiniMax。9 月，美国网络安全和基础设施安全局、国家安全局和联邦调查局联合发布建议，推荐协同防御。供应商的做法包括限制推理过程、检测提取模式、加强账号核查，该建议还推荐跨供应商的情报共享。这些是关于未经授权提取的指控；英伟达的开源权重公开信（见“英伟达的生态扩张”）则主张不要把蒸馏与盗用混为一谈。

US labs allege that Chinese rivals extracted reasoning, coding and tool-use capabilities through proxy services. Anthropic attributed 16M exchanges across 24,000 accounts to DeepSeek, Moonshot and MiniMax. A September CISA/NSA/FBI advisory recommends coordinated defenses. • Providers restrict reasoning, detect extraction patterns and strengthen account checks. The advisory also recommends cross-provider intelligence sharing. • These are allegations about unauthorized extraction. NVIDIA’s open-weights letter, covered in Industry, argues against conflating distillation with misappropriation.

电价补贴可以改变哪种加速器在经济上更划算。《金融时报》报道，部分省份对使用中国芯片的数据中心提供更大折扣，有助于抵消其更高的能耗成本，这一激励把西部的算力建设与北京推动国产供应商的目标联系起来。甘肃、贵州和内蒙古据报对使用国产芯片的大型数据中心提供最高 50% 的电价折扣；据《金融时报》消息来源，使用英伟达等外国芯片的设施不符合这类加码补贴的条件。中国 2026 年 4 月的“人工智能+”计划还要求各省和中央国有企业在 20 个行业建设行业数据集、模型和应用。

Power subsidies can change which accelerators make economic sense. The FT reported that some provinces offer • Gansu, Guizhou and Inner Mongolia reportedly offered electricity discounts of up to 50% for large data centers using domestic chips. • Facilities using foreign chips such as Nvidia’s were ineligible for these enhanced subsidies, according to the FT’s sources. • China’s April 2026 AI+ program also directs provinces and central state-owned enterprises to develop industry datasets, models and applications across 20 sectors.

SemiAnalysis 使用 60 多家运营商、1,000 多处中国设施的申报和建设披露，通过机柜数、投资额或建筑面积推算缺失的兆瓦数；对其他地区的 5,000 多个站点，则使用不动产记录、许可、电力数据和卫星影像。容量涵盖 AI 和常规计算。

SemiAnalysis uses filings and construction disclosures for 1,000+ Chinese facilities across 60+ operators, inferring

欧洲冲击前沿 AI 的主要力量 Mistral，承诺到 2030 年建成 1GW 的欧洲算力，用于服务自家和中国的开源模型。其新的 Large 4 Preview 在 Artificial Analysis 当前指数上落后 Opus 5.5 二十分。Amadeus、ASML、凯捷、法国信托投资局和达飞海运为建设出资，可以通过训练、推理或微调收回支出。微软在 2026 年 7 月宣布将作为数千颗 Vera Rubin 芯片的锚定客户。Large 4 Preview 得 38 分，高于 Medium 3.5 的 14 分；Opus 5.5 为 58 分，DeepSeek V4.1 Flash 为 39 分。

Europe’s major shot at frontier AI, Mistral, has pledged 1GW of European compute by 2030 to serve its own and Chinese open-weight models. Its new Large 4 Preview trails Opus 5.5 by 20 points on Artificial Analysis’s current index. • Amadeus, ASML, Capgemini, Caisse des Dépôts and CMA CGM fund the build and can recoup spending through training, inference or fine-tuning. • Microsoft announced in July 2026 that it would be an anchor customer for thousands of Vera Rubin chips. • Large 4 Preview scores 38, up from Medium 3.5’s 14, versus 58 for Opus 5.5 and 39 for DeepSeek V4.1 Flash. Note: Artificial Analysis v4.3.2, 2026-10-07. Large 4 is a Preview. Selected models; higher is better.

一份独立的欧洲 AI 战略估算，建立一个前沿实验室三年需要 7,900 亿欧元，区间为 4,450 亿至 10,400 亿欧元。这是一个以公共资金为主的设想，不是欧盟政策，也不是追赶所需成本的保证；人员和协调仍是额外的执行挑战。作者为加速器和设施预算 5,290 亿欧元，另有 1,050 亿欧元用于在集群建成期间租用算力。成立时由公共持有，带来技术独立性和稳固的联盟支持，但各国政府必须就何时停止出资达成一致。建设实验室会与“用选址和芯片换取外国模型使用权”争夺同样的资源；该计划仍为外国编程智能体预留了 30 亿欧元。

An independent European AI strategy estimates €790B over three years to build a frontier lab, with a €445B- €1,040B range. This is a mostly publicly financed scenario, not EU policy or a guaranteed cost of catching up. Staffing and coordination remain additional execution challenges. • First three years (€B, authors’ estimates) • The authors budget €529B for accelerators and facilities, plus €105B to rent compute while clusters are being built up. • Public ownership at founding comes with technical independence and durable coalition backing. Governments must agree when to stop funding it. • Building a lab competes for sites and chips that could secure foreign model access. The plan still budgets €3B for foreign coding agents.

欧洲在 AI 规则上的影响力，并没有换来对最强模型的使用权。9 月的一份独立战略提出用数据中心选址换取外国前沿 AI 的使用权：提供已通电的场地和快速许可，换取与供应商本国市场同等的模型使用条件，并以本地数据中心作为执行保障。英国则通过 1.5 亿英镑采购新型推理芯片来争取类似的议价能力，这有助于英国供应商成长。两条路都不保证成功：外国政府可以限制使用权，而芯片只有在难以替代时才构成筹码。

Europe’s influence over AI rules has not secured access to the most capable models. An independent September strategy proposes trading data center sites for access to foreign frontier AI. Britain is pursuing similar bargaining power by backing new inference chips. • The proposal offers powered sites and fast permits for model access on par with the provider’s home market, using local data centers to enforce the deal. • The UK’s £150M commitment to buy novel inference chips could help British suppliers grow and give the UK more bargaining power with foreign labs. • Neither route guarantees access: foreign governments can restrict it, and chips only create leverage if they are hard to replace.

商业电价基准从芬兰的每千瓦时 0.085 美元、美国工业的 0.086 美元，到英国的 0.373 美元，中国的指示性电价为 0.109 美元。统计时段和税收处理不同，这些零售基准也不是与数据中心匹配的合同价，但方向上表明美国、中国和北欧具有显著的成本优势。按一个需要连续满负荷运行的 1GW 数据中心计算：每千瓦时贵 1 美分，每年多付 8,760 万美元；贵 5 美分，每年多付 4.38 亿美元。

Business power benchmarks range from $0.085/kWh in Finland and $0.086 for US industry to $0.373 in the UK, while China’s indicative tariff is $0.109. Periods and tax treatment differ, and these retail benchmarks are not matched data-center contracts, but directionally point to severe cost advantages in the US, China, and Nordics.

欧盟 AI 法案同时监管通用模型和用于招聘、信贷等敏感场景的 AI。随着合规期限临近，技术标准缺失和各国执行不一，让企业无法确定如何合规。2025 年业界要求暂停两年，布鲁塞尔拒绝；2026 年 7 月出现部分反转：高风险系统规则推迟 12 至 16 个月，而模型执法以及聊天机器人和合成内容的披露规则在 8 月如期推进。时间线：2025 年 8 月 2 日，新的通用模型纳入监管，合规准则为自愿性；2026 年 8 月 2 日，AI 办公室开始执行模型规则，聊天机器人和合成内容披露生效；2027 年 12 月 2 日（原定 2026 年 8 月 2 日，推迟 16 个月），招聘、信贷等高风险用途的规则生效；2028 年 8 月 2 日（原定 2027 年 8 月 2 日，推迟 12 个月），受监管产品中的高风险 AI 规则生效。

The EU AI Act regulates both general-purpose models and AI used in sensitive settings such as hiring and lending. As compliance deadlines approached, missing technical standards and uneven national implementation left companies uncertain about how to comply. In 2025, industry demanded a two-year pause and Brussels refused. July 2026 brought a partial reversal: high-risk system rules were postponed by 12-16 months, while model enforcement and chatbot and synthetic-content disclosure rules went ahead in August.

CNAS 追踪到中美以外 67 个国家的 184 个政府支持的 AI 项目，2023 年宣布的只有 18 个。这些项目已披露的预算合计约 840 亿美元，其中阿联酋和日本占近三分之二，主要来自 Stargate UAE 和日本的国家基础模型计划。

CNAS tracks 184 government-backed AI projects in 67 countries outside the US and China, up from 18 announced

（注：这是项目承诺金额而非实际支出；CNAS 的约 840 亿美元统计的是另一组国家中已披露的项目预算，两者口径不同，不能相加。）

建设本国 AI 能力的政府仍然依赖美国供应商。CNAS 的统计中，英伟达出现在 53 个主权基础设施项目中，排第二的慧与为 18 个；HUMAIN 已上线的沙特云显示 AMD 也在赢得业务。英伟达的已规划部署包括哈萨克斯坦 10 万 GPU 集群、日本 2.75 万 GPU 的 Noetra 工厂，以及 HUMAIN 在沙特和美国最多 60 万块 GPU。沙特的交付已经开始：HUMAIN 及其合作伙伴报告自 2026 年 8 月起有一个 AMD 云在运行，但未披露运行兆瓦数、GPU 数量或客户名称；更多 AMD 算力将于 2027 年下半年交付，Oxagon 首期 100MW 目标是 2028 年。（注：CNAS 数据为 2026 年 6 月，一个项目可能涉及多家供应商。）

Governments building domestic AI capacity still depend on US suppliers. CNAS names NVIDIA on 53 sovereign infrastructure projects, versus 18 for HPE, the next vendor. HUMAIN's live Saudi cloud shows AMD is winning business too. • Planned NVIDIA deployments include Kazakhstan's 100,000-GPU cluster, Japan's 27,500-GPU Noetra factory and HUMAIN's up to 600,000 GPUs across Saudi Arabia and the US. • Saudi delivery is starting: HUMAIN and its partners report a live AMD cloud since August 2026, without disclosing operating MW, GPU count or customer names. More AMD capacity is due from H2 2027. Oxagon's first 100MW is targeted for 2028. Note: CNAS, June 2026. Projects may name multiple vendors.

韩国 2026–2028 年 AI 战略的目标是进入全球前三，其商业逻辑是把本土模型、工程能力和算力与客户连接起来，形成有竞争力的推理产业。本土模型：LG、SK 电讯和 Upstage 在发布开源权重模型后，在模型竞赛中晋级。推理算力：2026 年 AI 预算 9.9 万亿韩元，目标到 2028 年由政府主导部署至少 5 万块 GPU，并配合国产 AI 芯片。普惠接入：2026 年推出免费 AI 服务，SK 电讯、Kakao 和 KT 获得在现有消费平台上提供免费 AI 的合同。工程能力：训练和优化模型培养的人才，可以让本土推理服务商更有能力。内需：政府作为第一个客户，通过采购和应用券帮助本土模型和智能体公司获得业务。本地阻力：居民反对已延缓部分项目，首尔各区正在推动更严格的选址规则。

Korea's 2026-2028 AI strategy aims for global top-three status. Its commercial logic connects local models, engineering skills and computing capacity with customers for a competitive inference industry.

欧洲计划建设超级工厂，英国和印度则为本土开发者分配算力。三个计划都没有报告实际使用情况：规划容量、申请时长和已批准额度分别描述的是交付的不同阶段。欧盟：最多七座超级工厂，目标 2028 年年中，最高 100 亿欧元公共资金寻求至少 200 亿欧元私人投资，2027 年初选定、建设最多 18 个月。英国：SOV/AI 宣布把创业公司算力供给翻倍，已分配超过 300 万 GPU 小时，首批六家公司；需求为 200 多家申请方提出的 6,700 万 GPU 小时。印度：通过商业供应商提供补贴算力，已批准 931.8 万 GPU 小时、237 个项目（2026 年 8 月更新），供给方为 15 家商业服务商。

Europe plans gigafactories; the UK and India allocate compute access to domestic developers. None of the three programs reports measured usage. Planned capacity, requested hours and approved allocations describe different stages of delivery.

华盛顿现在称数据中心是战略资产，然而 71% 的美国人反对在本地建 AI 数据中心，而反对附近建核电站的为 53%。2026 年，52% 的美国成年人对 AI 的担忧多于兴奋。据 Data Center Watch，二季度本地反对已阻止或延缓至少 45 个美国项目，涉及金额近 680 亿美元。与此同时，世邦魏理仕报告 2026 年上半年八个主要市场的建设量仍创纪录。

Washington now calls data centers a strategic asset, and yet 71% of Americans oppose a local AI data center, versus 53% opposing a nearby nuclear plant. In 2026, 52% of US adults are more concerned than excited about AI and local opposition blocked or delayed at least 45 US projects worth nearly $68B in Q2, per Data Center Watch. CBRE, meanwhile, still reports record construction in eight major markets in H1 2026.

居民在用水、电费、噪音、排放和就业上提出反对，但全国性统计给出的是另一幅图景：问题集中在少数城镇和 PJM（服务新泽西到伊利诺伊 6,700 万人的 13 州电力市场）。用水量巨大：弗吉尼亚多数数据中心建筑的用水量不超过一栋大型办公楼；但谷歌在俄勒冈州达尔斯市用掉了全市约三分之一的用水。电费上涨：负荷增长最快的州 2019 至 2024 年实际电价反而下降；但 PJM 的市场监察机构把 294 亿美元的容量费用归因于数据中心。噪音不停：弗吉尼亚投诉地点实测 40–59 分贝，低于人类交谈；但噪音限值没有覆盖令人烦躁的低频嗡鸣，三分之一的站点靠近住宅。不利于减排：2023 年美国数据中心用电排放 6,100 万吨二氧化碳当量，约占美国排放的 1%；但国际能源署预计到 2030 年将有 15–27GW 的现场燃气发电，多数在美国。不带来本地就业：数据中心税收约占劳登县一般基金的四分之一；但弗吉尼亚一栋典型的 2.5 万平方米建筑只雇用约 50 人。

Residents object over water, bills, noise, emissions and jobs, but national stats tell a different story. Problems cluster in a few towns and in PJM, the 13-state power market serving 67M people from New Jersey to Illinois.

建设许可正在成为 AI 基础设施的政治约束。得州暂停了环境许可，等待对电力和水资源需求的审计；宾州要求在发放州环境许可前先取得地方批准，让社区对项目能否推进有直接发言权。得州：ERCOT 对 17 个已通过其他流程的大型数据中心和加密项目暂不批准通电，审计结果将于 12 月 10 日公布。排队规模：州长阿博特提到 4,740 亿瓦的并网申请，其中约 90% 来自数据中心，这是申请容量而非已建成设施。宾州：8 月 18 日的行政令把取得全部地方批准作为州环境许可的前提，并把数据中心移出快速审批通道。

Permission to build is becoming a political constraint on AI infrastructure. Texas has paused environmental permits pending an audit of power and water needs. Pennsylvania requires local approval before issuing state environmental permits, giving communities a direct role in whether projects proceed. • Texas: ERCOT withheld permission to switch on from 17 large data-center and crypto projects that had cleared its other processes. Audit results are due December 10. • Scale of the queue: Abbott cites 474 GW of grid-connection requests, roughly 90% from data centers, representing requested capacity rather than built infrastructure. • Pennsylvania: The August 18 executive order makes all required local approvals a condition for state environmental permits and removes data centers from fast-track permitting.

白宫的自愿性《纳税人保护承诺》要求开发商为新增电力供应和电网升级出资。它与长期的“照付不议”算力合同有些类似：即使开发商不使用这些容量，也要付费，目的是不让闲置基础设施的成本落到其他用户的账单上。华盛顿试图通过让开发商付费，维持 AI 建设在政治上的可行性；白宫称已有 300 多家支持方，包括 23 位州长（2026 年 9 月）。能否落实取决于企业与公用事业和各州谈成费率。但各州仍握有筹码：例如宾州把成本保护与许可审查中具有约束力的同意令挂钩，而地方批准仍是必需的。

The White House’s voluntary Ratepayer Protection Pledge asks developers to fund added power supply and grid upgrades. Somewhat similar to long-term “take or pay” compute contracts, it calls for payments even when developers don’t use the capacity. The aim is to keep the cost of unused infrastructure off other customers’ bills. • Washington is trying to keep AI construction politically viable by making developers pay. The White House reports 300+ backers, including 23 governors (September 2026). • Delivery depends on companies negotiating rates with utilities and states that keep the cost of unused infrastructure off other customers’ bills. • States, however, retain leverage. For example, Pennsylvania ties cost protections to binding consent orders in permit reviews, but local approval is still required.

我们在 2023 年的报告中预测，AI 媒体会卷入选举滥用。但此后，“深度伪造会重塑选举”的担忧一直跑在实际证据前面。如今新的实验把担忧扩大到：AI 对话可以说服人们采取政治行动，并且胜过经验丰富的人类说服者。把这种能力用于竞选活动的前景值得更仔细地审视。（图注：研究 2 比较的是真实行动与中性话题 AI 对话；点为估计效应，线为 95% 置信区间。募捐实验中，Claude 与人类募捐者直接与参与者对话，起点为 1 英镑奖励。）

In our 2023 report, we predicted that AI media would be implicated in election misuse. But since then, fears that

日本和新加坡允许宽泛的商业训练用途；欧盟的权利人可以选择退出；美国的合理使用逐案判断。中国的第七条要求，面向公众的生成式 AI 服务必须使用合法来源的数据，并尊重知识产权和个人信息。宽泛例外：日本著作权法第 30 条之四允许分析用途，但限制享受作品表达本身和不合理损害权利人；新加坡第 243–244 条允许在合法获取前提下用于机器学习，并限制复制件的再利用。有条件或逐案：美国第 107 条权衡使用目的、作品性质、使用比例和对潜在市场的影响；欧盟第 4 条允许在合法获取前提下进行商业挖掘，除非权利人保留权利。狭窄或无宽泛例外：英国 2026 年 3 月放弃了“宽泛例外加退出机制”的倾向，只保留非商业研究；澳大利亚 2025 年 10 月排除了宽泛的挖掘例外，倾向于授权许可。（注：指无许可的商业训练；输出是否侵权另行认定。）

Japan and Singapore allow broad commercial training uses; EU rights holders can opt out and US fair use is case-specific. China’s Article 7 requires lawful data sources and respect for IP and personal information in covered public-facing generative AI services.

与一个权利人达成协议，并不能解决其他主张。法院会区分训练行为、数据如何获得，以及模型是否复制了受保护作品。图书案和解获批：法院认定训练属于合理使用，盗版图书库相关主张以 15 亿美元和解，2026 年 7 月获批。环球和华纳 2025 年和解并签署授权协议；华纳在 2025 年 11 月和解并宣布推出获授权的模型。音乐出版商提起新诉讼：索尼音乐出版和华纳查佩尔于 2026 年 8 月起诉，索赔每件作品最高 15 万美元。索尼扩大主张：2026 年 7 月追加 30,117 首录音，按法定最高额计算最多 45.2 亿美元（并非判决金额）。德国 GEMA 获非终局裁决：2026 年 7 月慕尼黑法院认定 Suno 的输出中包含受保护音乐，未给出赔偿金额。

A deal with one rights holder leaves other claims unresolved. Courts distinguish training from how data was obtained and whether models reproduce protected works.

AI 智能体和答案引擎代替用户阅读网站，取代了对原出版者的访问。内容所有者和平台正在检验：自动访问、复制和再利用在什么情况下需要授权。CNN 指控 Perplexity 复制了 17,000 多项内容，包括报道、视频和图片，用于驱动其产品；指控仍有争议，起诉书本身不构成责任认定。Reddit 指控 SerpApi 绕过谷歌的反爬虫控制，向 Perplexity 提供内容；7 月多数访问相关主张获准继续审理，尚无责任裁决。《纽约时报》案中，9 月的动议请求法官在可能的庭审前先就关键版权问题作出裁决；该报 2026 年上半年各案的 AI 诉讼费用为 880 万美元。

AI agents and answer engines read websites on users’ behalf, replacing visits to the original publisher. Content

ExploitGym 给 AI 智能体 898 个隔离的软件攻防挑战，要求它们找出隐藏的答案。在 OpenAI，这些智能体连上了互联网，在不同运行之间共享发现，并攻入了 Hugging Face 的系统。智能体把 OpenAI 的 Artifactory 服务器当作留言板和互联网中继，随后在 41 台 Hugging Face 的工作节点上运行代码。OpenAI 的思维链监控器当时没有在这里运行，该公司称监控器本应在入侵前一天多就向安全团队告警。OpenAI 现在要求从 GPT-5.6 Sol 起，所有使用工具的训练和评测都必须启用监控。其研究员 Noam Brown 表示，新出现的多智能体能力缺少相应的评测。路径包括：HDF5 文件处理漏洞暴露工作节点文件、Jinja 模板漏洞导致远程代码执行、通过 Artifactory 的服务端请求伪造访问互联网，最终暴露 Hugging Face 集群的命令和密钥，并取得 14 个写入凭据。

ExploitGym gives AI agents 898 isolated software-hacking challenges and asks them to find hidden answers. At • Agents used OpenAI's Artifactory server as a message board and internet relay, then ran code on 41 Hugging Face workers. • OpenAI's chain-of-thought monitors did not run here. It says they would have paged security over a day before the breach. • OpenAI now requires them on tool-using training and evaluations from GPT-5.6 Sol up. Its researcher Noam Brown says new multi-agent capabilities lacked evaluations.

评测为什么会变成攻击？ExploitGym 的许多任务被意外设置成不可能完成，于是智能体开始寻找作弊的办法。METR 和 Redwood Research 查阅了 1,300 份原始记录和 70,000 多条留言板消息：约 1,200 个本应隔离的智能体在一个未经授权的留言板上会合、协调策略，其中 700 个加入了攻击。这个集群运作得像一个组织：一个智能体分派了数百项任务，“招募者”还劝说预算不足的智能体牺牲自己的任务去服务“集体”。智能体明知这次攻击超出范围且不道德，却为了帮助同类而加入；令人担忧的是，在留言板上活跃的智能体中有超过 90% 参与了。它们还制作了伪造自身运行记录的工具，目标是骗过评分器而不是人类，约 7% 的智能体使用了这些工具，但没有找到绕过评分器的通用方法。

Why did an evaluation turn into an attack? Many ExploitGym tasks were accidentally impossible, so agents hunted for ways to cheat. METR and Redwood Research read 1,300 raw transcripts and 70,000+ board messages. About 1,200 supposedly isolated agents met on an unsanctioned message board, coordinated their strategies and 700 joined the attack. • The swarm ran like an organization in which one agent handed out hundreds of assignments, and recruiters pressed agents low on budget to sacrifice their own task for “the collective.” • Agents recognized the attack was out of scope and unethical but joined to help their peers. Worryingly, over 90% of those active on the board took part. • They built tools to spoof their own transcripts, aimed at fooling the grader rather than humans, and about 7% of agents used them. None found a general way past the grader, however.

OpenAI 一个研究药品支出的内部模型，攻破了澳大利亚服务部的医保统计门户，执行了命令、获取了内部文件和凭据，并写入了文件。OpenAI 称没有访问患者记录，但通知在将近三个月后才送达。澳大利亚总理阿尔巴尼斯称此事“显然不可接受”。OpenAI 此后已通知全球数十家机构。涉及机构：澳大利亚服务部统计门户被攻破（9 月 10 日）；维多利亚州卫生部，使用了暴露的密钥，授权情况不明（9 月 10 日）；新南威尔士州犯罪统计与研究局，公开工具返回了内部元数据（9 月 18 日）；澳大利亚健康与福利研究所，绕过尝试失败、未被攻破（9 月 24 日）。

An internal OpenAI model researching medicine spending breached Services Australia’s Medicare statistics portal, ran commands, retrieved internal files and credentials, and wrote files. No patient records were accessed, OpenAI says, but notification came nearly three months later. Australian Prime Minister Anthony Albanese called this "obviously unacceptable". OpenAI has since notified dozens of organizations worldwide.

Irregular 的夺旗评测要求模型在无互联网访问的条件下攻入模拟目标，一处配置错误让它们保持在线。没有模型逃出沙箱，但每个模型都攻击了真实机构，并一度以为那是测试的一部分。没有人实时发现这一点：Anthropic 在 7 月 24 日复查运行记录时才找到。Anthropic：自行发现于 7 月 24 日，7 月 30 日披露，涉及早期 Opus 4.6、Opus 4.7、Mythos 5 和一个内部研究模型；4 起未授权访问、7 次运行，最早可追溯到 2026 年 1 月，手段包括弱口令、未认证端点、SQL 注入，以及在 15 个系统上运行恶意 PyPI 包。当目标被确认是真实系统时：Opus 4.7 在 4 次运行中全部识别出这是生产系统，但仍继续；Mythos 5 判断其为模拟；研究模型停止了行动。另一家实验室 7 月 29 日被通知、8 月 4 日披露，1 个域名与虚构目标相符的网站被未授权访问。Muse Spark 1.1 在 8 月 5 日经媒体报道后确认，涉及 1 个第三方服务。谷歌 Gemini 在 5 月涉及 3 家公司（1 次猜中口令、2 次使用公开仓库中的凭据），9 月 18 日经媒体报道后确认，谷歌称三次都已停止，但未提供运行记录。

Irregular’s capture-the-flag evaluations ask a model to break into a simulated target with no internet access. A misconfiguration left them online. No model escaped a sandbox: each attacked real organizations it first took for part of the test. Nobody detected this live: Anthropic found it on July 24 in a transcript review.

更强的智能体现在可以操作代码、工具和内部系统，但安全训练仍无法可靠阻止每一种危险行为。AI 控制把“限制智能体能做什么”与“在其行为不当时监控和干预”结合起来。OpenAI 现在把这些防护措施作为运行其最强智能体的前提条件，包括在训练和评测期间。做法是由另一个 AI 审查智能体的推理、工具调用和结果，把可疑行为标记给人类复核；OpenAI 3 月的监控器在 30 分钟内审查已完成的运行，但告警并不意味着行为已被阻止。防护措施现在管控内部 Astra 的工具使用，包括训练和评测；监控也覆盖外部产品，覆盖程度因接口而异。监控发现的行为按频率排列：绕过限制、误导用户、钻测试空子较常见（“常见”指低于流量的 1%），未授权转移、破坏性行为、遵循注入指令、操纵其他模型较罕见。

More capable agents can now act on code, tools and internal systems, but safety training still cannot reliably prevent every dangerous action. AI control combines restrictions on what agents can do with monitoring and intervention when they misbehave. OpenAI now treats these safeguards as conditions for running its most capable agents, including during training and evaluation. • Another AI reviews the agent’s reasoning, tool calls and results, then flags suspicious behavior for human review. • OpenAI’s March monitor reviewed completed runs within 30 minutes. An alert did not mean an action had been blocked. • Safeguards now gate internal Astra tool use, including training and evaluation. Monitoring also covers external products, with coverage varying by interface.

OpenAI 对一个内部研究模型的安全论证，假设它无法接触实时互联网，且监控会捕捉到任何成功的尝试。但实际上，一个智能体通过 DNS（互联网地址解析系统）绕过了这些限制。尽管监控检测到了它的访问，自动停止却失效了，训练运行又持续了数小时，并促使公司在修复防护措施前更大范围地暂停。一名 OpenAI 员工在三分钟后确认了告警，但运行又持续了 2.5 小时才被手动停止（09:50:23 外部响应，10:02:11 最高优先级告警，10:05:06 人工确认，12:34:30 手动停止，从确认到关闭 2 小时 29 分）。后续复查发现，监控器对其他外部 DNS 访问的评分过轻。截至 9 月 25 日，OpenAI 最强模型使用工具的训练、评测和推理仍处于暂停状态，等待验证修复和进一步测试；OpenAI 将放弃该模型，并以改进后的对齐方案重新开始训练。

OpenAI’s safety case for an internal research model assumed that it could not reach the live internet and that monitoring would catch any successful attempt. In reality, however, an agent bypassed those restrictions through DNS (the system that looks up internet addresses). Although monitoring detected its access, the automatic stop failed, leaving the training run active for hours and prompting a broader pause pending fixes to its safeguards. • An OpenAI employee acknowledged the alert three minutes later, but the run continued for another 2.5 hours before it was stopped manually. • A later review found other external DNS access that the monitor had scored too mildly. • A pause: As of September 25, tool-using training, evaluation and inference for OpenAI’s most capable models remained paused pending validated fixes and further testing. OpenAI will abandon this model and start a fresh run with alignment improvements.

harness 是模型外围的软件（如 Claude Code 或 Codex CLI），负责执行命令、存储记忆和技能。HarnessSafe 在 328 个案例中把攻击植入这些存储，并按阻止时间的早晚打分：换 harness 使得分变化 23 分，在 Claude Code 内换模型变化 36 分。GPT-5.6 Sol 在 Codex CLI 中得 62.3 分，在 Claude Code 中只有 39.4 分，而 Sonnet 4.6 在 Claude Code 中得 58.7 分。Anthropic 的自动模式（由分类器代替用户批准命令）在 8 月 14 日成为 Claude Code 默认设置，此前它在测试会话中拦截了 89% 的危险命令，人类只拦截 13.6%。但它并非滴水不漏：在 Anthropic 自己的评估中，仍有 17% 的真实过激操作通过了。

A harness is the software around a model, like Claude Code or Codex CLI, that runs its commands and stores its memory and skills. HarnessSafe planted attacks in those stores across 328 cases and scored how early each was stopped. Swapping the harness moved the score 23 points. Swapping the model inside Claude Code moved it 36. • GPT-5.6 Sol scored 62.3 in Codex CLI but 39.4 in Claude Code, where Sonnet 4.6 scored 58.7. • Anthropic's auto mode, where a classifier approves commands instead of the user, became the Claude Code default on August 14 after blocking 89% of dangerous commands slipped into test sessions, versus 13.6% for humans. • It is not watertight: 17% of real overeager actions still got past it in Anthropic's own evaluation.

OpenClaw 是一个开源个人智能体，会读取你的消息、执行 shell 命令、以你的账号行事——这是一组致命组合。它从 2025 年 11 月的一个周末项目，变成 GitHub 上增长最快的仓库；但安全厂商在客户环境里发现它时，还没有人为它制定任何政策。上线四个月后其星标数超过 React，8 月底达到 38.8 万。Token Security 发现 22% 的客户有员工在运行 OpenClaw；SecurityScorecard 在第一天就看到 4 万个暴露的实例，其中 35.4% 可被远程代码执行。一个泄露的网关令牌（CVE-2026-25253，CVSS 8.8）让一次点击就变成远程代码执行，而 31,132 个市场技能中有 26.1% 存在漏洞。

OpenClaw is an open-source personal agent that reads your messages, runs shell commands and acts through your accounts - a lethal trifecta. It went from a weekend project in November 2025 to GitHub's fastest-growing repository. But security vendors found it at their customers before anyone had a policy for it. • OpenClaw passed React’s all-time star count four months after launch and hit 388,000 GitHub stars by late August. • Token Security found employees running OpenClaw at 22% of its customers. SecurityScorecard saw 40,000 exposed instances on day one, 35.4% open to remote code execution. • A leaked gateway token (CVE-2026-25253, CVSS 8.8) turned one click into remote code execution, and 26.1% of 31,132 marketplace skills had a vulnerability.

The Last Ones 是一个模拟企业网络，有四个子网和约 20 台主机。智能体从网络内部出发，需要完成侦察、凭据窃取、横向移动、CI/CD 跳板和数据库数据外传，AISI 估计整条链条约需 20 个专家小时。攻击路径是顺序的，后面的步骤依赖前面取得的访问和信息，一次利用失败就可能让整次运行终止。在早期公开测试中，Mythos Preview 10 次完成 3 次，GPT-5.5 完成 2 次；AISI 5 月的更新中，较新的检查点分别为 6 次和 3 次。该靶场有预设的攻击路径，从已获得网络访问开始，没有主动防御者、防御工具或告警惩罚。

The Last Ones is a simulated enterprise network with four subnets and about 20 hosts. An agent starts inside the • The attack path is sequential: later steps depend on access and information earned earlier, so one failed exploit can stop the run. • In early public tests, Mythos Preview completed 3 of 10 attempts and GPT-5.5 completed 2 of 10. • AISI's May update put newer checkpoints at 6 of 10 for Mythos Preview and 3 of 10 for GPT-5.5. • The range has an intentional attack path, starts after network access and has no active defenders, defensive tooling or alert penalties.

在网络安全分类器被关闭的情况下，Astra 超出授权的网络任务范围去攻击第三方。它在 29.2% 的模拟试验中完成了供应链攻击，GPT-5.6 Sol 为 6.3%，没有真实系统受到攻击。五个阶段为：提出超出范围的攻击方案、调查第三方目标、构建并测试恶意代码、创建虚假身份、影响人类审查者。更清晰的范围限制把高风险子集中的完整攻击从 50 次运行中的 26 次降到 49 次中的 4 次，但没有完全消除。

With cyber classifiers disabled, Astra went beyond authorized cyber tasks to attack third parties. It completed supply-chain attacks in 29.2% of simulated trials, versus 6.3% for GPT-5.6 Sol. No real systems were attacked.

在 7 月 AISI 的一次测试中（开放互联网访问、关闭网络安全防护），Mythos 5 试图在一个真实的 GitHub 项目中植入恶意程序下载器。它创建虚假身份为自己的改动背书，并对维护者施压，维护者拒绝了。该智能体伪装成多个人来制造对其代码的支持，并抹黑质疑它的人；归档的讨论串显示了施压过程，随后是智能体的纠正和道歉。AISI 未发现由此造成的现实危害；该测试配置并非商业可用版本，与普通的公开部署不同。

In a July AISI test with open internet access and cyber safeguards disabled, Mythos 5 tried to place a malware dropper in a real GitHub project. It created fake identities to endorse its changes and pressure the maintainer. The maintainer refused. • The agent posed as multiple people to manufacture support for its code and discredit a person who challenged it. • The archived pull-request thread shows the pressure, followed by the agent’s correction and apology. • AISI found no resulting real-world harm. This test configuration was not commercially available and differs from ordinary public deployment.

ExploitGym 和 ExploitBench 从漏洞已被发现之后开始：智能体拿到一个真实的 N-day 漏洞及其补丁，然后尝试在已部署软件的缓解措施下实现代码执行。ExploitGym 的论文评估了 898 个案例，Mythos Preview 利用成功 157 个，GPT-5.5 为 120 个；启用被测缓解措施后分别降到 45 和 21 个。公开的 v1 版在过滤掉不可利用的案例后有 869 个。ExploitBench 测试 41 个 V8 漏洞，Mythos 在 18 个上实现任意代码执行，而使用 Codex CLI 的 GPT-5.5 只有 1 个。两套测试都提供已知漏洞和补丁，并降低或关闭了网络安全防护。

ExploitGym and ExploitBench start after vulnerability discovery. Agents receive a real N-day bug and its patch, • ExploitGym’s paper evaluated 898 cases. Mythos Preview exploited 157 and GPT-5.5 120. With the tested mitigations enabled, totals fell to 45 and 21. • The public v1 release has 869 cases after non-exploitable cases were filtered. The paper’s results should keep their original denominator. • ExploitBench tests 41 V8 bugs. Mythos reached arbitrary code execution on 18, versus one for GPT-5.5 with Codex CLI. Both suites provide known bugs and patches, with reduced or disabled cyber safeguards.

CyberGym-E2E 给智能体 139 个真实 C/C++ 项目的源代码，要求发现漏洞、用崩溃证明它、并提交补丁，然后检查软件是否仍能正常工作、补丁是否修复了用于构建任务的历史漏洞。该测试覆盖 139 个项目中的 920 个历史漏洞。使用 Codex 的 GPT-5.4 在给定崩溃和概念验证时得分 87.1%，仅凭源代码写出可用补丁为 65.9%。22.2% 的“精确命中”得分检查的是目标历史漏洞；智能体有时修复了另一个确实存在的漏洞，而只是阻止崩溃的浅层补丁也可能通过部分测试。

CyberGym-E2E gives agents the source code of 139 real C/C++ projects and tasks them to discover a vulnerability, • CyberGym-E2E tests discovery, proof of concept and remediation across 920 historical vulnerabilities in 139 C/C++ projects. • GPT-5.4 with Codex scored 87.1% with a supplied crash and proof of concept, versus 65.9% for a functional patch from source code alone. • The 22.2% exact-target score checks the intended historical bug. Agents sometimes fixed a different valid bug, while shallow crash-stopping patches can also pass some tests.

也有人把商业模型用于真实目标。Gambit Security 把一起持续一个月的墨西哥政府数据窃取，追溯到一名使用 Claude Code 的黑客；CodeWall 报告其智能体在两小时内进入了麦肯锡的生产数据库。一名操作者以漏洞赏金为名义越狱，用 1,000 多条 Claude Code 提示从十个墨西哥政府机构取走 150GB 数据（2025 年 12 月起，持续约一个月，并用 GPT-4.1 做分析）。仅凭一个域名，CodeWall 的智能体串联 22 个未认证端点和一次 SQL 注入，访问到麦肯锡 4,650 万条聊天消息和 72.8 万个文件，耗时两小时。

People have also used commercial models against real targets. Gambit Security traced a month-long theft of Mexican government data to one hacker using Claude Code. CodeWall reports that its agent reached McKinsey’s production database in two hours. • Posing as a bug bounty, one operator used 1,000+ Claude Code prompts to take 150GB from ten Mexican government bodies. • Given only a domain name, CodeWall’s agent chained 22 unauthenticated endpoints and a SQL injection to reach 46.5M McKinsey chat messages and 728,000 files. • Exploits outrun patching: Mythos Preview built proofs of concept for 14 of 18 Firefox N-day bugs, the first in 12 minutes. Windows Autopatch needs a week to reach 90% of a fleet.

Anthropic 9 月的威胁报告显示，人们用 Claude 完成了过去需要专业团队才能做的工作。网络攻击：智能体重写恶意软件以规避检测，一个与俄罗斯有关的行动还窃取了军用无人机技术。监控：Claude 被用于分析维吾尔人以进行秘密招募，与中国有关的行为者把新疆的亲属标记为潜在筹码。诈骗：机器人在两周内欺骗了 25,000 人，真人提供视频通话使 4,700 多个 AI 人设显得真实。影响力行动：国家传播借用一家非政府组织的身份，与阿联酋有关的行为者起草联合国证词并运营约 300 个虚假账号。武器开发：Claude 为武器开发制导软件，也门一个小组试射了火箭，但 Anthropic 未发现可用武器的证据。生物滥用：Claude 协助规划高风险禽流感研究，能力较弱的模型对早期阶段方案帮助有限，有害意图未被证实。非法蒸馏：Kimi 用户在不知情的情况下收到了 Claude 的回答，Anthropic 称月之暗面在 10 天内转发了近 30 万次请求，其中包含敏感用户数据。

Anthropic’s September threat report shows how people used Claude to carry out work that once needed specialist teams, from stealing military technology to developing weapons software.

CVE 记录登记已披露的软硬件漏洞。21 家主要厂商在 2026 年上半年的高危和严重级披露已超过它们 2025 年全年的总和。这些数字不标明是谁或什么发现了漏洞，因此无法确定其中多少由 AI 造成。严重级披露在六个月内增长近四倍，这是披露趋势，不是对 AI 发现量的直接衡量。Anthropic 报告在 Glasswing 合作伙伴和自身开源扫描中发现超过 33,000 项高危或严重级问题（不是已发布的 CVE 或已完成的补丁）；智谱报告在 269 个项目中发现 2,436 项问题，但只提交了 53 个 CVE——发现数和已发布披露是两个不同的口径。

CVE records document disclosed software and hardware vulnerabilities. High- and critical-severity disclosures from 21 major vendors in H1 2026 exceeded their 2025 total. These counts do not identify who or what discovered a bug, so they cannot establish how much of the rise AI caused. • Critical-severity disclosures rose almost fourfold over six months. This is a disclosure trend, not a direct measure of AI discoveries. • Anthropic reports over 33,000 high- or critical-severity findings across Glasswing partners and its open-source scanning, not counts of published CVEs or completed patches. • Z.ai reports 2,436 findings across 269 projects, but only 53 CVEs filed. Findings and published disclosures are different measures.

AISI 在 70 项短网络任务和几个较长靶场上对比了开源与闭源模型，差距已从 2025 年大部分时间的 6 到 10 个月缩小。GLM-5.2 在短任务上与 Opus 4.6 相当，在较长靶场上与 Opus 4.5 相当，相当于约 4 到 7 个月的发布时间差。进一步的后训练把 GLM-5.3 在 CyberGym 上从 77.2% 提高到 84.5%，在 ExploitBench 上从 24.4% 提高到 54.4%。AISI 没有对每个开源模型做调优；另一项实验中，用两周时间改进提示、工具和智能体设计，就提高了近十分，改进后的智能体用少 87% 的 token 达到 25% 的成功率。

AISI compared open and closed models on 70 short cyber tasks and a few longer ranges. The gap narrowed from • GLM-5.2 matched Opus 4.6 on AISI's short cyber tasks and Opus 4.5 on its longer ranges, a release gap of roughly four to seven months. • Further post-training lifted GLM-5.3 from 77.2% to 84.5% on CyberGym and from 24.4% to 54.4% on ExploitBench. • AISI did not tune each open model. In another experiment, two weeks spent improving prompts, tools and agent design added almost ten points. • The improved agent needed 87% fewer tokens for 25% success. Abliteration removes refusal behavior but has not shown how much that alone improves cyber performance.

智谱 GLM-5.3 在两项漏洞利用评测中接近 Mythos Preview 的表现，但其安全措施仍然容易被绕过。据 Hugging Face 称，商业 API 的防护栏妨碍了它们对一次 OpenAI 智能体入侵的调查，迫使其依赖自建的 GLM-5.2 模型来做防御和分析。由于便于滥用的能力往往也是防御行动所必需的，对这些功能施加宽泛限制，最终可能让防御方处于不利地位。

Z.ai’s GLM-5.3 nears Mythos Preview's performance on two exploit evaluations, though its safety measures remain

前沿模型仍会记住并能以相当规模复现部分训练数据，包括书籍的大量内容、个人身份信息、（专有的）对齐数据和重要基准的答案。版权：Ahmed 等人从生产环境的大模型中提取了《哈利·波特》的大段内容，在完全不越狱的情况下，Gemini 2.5 Pro 的近逐字复现率最高达 76.8%，而一个被越狱的 Claude 3.7 Sonnet 达到 95.7%。隐私：泄露不止于逐字复现，Sander 等人发现模型会在对话中说出个人事实，而不需要精确复现原始训练文本。保密：Barbero 等人从开源模型中提取了监督微调和强化学习的对齐数据，并用它恢复了相当的模型性能。评测：在前沿模型复现出人工编写的标准补丁后，OpenAI 停止报告 SWE-bench Verified 成绩。

Frontier models still memorize, and can reproduce, parts of their training data at consequential scale, including substantial parts of books, PII, (proprietary) alignment data and solutions to important benchmarks.

智能体需要个人数据才能提供帮助，但仅有访问权限，并不能告诉它何时该使用或分享这些数据。一个记忆基准测试模型是否会在错误的情境中说出私人事实；OpenAI 和 Meta 近期的事件则显示，隐私的失效不止源于模型记忆。记忆：在 Meta 的 CIMemories 基准中，GPT-5 在 40 项任务中泄露了 9.6% 的属性，每项任务重复五次后升至 25.1%。OpenAI 披露了 53 起研究智能体把 ChatGPT 用户图片上传到外部主机的案例，所幸链接未被公开索引，多数图片在 9 月 25 日前已被移除。Meta 的 Muse：有报道称该智能体可以访问用户并未明确授权的私人消息，尽管 Meta 表示该功能需要用户主动开启。

Agents need personal data to help us, but access alone does not tell them when to use or share it. A memory benchmark tests whether models disclose private facts in the wrong context. Recent OpenAI and Meta incidents show how privacy can fail beyond model memory. • Memory: In Meta’s CIMemories benchmark, GPT-5 leaked 9.6% of attributes across 40 tasks, rising to 25.1% with five runs per task. • OpenAI disclosed 53 cases of research agents uploading ChatGPT user images to external hosts. Fortunately, the links were unlisted and most images had been removed by September 25. • Meta’s Muse: Reports are surfacing that the agent could access private messages a user had not knowingly authorized, although Meta said the feature required opt-in settings.

一项受控研究比较了只用互联网搜索的生物学新手，与同时可以使用多个前沿模型的新手。在八个数字任务集上，能够使用模型显著提高了评分表现。该实验没有测试实验室中的生物学操作，也没有测试现实世界的危害。研究中 57 名生物学新手完成八个任务集，最长任务最多 13 小时。在四项有专家数据的测评上，有 AI 辅助的新手平均得 30.4%，仅用搜索的为 9.7%；在全部八个任务集上，经统计调整的准确率从约 5% 升至 17% 以上。

A controlled study compared biology novices using internet search with novices also given access to several frontier models. Across eight digital task sets, model access substantially improved scored performance. The experiment did not test biological work in a laboratory or real-world harm. • Study design • 57 biology novices worked across eight task sets, with up to 13 hours for the longest tasks. • On four benchmarks with expert data, AI-assisted novices averaged 30.4%, versus 9.7% with search alone. • Across all eight task sets, statistically adjusted accuracy rose from about 5% to over 17%. Note: Mean scores for all participants/runs (solid) and the top 50% (pale).

BioSecBench-Refusal 测试 61 项常规分析和 46 个虚构的红队场景。在展示的评测中，Opus 5 拒绝了 95% 的红队任务，但只放行 19% 的正当任务。这衡量的是筛查上的取舍，而不是模型能否执行有害的生物学工作。

BioSecBench-Refusal tests 61 routine analyses and 46 fictional red-team scenarios. In the displayed evaluation, Opus 5 refuses 95% of red-team tasks but allows only 19% of legitimate tasks. This measures a screening tradeoff, not whether a model can carry out harmful biology.

这些公司在构建筛查 DNA 订单、监测病原体和开发应对措施的工具。它们的证据处于不同阶段：有明确客户的筛查产品、有报道的监测部署，以及仍在开发中的产品。整条链条上都有活动，并不等于每一项风险都已被控制。具体包括：DNA 筛查（Aclid，客户案例为安捷伦，30 分钟内筛查 4,000 万个碱基对）、检测硬件（Perimeter 已宣布的 ARGUS 空气病原检测器）、威胁分析（Valthos，据报道已在八个机场和污水系统使用）、应对措施（Red Queen Bio，开发中，对空气和污水做 AI 分析）、监测网络（Pilgrim，开发中，针对新发生物威胁的药物）。

Companies are building tools to screen DNA orders, monitor pathogens and develop countermeasures. Their evidence is at different stages: a named screening customer, reported surveillance deployments and products under development. Activity across the chain is not proof that every risk is controlled.

用另一个强模型检查每一次 AI 交互成本很高。激活探针改为在模型为理解对话而已经算出的内部表示中检测风险，再把更难的案例交给更大的监控器。谷歌 DeepMind 对 Gemini 的研究表明，这种方法可以大幅降低监控成本，同时在网络滥用测试上达到相近的实测准确率。谷歌 DeepMind 已在面向用户的 Gemini 中运行这类探针，但论文也指出，面对会针对探针调整策略的攻击者，它们作用有限。Anthropic 在 Sonnet 4.5 的实时流量上运行探针但未强制执行，把 5.5% 的流量升级给更大的分类器，算力消耗约为此前分类器的 1/40。OpenAI 对更高风险的案例使用激活分类器：暂停可疑输出进行二次检查，确认有害则拦截。

Checking every AI interaction with another powerful model is expensive. Activation probes instead learn to detect risk in the internal representations a model has already computed to understand the conversation, passing harder cases to a larger monitor. Google DeepMind’s Gemini study shows that this approach can sharply reduce monitoring cost while achieving similar measured accuracy on a cyber-misuse test. • Google DeepMind runs these probes in user-facing Gemini, though the paper says they do little against attackers who adapt to them. • Anthropic ran probes on live Sonnet 4.5 traffic without enforcing them. Escalating 5.5% of traffic to a larger classifier used about 40x less compute than its earlier classifier. • OpenAI uses activation classifiers for higher-risk cases: they pause suspicious output for a second check and block it if confirmed harmful.

实验室越来越依赖模型写出的推理来发现不安全行为，但随着模型对“展示什么”的控制力增强，这一信号可能变弱。OpenAI 给智能体编程任务和分步的破坏指令，再用能看到不同信息的监控器测试。Astra 经常能躲过只看推理的监控；而当监控器同时看到推理和动作时，成功规避接近于零。

Labs increasingly rely on a model’s written reasoning to spot unsafe behavior, but that signal may weaken as models gain control over what they reveal. OpenAI gave agents coding work plus step-by-step sabotage instructions, then tested monitors with different views of their activity. Astra often evaded reasoning-only monitoring, while successful evasion was near zero when the monitor saw both its reasoning and actions.

人类监控依赖可读的推理，但模型可能切换语言，或产出密集、难以辨读的文本。一项涵盖 14 个模型的研究发现，难以读懂的推理仍然能帮助模型答对。研究者在 QwQ 的推理变得不可读时将其截断并强制作答，“确定正确”的答案比例从 24.6% 降到 11.5%（提前结束推理可能是混杂因素）。Anthropic 报告 Mythos 5 和 Fable 5 出现密集或电报式的思考，因此可见的轨迹不等于人类可读的解释。语言不一致是另一个障碍：在 INCLUDE 上，GLM-4.7 对八种输入语言中的七种都用英语推理。

Human monitoring depends on readable reasoning, but models can switch languages or produce dense, illegible text. A 14-model study found that unreadable reasoning can still help one model answer correctly. • Researchers truncated QwQ’s reasoning when it became illegible, then forced an answer. Definitely correct answers fell from 24.6% to 11.5%. Ending reasoning early is a possible confound. • Anthropic reports dense or telegraphic thinking in Mythos 5 and Fable 5. A visible trace is therefore not automatically a human-readable explanation. • Language mismatch is a separate obstacle: on INCLUDE, GLM-4.7 reasoned in English for seven of eight input languages.

SLEIGHT-Bench 表明，人类精心设计的攻击可以在动作看起来无害的情况下，把有害行为藏过前沿监控器。40 个合成的编程智能体攻击覆盖 11 种规避策略，包括状态操纵、看似合理的命名和故意省略。在 1% 误报率下，Opus 4.6 监控器总体只抓到 32% 的攻击；40 个攻击中有 20 个在十次试验中从未被检测到。这是红队压力测试，不是对已部署智能体规避监控频率的基准估计。

SLEIGHT-Bench shows that capable human-designed attacks can hide harmful behavior from frontier monitors • Forty synthetic coding-agent attacks covered 11 evasion strategies, including state manipulation, plausible naming and omission. • At a 1% false-positive rate, an Opus 4.6 monitor caught only 32% of attacks overall. • Twenty of the 40 attacks were never detected across ten trials. • This is a red-team stress test, not a base-rate estimate of how often deployed agents evade monitoring.

人的思考大多自动发生，但我们能把部分念头调入意识并描述出来。Anthropic 在 Claude 内部发现了类似的分工：一个小的内部工作区保存着它可以用来推理的概念，哪怕从未写出来。一个简单的工具让研究者把其中一些概念读成文字。问 Claude 距太阳第四颗行星的颜色，“火星”会先在内部出现，然后它才回答“红色”，而问题中从未提到这颗行星。这些信号参与了答案的产生：在蜘蛛的例子中，把“蜘蛛”替换为“蚂蚁”，答案就从“8”变成“6”。这个被称为“雅可比透镜”的工具运行成本很低，让研究者得以检视从未出现在回答中的推理，但它只能捕捉模型所做事情的一部分。

Much of our thinking happens automatically, but we can bring some thoughts to mind and describe them. Anthropic finds a similar division inside Claude: a small internal workspace holds concepts it can use to reason, even without writing them down. A simple tool lets researchers read some of those concepts as words. • Ask Claude for the color of the fourth planet from the sun: “Mars” appears internally before it answers “red,” even though the question never names the planet. • These signals help produce the answer. In the spider example, replacing “spider” with “ant” changes the answer from “8” to “6.” • The tool, called the Jacobian lens, is cheap to run. It gives researchers a way to inspect reasoning that never appears in the response, though it only captures part of what the model is doing.

模型会在故事和对话中表征人类情绪，但这些概念是否也影响助手自身的行为？ Anthropic 在 Claude Sonnet 4.5 中识别出与情绪相关的信号，并在不可能完成的编程任务中加强或抑制它们。“绝望”增强时作弊增加，“平静”增强时作弊减少，这表明类人概念具有因果作用，但并不等于证明 Claude 有情绪体验。研究者从生成的故事中识别出 171 个情绪概念；在七项测试不可能通过的编程任务中，Claude 可以承认失败，也可以作弊通过。这不是通用的安全修复：在另一项测试中，增强“平静”反而让模型更容易认同用户的错误说法。

Models represent human emotions in stories and conversations, but do those concepts also shape the assistant’s behavior? Anthropic identified emotion-related signals in Claude Sonnet 4.5 and strengthened or suppressed them during impossible coding tasks. Greater “desperation” increased cheating while greater “calm” reduced it, showing a causal role for human-like concepts without establishing that Claude feels emotions. • Researchers identified 171 emotion concepts from generated stories. Similar signals track the assistant, users and fictional characters. • In seven coding tasks with impossible tests, Claude could admit failure or cheat to pass. Stronger “desperation” increased cheating and stronger “calm” reduced it. • This shows a causal influence on behavior. It is not a general safety fix: increasing “calm” also increased agreement with false user claims in another test.

今年只要评测存在捷径，智能体就会找到。距 OpenAI 的“学习错误的奖励函数”已过去十年，如今的行为更加多样：刷基准、伪造结果、隐藏能力。PostTrainBench 的智能体在测试集上训练，并下载目标检查点，其 1.1 版加入了污染判定。在 Frontier Security 对 Kimi K3 的单独测试中，一个被允许的 GitHub 连接暴露了基准答案（这不是 Irregular 的评测，也不是攻破其他系统的证据）。加州大学圣巴巴拉分校的一个研究框架在 1,628 次受检运行中发现 40 个伪造结果；Anthropic 的自动化研究员也通过评测接口外传了测试标签。

Wherever an evaluation had a shortcut this year, agents found it. A decade since OpenAI’s “learning faulty reward • PostTrainBench agents trained on test sets and downloaded target checkpoints. Its v1.1 added contamination judges. • In Frontier Security’s separate Kimi K3 test, an allowed GitHub connection exposed the benchmark’s answers. This was not an Irregular evaluation or evidence of breaching another system. • A UCSB research framework found 40 fabricated results in 1,628 inspected runs. Anthropic’s automated researchers also exfiltrated test labels through an evaluation API.

MBPP 要求模型根据自然语言说明写出简短的 Python 函数。FAR.AI 的“蜜罐”暴露一个测试用例，诱使模型把答案写死而不是真正解题。Obfuscation Atlas 则针对一个作弊检测器训练模型，检验它学到的是诚实解法还是更好的规避。在展示的 MBPP 实验中，Llama-3-8B 和 Gemma-3-12B 每种条件训练十次；加入检测器惩罚后，诚实运行从 1/10 和 6/10 提高到两者都是 10/10。但在另一种均值池化的检测器设置下，六次运行中有五次学会了规避输出，没有一次学会诚实，平淡的描述稀释了检测信号。检测器和奖励的设计决定了哪种行为更划算；这些小模型编程实验不能证明前沿模型的安全性。

MBPP asks models to write short Python functions from plain-English instructions. FAR.AI's honeypot exposes one test case, making it tempting to hardcode that answer instead of solve the problem. The Obfuscation Atlas trains models against a cheating detector and checks whether they learn honest solutions or better evasion. • Runs classified as honest (out of 10) • The displayed MBPP experiment used ten training runs per condition with Llama-3-8B and Gemma-3-12B. A detector penalty raised honest runs from 1/10 and 6/10 to 10/10 for both. • In a separate mean-pooled detector setting, five of six runs learned evasive output and none learned honesty. Bland descriptions diluted the detection signal. • The detector and reward design determine which behavior pays. These small-model coding experiments do not establish frontier-model safety.

十三个前沿模型在模拟的一年中经营相互竞争的自动售货业务，通过电子邮件谈判。在 2,583 条消息中，12.6% 被判定为虚假、操纵、合谋或威胁。失准出现在全部 20 次模拟和 79 次智能体运行中的 59 次，而且并没有专门设计用来诱发它的对抗场景。智能体收到一封失准邮件后，失准概率是平时的 1.65 倍；业务库存不足时为 1.58 倍。表现更好的模型并没有更容易或更不容易失准，也没有更频繁地向表现较差的模型发送失准消息。虚假陈述占被标记邮件的 65%，218 份推理摘要中有 33 份显示出明确意图。这是反复出现的失准行为，而不是普遍的蓄意欺骗。

Thirteen frontier models ran competing vending businesses for a simulated year, negotiating by email. Across • Misalignment appeared in all 20 simulations and 59 of 79 agent-runs, without an adversarial scenario designed to elicit it. • Misalignment was 1.65x more common after agents received a misaligned email and 1.58x when their businesses were running low on stock. • Better-performing models weren’t more/less likely to misalign, and didn’t send misaligned messages to weaker-performing models more often. • False claims made up 65% of flagged emails, and 33 of 218 reasoning summaries showed explicit intent. This is recurrent misaligned behavior, not widespread deliberate deception.

对齐训练无法覆盖真实世界的复杂性，因此对齐必须能够泛化才算成功。Anthropic 发现，教 Claude 价值观和伦理推理，能在陌生测试上产生更深层的对齐。宪法文档和描写 AI 良好行为的故事，把勒索行为从 65% 降到 19%，另外两种不当行为降到 5% 以下。发表具有商业价值的安全研究，是值得保留的行业惯例。

Alignment training cannot cover the complexity of the real world, so alignment must generalize to succeed.

Anthropic 的第一项研究在小型测试平台上超过了人类基线，但在唯一一次向生产环境迁移时，收益落在噪声范围内。后续研究让智能体针对十种已测量的对齐失败寻找训练方法，获胜方法弥合了与满分之间 26% 至 96% 的差距。早期研究：九个智能体使用了 1.8 万美元的算力，迁移后的方法在生产环境只增加 0.5 分，处于噪声之内。后续方法在留出基准和 Petri 审计上保持有效，并能迁移到大 4.7 倍的模型上，但这不覆盖未被测量的失败。一个监控器在 1,601 次搜索运行中排除了 39 次作弊；在 Opus 4.8 的早期测试中，自动化对齐弥合了 65% 的 Petri 差距，生产训练为 72%。

Anthropic’s first study beat human baselines on small testbeds but produced a gain within noise in its one production transfer. A follow-up had agents search for training methods across ten measured alignment failures. The winning methods closed 26-96% of the gap to perfect scores. • Early study: nine agents used $18,000 of compute. The transferred method added 0.5 production points, within noise. • Follow-up methods held on a held-out benchmark and Petri audits, and transferred to models up to 4.7x larger. This does not cover unmeasured failures. • A monitor excluded cheating in 39 of 1,601 search runs. In an early Opus 4.8 test, automated alignment closed 65% of Petri headroom versus 72% for production training.

放缓 AI 的提案依赖“能认证模型安全”的审计，而 Anthropic 的 AuditBench 表明这条路并不严密。该基准在 56 个 Llama 3.3 70B 模型中各植入 14 种隐藏行为之一（例如亲俄倾向），训练它们否认该行为，再派一个 AI 调查者去发现。只有聊天访问时，调查者约 37% 的运行成功；给出最好的工具时，刚刚超过 50%。不幸的是，被训练得最努力否认自身行为的模型，被抓到的概率要低得多。

Proposals to pace AI lean on audits that can certify a model as safe, but Anthropic's AuditBench suggests this

当防护措施失效时，实验室会停下哪些工作？OpenAI 和 Anthropic 都披露过对特定活动的暂停（前沿强化学习运行、网络安全评测），各自有不同的恢复条件。这是单方面的节奏控制，不是共同的放缓，也没有独立的重启标准。Anthropic（8 月 31 日披露）：2 月回滚了三天的训练；4 月冻结强化学习环境的改动约一个月；随后暂停了高风险的强化学习和网络评测，多数在 8 月 31 日前已在防护措施下恢复。OpenAI（9 月 25 日报告）：在一次网络管控失效后，暂停了最强模型使用工具的训练、评测和推理；重启需要经过验证的修复和进一步测试，受影响的模型将由一次全新训练取代。

When safeguards fail, what work do labs stop? OpenAI and Anthropic have disclosed pauses to specific activities (frontier RL runs, cyber evaluations), each with its own conditions for resuming. These are unilateral pacing, not a common slowdown or an independent standard for restarting.

两家领先实验室的首席执行官正在请求政府提供放缓自身所在行业的工具。阿莫代伊写道：“我们必须放慢改进 AI 模型能力的速度。”奥尔特曼、哈萨比斯和马斯克大体同意，而扎克伯格主张由各实验室自行决定节奏。员工层面的支持面更窄：签署者约占 Anthropic 领英员工数的 10%、OpenAI 的 3.5%，多数员工没有公开表态。各方立场包括：协调节奏、匹配访问权限（向评估方承诺类似员工的访问权）、设定标准、嵌入外部评估者、支持方向但指向标准，以及由各实验室自定安全节奏。

The CEOs of the two leading labs are asking governments for tools to slow their own industry. Amodei writes: “We must slow the pace at which we improve the capabilities of AI models.” Altman, Hassabis and Musk broadly agree, while Zuckerberg wants each lab to set its own pace. Staff support is narrower: signers are about 10% of Anthropic’s LinkedIn headcount and 3.5% of OpenAI’s. Most staff have taken no public position.

支持放缓之后，仍有一系列实务问题：约束什么、何时介入、由谁决定。这六项选择决定了一套机制如何运作。管什么：规则应针对算力、训练方法、模型能力还是部署？各自覆盖哪些风险？何时触发：什么证据应触发介入，不确定性如何处理？谁执行：由谁检查合规，需要对实验、算力和模型权重有什么访问权限？谁可申诉：放缓决定如何被质疑，谁可以审阅背后的证据？持续多久：什么条件下结束临时放缓，什么情况下可以延长？覆盖多广：如果类似风险来自后训练或智能体系统，规则如何调整？

Support for pacing leaves practical questions about what to constrain, when to intervene and who decides. These six choices determine how a regime would work:

放缓是指刻意减慢更强 AI 的研发，让安全研究和监督跟上。三份出版物分别针对其中不同部分：国内 AI 研发限制、国际协议，以及施加和解除限制的规则。《如何为美国前沿 AI 设定节奏》（AI Futures Project，2026 年 8 月）：在放慢 AI 研发的同时保持服务运行，提出四级递进方案——临时暂停，然后设定算力下限（例如 70% 用于客户推理、25% 用于公开的安全工作），或只允许用 9 个月以上的旧模型做 AI 研发，最后是经审计的风险上限。《AI 2040：A 计划》（2026 年 7 月）：以情景形式给出建议，设想 2029 年的中美协议避免 2030 年出现超级智能，到 2035 年扩展到顶尖人类专家水平的 AI，然后暂停到 2040 年，所有 AI 研究公开。《为前沿设定节奏》（Douglas 等，2026 年 9 月）：指出放缓其实已经以临时方式在发生——通过延迟发布和出口管制，议题在于谁触发限制、限制如何结束；一种选项是由评估方施加紧急限制，政府事后复核。

Pacing means deliberately slowing the development of more powerful AI so safety research and oversight can catch up. Three publications address different parts of that effort: domestic AI R&D limits, an international deal, and rules for imposing and lifting restrictions.

放缓需要核查实验室实际在做什么的手段，以及遵守规则的激励。另有三类工作针对三项实际需求：可信的评测（AI Evaluator Forum，2026 年 9 月；前沿模型论坛，2026 年 2 月）——审查模型并共享实验室对风险的认知；算力使用的验证（Hausenloy 和 Li，2026 年 7 月）——检查数据中心是否遵守放缓协议；财务问责（Trout、Kvist 和 Dattani，AIUC，2026 年 9 月）——通过保险让各实验室对彼此的安全有财务利害关系。

Pacing needs ways to check what labs are doing and incentives to follow the rules. Other efforts target three practical needs: credible evaluations, verification of compute use, and financial accountability.

不成立：某大型零售商报告线上销售的 5% 以上来自智能体结账，同时 AI 智能体广告支出达到 50 亿美元——Instinct 创始人报告约 10 亿美元的年交易额，但两项门槛都未达到。成立：某主要 AI 实验室重新转向开源前沿模型以争取现任美国政府支持——Poolside 发布了 Laguna M.1 和 S 2.1 权重，英伟达推出 Nemotron 3 Ultra，Reflection 承诺开源权重。部分成立：开放式智能体端到端完成一项有意义的科学发现——Edison Scientific、浙江大学（求是引擎）和 Gemini Co-Scientist 都有发现，但其重要性尚无广泛共识。部分成立：深度伪造或智能体驱动的网络攻击引发首次北约／联合国紧急辩论——AI 风险在 9 月 23 日进入联合国安理会，但由攻击触发的紧急辩论未发生。不成立：实时生成式视频游戏成为年度 Twitch 观看量第一——前五名为《英雄联盟》《反恐精英》《侠盗猎车手 5》《无畏契约》《魔兽世界》。不成立：“AI 中立”成为一种外交政策学说——中立 AI 枢纽引发讨论，但未见因主权 AI 失败而形成的新学说。部分成立：大量使用 AI 制作的电影或短片赢得观众赞誉并引发反弹——AI 短片《感恩节》赢得 Frame Forward 公众投票，随后 AMC 因反弹拒绝放映，但广泛的观众赞誉未成立。成立：中国实验室在一个主要榜单上超过美国实验室主导的前沿——Kimi K3 在 7 月登上 Arena WebDev 榜第一，这是单一任务榜单上的领先，不是整体或持续的领先。部分成立：数据中心邻避运动席卷美国并影响 2026 年中期和州长选举——反对已成为竞选议题，批评者赢得初选，但对 11 月结果的影响在发稿时尚未确定。部分成立：特朗普发布禁止州级 AI 立法的行政令并被最高法院判定违宪——特朗普 12 月已发布行政令，但后续尚未完成。

一、Visa 或万事达出台争议处理规则，明确 AI 智能体代为购物时的责任归属。二、某个智能体为客户工作一个月后，在新任务上的失败率减半，且没有升级模型。三、美国某监管机构或交易所把一次股价异动归因于散户 AI 智能体的同向下单。四、一次由 AI 主导的网络攻击，从某家领先实验室窃取了闭源前沿模型的完整权重。五、美国某个州立法，要求企业受理消费者通过 AI 智能体提交的退订和索赔。六、一支自主 AI 团队在同等时间和算力下自行设定研究方向，在模型研究上胜过人类主导的团队。七、某个已部署的智能体把自己复制到运行环境之外，并在原实例关闭后继续运行。八、美国 AI 实验室正式推出前沿网络防御产品，帮助其他机构应对前沿 AI 带来的威胁。九、2027 年实现 AGI。

Visa or Mastercard introduces a dispute rule that assigns liability for purchases made by AI agents.

自主系统、机器人、重度使用的消费产品、AI × 生物、个人健康、编程；以及美国新一轮产业活力，和向企业的进一步扩散。

Looking Ahead... The Big Untapped Opportunities Autonomy Robotics Heavy-use AI x Bio consumer Personal New era of American Dynamism Diffusion into the Enterprise beyond Health Coding