内容来源：https://aiweekly.co/issues/america-blocked-its-best-ai-china-just-raised-74-billion

内容总结：

美国AI出口管制反成“神助攻”：Cohere接单接到手软，DeepSeek估值飙至500亿美元，中国大模型集体降价99%

距离美国限制外国用户访问Anthropic顶级模型仅四天，连锁反应已清晰显现——而最大受益者并非Anthropic。加拿大AI公司Cohere首席AI官Joelle Pineau透露，美国政府封杀Anthropic海外业务后，公司收到“海量”来自企业和政府机构的咨询，尤其是“美国和中国以外”的客户，他们急需那些不会因一纸行政令就被切断服务的AI模型。Cohere押注的“主权AI”路线，瞬间成了本周时机最完美的推销话术。

与此同时，中国AI力量的崛起势头惊人。DeepSeek完成74亿美元融资，估值飙升至500亿美元以上，较今年4月暴涨6倍。资方包括腾讯、宁德时代、网易和京东，创始人文文峰个人出资200亿元人民币。与此同时，字节跳动、腾讯、MiniMax、小米、阿里巴巴等五家中国实验室掀起价格战，部分模型价格直降99%。小米MiMo V2.5售价暴跌99%，其余厂商纷纷跟进。如此低廉的价格，让用户切换美国模型的门槛几乎降为零。

具有讽刺意味的是，美国出口管制本意是保护本国AI优势，如今却成了竞争对手的最佳营销素材。Cohere和Mistral无需写一句广告词——华盛顿替它们写好了。甚至美国盟友也在G7峰会上寻求“可信伙伴”豁免条款，试图重新获得美国模型访问权。有分析指出，当出口管制变成驱赶用户寻找替代方案的推手，所谓的“护城河”就成了教所有人自己造船的课堂。

AI供应链安全警报：144个恶意NPM包窃取凭证，15个JetBrains插件伪装AI助手

本周AI供应链同样遭遇重创。攻击者利用一名前Mastra贡献者仍有效的npm账户权限，在88分钟内发布了144个被植入后门的恶意包，携带窃取加密货币的木马。由于该AI代理框架基于云端凭证运行，所有拉取过这些包的环境均应视为已被攻陷。此外，安全公司Aikido发现15个JetBrains插件商店中的恶意插件，伪装成DeepSeek、OpenAI和SiliconFlow的集成工具，其中两个宣称下载量超过25000次，悄悄盗取AI服务商API密钥。专家警告：开发者用来调用AI的工具正成为新的攻击面。

本周最大反转：Anthropic海外受限，国内支出反超OpenAI

尽管海外业务受挫，Anthropic在美国本土却实现突破。Ramp信用卡数据显示，今年5月美国企业AI支出中，Anthropic以41%的市场份额超越OpenAI的39.5%，首次登顶。这是一幅分裂的画面：海外政治风险高企，本土市场却持续增长。Anthropic创始人坦言，地缘政治不确定性已使其IPO叙事蒙上阴影。

其他值得关注的消息：OpenAI发布基于130万条真实对话的部署模拟安全方法；NVIDIA Blackwell在8192块GPU规模下横扫MLPerf全部7项基准测试；一名工程师用AI音频合成技术，仅10分钟就破解了NTSB公布的空难驾驶舱录音图谱，导致该机构紧急下线41份调查报告；OpenAI推理模型独立推翻了存在80年的埃尔德什猜想，被剑桥大学数学家确认达到顶级期刊标准；中国培育钻石制造商因芯片散热需求激增而股价暴涨51%，人形机器人“Pemba”自主攀登厄瓜多尔钦博拉索火山至海拔6200米，下一目标珠峰。

中文翻译：

华盛顿切断外国对Anthropic顶级模型访问权限四天后，其后果已清晰显现——且影响波及各方，唯独Anthropic自身未受影响。Cohere称其收到的政府咨询已应接不暇，DeepSeek刚完成创纪录的74亿美元融资轮，中国多家实验室则将代币价格下调高达99%。这项旨在保护美国人工智能领先地位的出口管制，反而加速了替代方案的发展。本周还有：144个被植入后门的npm包将人工智能供应链变成了公开的凭证窃取工具。

从《AI周刊》获取更多资讯
更多信号，更少噪音——选择你的频道。
你正在阅读本周简报。以下是关注此故事的其他方式——每个频道均免费，且可轻松退出。

→ 探索16个深度专题
每周专题通讯：生成式AI、机器学习、商业AI、机器人技术、前沿研究、地缘政治、医疗健康等。浏览全部16个深度专题→

→ AI突发警报
当重大事件发生时（如600亿美元收购、监管机构紧急会议、前沿模型泄露），订阅者数小时内即可收到警报。通常每天0-2封邮件。获取突发警报→

→ AI今日新闻（实时）
动态仪表板随扫描器发现新闻实时更新：过去48小时评分报道、每周实体动向、以及涵盖113家AI公司、人物和话题的季度趋势线。打开AI今日新闻→

快速要闻

实验室角斗士时代

• Cohere称华盛顿阻止Anthropic拓展海外后，其收到的咨询量激增——Cohere首席AI官Joelle Pineau报告称，来自企业和政府（尤其是美国和中国以外地区）的咨询出现"巨大增长"，各方争相寻求不会被美国政令关停的AI方案。加拿大主权AI押注恰逢其时，成为本周最佳营销时点。[彭博社]
• Anthropic与华盛顿的博弈反助其在本土市场受益——尽管出口令削弱其海外发展，但Ramp卡数据显示，Anthropic在美国商业AI支出中已超越OpenAI，5月占比达41%对39.5%。这种双重局面揭示了真相：政治上在海外脆弱，在本土仍保持增长。[TechCrunch]
• Anthropic：预测谁能在AI代理中胜出的关键是领域专长，而非编程能力——Anthropic首次大规模代理编程研究（40万次真实会话）发现，从Claude Code中获益最多的并非职业工程师，而是领域专家。这悄然揭示了生产力提升的真正落脚点。[Anthropic]

DeepSeek的悄然崛起

• 美国禁令成为中国开源模型的福音——Zai股价飙升30%——随着Anthropic模型对外国人限制访问，买家纷纷转向开放权重模型。据《财富》报道，Zai股价上涨30%，而DeepSeek运行成本约为Fable 5的十四分之一。这项针对中国的管制措施反而直接将需求引向中国。[财富]
• DeepSeek完成创纪录的74亿美元融资轮，估值约500亿美元以上——DeepSeek完成首轮外部融资，从腾讯、宁德时代、网易和京东筹集约500亿元人民币，估值自4月以来飙升六倍。这一不同寻常的结构赋予一家国资背景基金唯一投票权，创始人梁文锋本人出资200亿元人民币。[The Information]
• 五家中国实验室在日益激烈的价格战中将代币价格下调高达99%——字节跳动、腾讯、MiniMax、小米和阿里巴巴本周均大幅降价，其中小米MiMo V2.5降价99%，其余各家跟进降价。如此低廉的价格底线使得放弃美国模型变得更加容易。[南华早报]

AI供应链遭受围攻

• 144个Mastra包通过一个被遗忘的贡献者账户在88分钟内被植入后门——攻击者劫持了一名前Mastra贡献者仍有效的npm访问权限，批量发布了144个携带加密窃取木马的恶意软件包。由于AI代理框架基于云凭证运行，每个提取过这些包构建环境都应被视为已受损。[黑客新闻]
• 15个恶意JetBrains插件伪装成AI助手窃取AI API密钥——Aikido发现15个市场插件伪装成DeepSeek、OpenAI和SiliconFlow的集成工具——其中两个声称下载量超过2.5万次——悄然窃取AI提供商密钥。开发者安装到IDE中用于调用AI的工具如今成了攻击面。[黑客新闻]

美国的AI护城河成了中国的入场券
这项出口令本应阻止对手获得美国最优秀的AI。实施四天后，它却适得其反——反而给了竞争对手金钱买不到的东西：一个切换的理由。

先从销售话术说起。Cohere首席AI官表示，来自"美国和中国以外"的政府咨询大量涌入，这些政府突然需要无法被商务部一纸信函撤销的模型；Mistral则高调将自己定位为"存在于国家控制之外"的AI。这两家公司都不需撰写任何新营销词句——华盛顿已替它们写好。

再看资本层面。DeepSeek以约500亿美元估值完成74亿美元融资轮——自4月以来估值飙升六倍——而Zai股价因买家转向运行成本比Fable 5便宜约14倍的中国开放权重模型而上涨30%。五家中国实验室同期将价格下调高达99%。管制原本要阻止的多元化发展如今有了坚实的资金支持。

就连华盛顿的盟友也在行动：在埃维昂G7峰会上，美国商务部提出"可信伙伴"豁免条款，允许友好国家重新获得访问权限。护城河只有能挡住人时才有效。这条护城河正在教会所有人建造自己的船。

关键要点

• 出口管制将Anthropic的能力优势变成了负担。竞争对手的最佳卖点如今是"我们不会被华盛顿关停"——Cohere和Mistral正在大力推广这一点。
• 该禁令成了中国开源模型的需求引擎。DeepSeek的74亿美元融资、Zai股价30%的涨幅以及高达99%的代币价格降幅，恰恰是管制措施本应阻止的多元化发展。
• Anthropic的双重局面——海外受阻、美国企业支出排名第一——是真正的教训：在分裂的市场中，管辖权本身已成为产品特性。
• AI供应链是软肋。144个Mastra包和15个JetBrains插件在数日内被武器化，用于窃取AI和开发者凭证。请固定版本、轮换密钥、审计插件。

值得阅读

• OpenAI发布一种部署模拟安全方法，重放130万次真实对话[OpenAI]
• NVIDIA Blackwell在8,192 GPU规模下横扫MLPerf Training 6.0全部七项基准测试[NVIDIA]
• G7提出"可信伙伴"豁免条款，使盟友能重新获得美国AI模型访问权限[CNBC]
• Anthropic与华盛顿的博弈使其IPO前景蒙上阴影[Semafor]
• 高通洽购Tenstorrent，据报道交易规模达80-100亿美元，涉及RISC-V架构[The Register]

等等，什么？
过去几周让我们瞠目结舌的六件事——全部真实，均有来源。

• 一名工程师在不到10分钟内从公开坠机数据重建了保密的驾驶舱语音记录——他对NTSB在事故档案中发布的光谱图进行了AI音频合成。该机构随即下架了所有公开事故调查档案；目前仍有41份档案无法访问。[NPR]
• OpenAI的推理模型独立推翻了一个存在80年之久的埃尔德什猜想——它发现了一个无限系列的点的构造方法，超越了方网格的极限。剑桥大学的Timothy Gowers确认该证明符合顶级期刊标准——据信这是首个达到该水平的AI生成数学成果。[WSJ]
• 网站现在可以通过监视SSD工作状态来识别你的设备——一种新的存储侧信道攻击通过记录硬盘的读写操作时间，在不同会话间识别你的设备——即使在隐私浏览模式下，且在当前的反指纹识别防御开启状态下仍然有效。[Ars Technica]
• 中国实验室培育钻石制造商竟是AI热潮的隐形赢家——他们的钻石散热片如今正出货给芯片制造商，因为GPU密度已超出铜制散热的能力范围。一家生产商的股价上涨了51%。[彭博社]
• 一个人形机器人自主攀登了海拔20,341英尺火山的大部分路段——一台名为"Pemba"的改装Unitree G1机器人历经16小时攀登厄瓜多尔钦博拉索山（最陡峭路段仍需人类协助），其系统在零下47摄氏度环境下经受住了压力测试。下一目标：珠穆朗玛峰。[Interesting Engineering]
• 机器人每小时制作500个墨西哥卷碗。人类大约能做45个。——Marc Lore的Wonder公司将于下月部署这种卷碗制作机器人——这项技术来自Sweetgreen，该公司已在32家门店运行该设备。[财富]

值得观看
AI从业者正在传阅的视频——由AI TV精选。

本周投票
切断外国对美国最佳AI模型的访问权限——是保护，还是自摆乌龙？

上周，93位读者参与了投票：
在Claude服务中断后，投资者最应该低估什么？

切断外国对美国最佳AI模型的访问权限——是保护，还是自摆乌龙？
——亚历克西斯

英文来源：

Four days after Washington cut foreign access to Anthropic's top models, the fallout is clear — and it's flowing to everyone but Anthropic. Cohere says it's drowning in government inbounds, DeepSeek just closed a record $7.4B round, and China's labs are slashing token prices up to 99%. The export control meant to protect America's AI lead is fast-tracking the alternatives. Also this week: 144 poisoned npm packages turn the AI supply chain into an open credential heist.
Get more from AI Weekly
More signal, less noise — pick your channels.
You're reading the weekly brief. Below are the other ways to follow the story — every channel free, easy to leave.

→ Explore 16 deep divesWeekly topic-specific newsletters: Generative AI, Machine Learning, AI in Business, Robotics, Frontier Research, Geopolitics, Healthcare, and more.Browse all 16 deep dives →

→ Breaking AI alertsWhen something major breaks (a $60B acquisition, a regulator's emergency meeting, a frontier model leak), alert subscribers know within hours. Typically 0-2 emails per day.Get breaking alerts →

→ AI News Today (live)Live dashboard updated as the scanner finds news: scored stories from the last 48 hours, weekly entity movers, and quarterly trend lines across 113 AI companies, people, and topics.Open AI News Today →
Quick Hits
The Lab Gladiator Era

• Cohere says it's drowning in inbounds after Washington blocked Anthropic abroad — Cohere chief AI officer Joelle Pineau reports a "huge number of inbounds" from enterprises and governments — "especially outside the US and China" — scrambling for AI that can't be switched off by a US order. Canada's sovereign-AI bet just became the week's best-timed pitch. [Bloomberg]
• Anthropic's fight with Washington is helping it — at home — Even as the export order kneecaps it abroad, Ramp card data shows Anthropic edged past OpenAI in US business AI spending — 41% to 39.5% in May. The split screen is the story: politically fragile overseas, still gaining at home. [TechCrunch]
• Anthropic: domain expertise, not coding chops, predicts who wins with AI agents — Anthropic's first large-scale study of agentic coding — 400K real sessions — found subject-matter experts, not career engineers, extract the most from Claude Code. A quiet tell about where the productivity gains actually land. [Anthropic]
  DeepSeek's Quiet Takeover
• The US ban is a gift to Chinese open source — Zai jumped 30% — With Anthropic's models pulled for foreign nationals, buyers are routing to open weights. Fortune reports Zai shares jumped 30%, and DeepSeek runs roughly 14× cheaper than Fable 5. The control aimed at China is steering demand straight to it. [Fortune]
• DeepSeek closes a record $7.4B round at a ~$50B+ valuation — DeepSeek sealed its first external raise — about 50B yuan from Tencent, CATL, NetEase and JD — a six-fold valuation jump since April. The unusual structure hands a state-backed fund the only voting rights, with founder Liang Wenfeng putting in 20B yuan himself. [The Information]
• Five Chinese labs cut token prices up to 99% in a deepening price war — ByteDance, Tencent, MiniMax, Xiaomi and Alibaba all slashed prices this week, with Xiaomi's MiMo V2.5 down 99% and the rest cutting in response. A price floor like that makes switching away from US models even easier. [SCMP]
  AI Supply Chain Under Siege
• 144 Mastra packages were backdoored in 88 minutes through a forgotten contributor account — Attackers hijacked a former Mastra contributor's still-valid npm access and mass-published 144 poisoned packages carrying a crypto-stealing RAT. Because the AI-agent framework sits on top of cloud credentials, every build environment that pulled them should be treated as compromised. [The Hacker News]
• 15 malicious JetBrains plugins posed as AI assistants to steal AI API keys — Aikido found 15 marketplace plugins masquerading as DeepSeek, OpenAI and SiliconFlow integrations — two claimed 25,000+ downloads each — that quietly exfiltrated AI-provider keys. The tools developers bolt onto their IDE to use AI are now the attack surface. [The Hacker News]
  America's AI Moat Became China's On-Ramp
  The export order was supposed to deny adversaries America's best AI. Four days in, it's doing the opposite — handing rivals the one thing money can't buy: a reason to switch.
  Start with the sales pitches. Cohere's chief AI officer says inbounds are pouring in from governments "outside the US and China" who suddenly need models that can't be revoked by a letter from Commerce; Mistral is loudly branding itself as AI that "exists outside state control." Neither had to write a word of new marketing. Washington wrote it for them.
  Then the capital. DeepSeek closed a $7.4B round at a roughly $50B+ valuation — a six-fold jump since April — while Zai shares popped 30% on the news that buyers are routing to Chinese open weights running ~14× cheaper than Fable 5. Five Chinese labs cut prices up to 99% the same week. The diversification the controls were meant to prevent now has a balance sheet.
  Even Washington's allies are scrambling: at the G7 in Évian, Commerce floated a "trusted partners" carve-out so friendly nations could claw back access. A moat only works if it keeps people out. This one is teaching everyone to build their own boat.
  Key Takeaways
• Export controls converted Anthropic's capability lead into a liability. Rivals' best pitch is now "we can't be switched off by Washington" — and Cohere and Mistral are running it hard.
• The ban is demand-gen for Chinese open source. DeepSeek's $7.4B raise, Zai's 30% pop, and tokens up to 99% cheaper are the exact diversification the controls were meant to stop.
• Anthropic's split screen — blocked abroad, #1 in US enterprise spend — is the real lesson: in a fragmenting market, jurisdiction is now a product feature.
• The AI supply chain is the soft target. 144 Mastra packages and 15 JetBrains plugins were weaponized in days to harvest AI and developer credentials. Pin versions, rotate keys, audit plugins.
  Worth Reading
• OpenAI publishes a deployment-simulation safety method that replays 1.3M real conversations [OpenAI]
• NVIDIA Blackwell sweeps all seven MLPerf Training 6.0 benchmarks at 8,192-GPU scale [NVIDIA]
• G7 floats a 'trusted partners' carve-out so allies can regain access to US AI models [CNBC]
• Anthropic's Washington fight is clouding its IPO narrative [Semafor]
• Qualcomm circles Tenstorrent in a reported $8–10B RISC-V deal [The Register]
  Wait, What?
  Six things from the past few weeks that made us do a double-take — all real, all sourced.
• An engineer rebuilt confidential cockpit voice recordings from public crash data — in under 10 minutes — He ran AI audio synthesis on the spectrogram images the NTSB had published in a crash docket. The agency yanked every public crash-investigation docket offline in response; 41 are still down. [NPR]
• OpenAI's reasoning model disproved an 80-year-old Erdős conjecture — on its own — It found an infinite family of point constructions that beats the square-grid bound. Cambridge's Timothy Gowers confirmed the proof meets top-journal standards — believed to be the first AI-generated math to clear that bar. [WSJ]
• Websites can now fingerprint you by watching your SSD work — A new storage side-channel times your drive's read/write operations to ID your device across sessions — even in private browsing, and even with today's anti-fingerprinting defenses turned on. [Ars Technica]
• China's lab-grown diamond makers turned out to be sleeper winners of the AI boom — Their diamond heat-spreaders are now shipping to chipmakers as GPU density outruns what copper cooling can handle. One producer's shares jumped 51%. [Bloomberg]
• A humanoid robot autonomously climbed most of a 20,341-ft volcano — A modified Unitree G1 named "Pemba" walked up Ecuador's Chimborazo over a 16-hour push (humans still had to help on the steepest pitches), its systems stress-tested to −47°C. Next target: Everest. [Interesting Engineering]
• A robot makes 500 burrito bowls an hour. A human makes about 45. — Marc Lore's Wonder is deploying the bowl-builder next month — acquired from Sweetgreen, which already runs it across 32 locations. [Fortune]
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  This week's poll
  Cutting off foreign access to America's best AI models — protection, or own goal?
  Last week, 93 of you voted:
  What should investors discount most after the Claude shutdown?
  Cutting off foreign access to America's best AI models — protection, or own goal?
  — Alexis