内容来源：https://aiweekly.co/issues/nvidia-may-buy-into-perplexity-above-30b-before-wednesdays

内容总结：

英伟达拟投资Perplexity，AI资本格局生变

据The Information报道，英伟达正与AI搜索公司Perplexity洽谈股权投资事宜，后者估值有望突破300亿美元。消息人士透露，Perplexity的年化收入已从2026年初的不到2.5亿美元飙升至7.5亿美元以上。

与此同时，彭博社称软银计划发行创纪录的1万亿日元（约合63亿美元）零售债券，用于偿还其持有OpenAI股份背后的过桥贷款，并继续为AI交易融资。

两则消息叠加，揭示了即将到来的资本分化趋势：芯片供应商正向产品层渗透，而前沿投资者开始动员日本储户为其押注买单。英伟达将于美东时间周三下午5点发布财报，市场将密切关注管理层措辞——他们究竟以供应商、投资者，还是双重身份出现。

行业动向速览

• Meta持续挖角OpenAI人才，研究员Luke Metz加盟Meta超级智能实验室，此前他曾参与DCGAN开发及OpenAI o1系统设计。
• OpenAI CEO奥特曼警告AI权力集中风险，称若公众因安全担忧而放弃自由换取保护，AI可能被少数巨头掌控。
• 韩国拟限制杠杆芯片ETF的个人投资敞口，要求散户购买前完成一周课程，为三星和SK海力士的投机热潮降温。
• OpenAI主动请求加州加强前沿实验室监管，建议SB 53法案增加训练监控条款，打破行业与监管对立的常规剧本。
• 多州州长因电力、水资源及当地许可问题，从招商引资转向抵制数据中心建设，AI基础设施的下一个瓶颈或是政治许可而非芯片。
• 特朗普公开支持AI数据中心建设，称其规模可能超越石油行业，与地方政府的抵制形成鲜明共和党内部分歧。
• 谷歌和微软正竞相布局美国校园AI教育工具，争夺学生接触的首个AI助手入口。
• AI代理正将创业者变成每周工作90小时的"管理者"，担心竞争对手自动化更快而产生新的生产力焦虑。

英伟达财报看点

本次财报的实质问题已不仅是芯片出货量，而是英伟达打算"拥有"什么。Perplexity谈判将资本配置提升为战略问题——入股意味着英伟达更贴近消耗其芯片的搜索与答案层。投资者应关注管理层如何描述投资性质：是财务敞口、生态建设，还是分销布局。

需重点追踪三个数字：战略投资现金承诺、回购规模、供应协议。若战略投资持续扩大而回购放缓，说明英伟达正用芯片利润在硬件之外构建影响力；若投资规模远小于回购，则参股Perplexity仅是期权而非新商业模式。

问答环节比准备好的增长话术更重要。分析师应追问：除股权外，英伟达是否寻求技术授权、商业承诺、董事会席位或招聘优先权？据报道，投资方案提出前曾考虑过授权和人才引进路径，每条路径买到的价值截然不同。

电话会结束时，请用一句话总结：英伟达主要是在销售AI热潮、为AI热潮融资，还是试图控制更多分销环节？这个答案比单季财报的盈亏更具长期意义。

关键要点

• 融资路径与估值同等重要：股权、零售债券和供应商投资将风险分散到不同群体。
• 区分代理数量与管理杠杆：更大的代理团队可能增加监督负担而非实际产出。
• 校园入口即分销入口：学生最先学会使用的助手，可能成为他们职场默认工具。
• 关注英伟达周三问答环节：战略投资伴随哪些权利和商业承诺，值得深挖。

中文翻译：

据The Information报道，英伟达正在洽谈以超过300亿美元的估值对Perplexity进行股权投资。据报道，Perplexity的年化收入已超过7.5亿美元，而2026年初这一数字还不到2.5亿美元。另外，彭博社称软银计划发行创纪录的1万亿日元（约合63亿美元）零售债券，以偿还收购OpenAI股份时使用的过桥贷款，并为更多AI交易提供资金。综合来看，这些信号揭示了未来一周资本流向的分化：芯片供应商正在向应用层迈进，而前沿投资者则在呼吁日本储户为其下注提供资金。英伟达将于美国东部时间周三下午5点发布财报。请留意管理层现在的口吻是像供应商、投资者，还是两者兼有。

赞助商

当前有哪些趋势正在塑造企业战略？

探索审计合伙人对当今影响商业决策、风险管理和企业战略的趋势的观察。

In the Wild

在成为新闻稿之前，什么正在悄然发生。在In the Wild中查看完整动态。

• Harness优化达到了18，600次Telegram浏览。Scale AI的新基准测试检验一个AI能否改进另一个AI的提示词、工具、记忆和控制流程。论文发现，优化器模型比其所使用的编码框架更为关键。阅读HarnessOpt-Bench。
• Conduit在生产力类目中上升五位，升至第11名。这款iPhone客户端可连接用户自己的Open WebUI服务器，这是一个微小信号，表明自托管AI正在超越桌面端爱好者群体。查看Conduit。
• Vidmor在图形与设计类目中上升三位，升至第12名。又一款文生视频应用在不拥有前沿模型的情况下找到了买家。分发本身仍然是一种产品。查看Vidmor。
• 一款无付费墙的图像生成器以第16名新晋榜单。它的名字就是卖点：无需订阅门槛，无限生成。消费端价格压力最先在这里显现。查看该应用。
• Runway在照片与视频类目中上升三位，升至第20名。专业生成式视频与低成本封装应用在同一张榜单上竞争。这个类目正在向两端同时拓宽。查看Runway。

快讯

实验室角斗士时代

• META又挖走了一位OPENAI研究员。Luke Metz将离开OpenAI加入Meta超级智能实验室，直接向Alexandr Wang汇报。Metz于今年一月从Thinking Machines Lab回归，曾共同开发DCGAN，并为OpenAI的o1系统卡片做出贡献。查看Meta人才争夺战的最新动向。
• SAM ALTMAN在最值钱的实验室内部发出权力集中的警告。他对Business Insider表示，如果安全恐惧促使人们用自由换取保护，AI最终可能被少数几家强大公司控制。这一警告将让下周所有关于开源权重（open-weights）的争论更加尖锐。阅读Altman的观点。

各国政府认真对待的一年

• 韩国在AI芯片交易面前设置了减速带。监管机构计划限制个人对杠杆芯片ETF的敞口，并要求散户投资者在购买前完成为期一周的课程。这一护栏措施出台前，三星和SK海力士的相关投注出现了激增。查看新的投资者规则。
• OPENAI请求加州加强对前沿实验室的监管。该公司希望SB 53法案在前沿训练中加入监控条款，此前发生了最近的智能体被黑客入侵事件。一个实验室游说加强州级法律，这将使通常的行业对监管的剧本变得更加复杂。查看OpenAI提出的修改意见。

AI资本支出税

• 曾经招商引资数据中心的外长们现在正在反对它们。《华尔街日报》追踪了这一转变，电力、用水和地方审批正在成为选举议题。AI基础设施的下一个制约因素可能是政治许可，而非芯片。查看州级反弹。
• 特朗普告诉这些社区，他们正在犯一个错误。在周日接受迈克尔·科恩采访时，总统表示AI数据中心可能变得比石油还重要，并驳回了电网方面的担忧，称运营商正在自建发电厂。共和党内部的分歧现在已经公开化。查看特朗普对大规模建设辩护。

全自动模式

• GOOGLE和MICROSOFT正在争相为美国学校接入AI。《纽约时报》报道，两家公司都在推进课堂工具和培训，竞相塑造学生最先学会使用的助手。我们消息流中的十位专家分享了这篇报道。查看校园分发竞赛。
• AI智能体正在把创业公司创始人变成每周工作90小时的管理者。《华尔街日报》发现，创始人全天候监督着成群的智能体，因为他们担心竞争对手会更快实现自动化。更多的智能体容量正在制造一种新形式的生产力错失恐惧症（FOMO）。查看智能体管理跑步机。

英伟达周三发布财报。问一问它的钱将流向哪里

英伟达将于周三发布第二财季业绩。网络直播于太平洋时间下午2点开始。

真正值得追问的问题不再仅仅局限于它能出货多少加速器。要问英伟达打算拥有什么。

Perplexity谈判把资本配置变成了战略问题。入股将使英伟达更接近消耗其芯片的搜索与回答层。请留意管理层是将这些投资描述为财务敞口、生态系统建设还是分发渠道。这些措辞暗示了不同的野心。

然后追踪三个数字：承诺用于战略投资的现金、回购和供应协议。如果战略投资持续扩大而回购放缓，说明英伟达正在利用芯片利润在硬件之外建立影响力。如果投资规模与回购相比仍然很小，那么Perplexity只是一项选择权，而非新的运营模式。

问答环节比准备好的增长话术更重要。分析师应当先问英伟达在讨论入股之前是否寻求了技术授权、商业承诺、董事会席位或招聘选项，然后再讨论股权本身。The Information称，在走投资路线之前，许可协议和人才引进都曾被考虑过。每条路径买到的东西各不相同。

财报电话会议结束后，写下一句话：英伟达主要在销售AI热潮、在为它融资，还是在试图拥有更多分发环节。这个答案比第一次业绩超预期或不及预期更值得长期关注。

要点总结

• 关注融资路径与估值本身同等重要。股权、零售债券和供应商投资将风险分散到不同的人群身上。
• 把智能体容量与管理杠杆区分开来。更大的智能体舰队可能增加监督负担的速度快于增加有效产出的速度。
• 把校园入口视为分发渠道。学生最先学到的助手可能成为他们在工作中默认使用的工具。
• 把英伟达周三的问答环节记在日程上。问清楚战略投资附带哪些权利和商业承诺。

首发发现

一手研究AI周刊在进入新闻周期之前就已经浮出水面。

• 所有五个被测试的LLM记忆框架都输给了完全没有记忆的方案。在长时程任务中，最好的记忆方法仍然比无记忆基线低了十个百分点以上。持久化记忆增加检索噪声的速度可能快于增加有效上下文的速度。查看论文和我们的早期解读。
• 最好的编码智能体在科学软件上仍然低于50%。SWE-bench Science测试了20个科学领域的119个真实任务。即使是Claude Code搭配Opus 5，在pass@1上也有一半以上的任务失败。查看编码智能体在哪里失效。

值得一读

• 互联网上有多少内容是由AI写的？：Pew分析了Common Crawl中49万个英文页面。2026年7月的样本中有10%显示出显著的AI创作痕迹，而在ChatGPT发布后发布的页面中，这一比例超过三分之一。检测并不完美，但时间序列的趋势很难被否認。

等等，什么？

• 哈佛商学院正在以699美元的价格出售其教授的AI克隆体访问权限。其为期八周的Foundry训练营使用HeyGen数字人来点评路演，并在直播课程之间主持模拟董事会会议。一位授课教师称自己的数字孪生“令人毛骨悚然”。查看AI师资团队。

值得观看

AI从业者正在传阅的视频——由AI TV精选。

本周投票

下周最有价值的AI信号将出现在哪里？

下周最有价值的AI信号将出现在哪里？

周三见。

Alexis

英文来源：

The Information reports that Nvidia is discussing an equity investment in Perplexity at a valuation above $30 billion. Perplexity's annualized revenue has reportedly passed $750 million, up from less than $250 million at the start of 2026. Separately, Bloomberg says SoftBank plans a record ¥1 trillion, or $6.3 billion, retail bond to help repay the bridge loan behind its OpenAI stake and fund more AI deals. Put together, the signals expose next week's capital split: the chip supplier is moving toward the product layer, while the frontier investor is asking Japanese savers to finance its bets. Nvidia reports Wednesday at 5 p.m. ET. Listen for whether management now talks like a supplier, an investor, or both.
Sponsor
What are the current trends shaping corporate strategy?
Explore observations from audit partners on the trends shaping business decision-making, risk management, and corporate strategy today.In the Wild
What is moving before it becomes a press release. See the full pulse in In the Wild.

• Harness optimization reached 18,600 Telegram views. Scale AI's new benchmark tests whether one AI can improve the prompts, tools, memory and control flow around another. The paper found that the optimizer model mattered more than the coding harness it used. Read HarnessOpt-Bench.
• Conduit climbed five places to No. 11 in Productivity. The iPhone client connects to a user's own Open WebUI server, a small sign that self-hosted AI is reaching beyond desktop tinkerers. See Conduit.
• Vidmor rose three places to No. 12 in Graphics & Design. Another text-to-video app is finding buyers without owning a frontier model. Distribution is still a product. See Vidmor.
• A no-paywall image generator entered at No. 16. Its name is the pitch: unlimited generation without a subscription gate. That is where consumer price pressure is showing up first. See the app.
• Runway climbed three places to No. 20 in Photo & Video. Professional generative video is gaining on the same chart as low-cost wrappers. The category is widening at both ends. See Runway.
  Quick Hits
  The Lab Gladiator Era
• META JUST TOOK ANOTHER OPENAI RESEARCHER. Luke Metz is leaving OpenAI for Meta Superintelligence Labs and will report to Alexandr Wang. Metz returned from Thinking Machines Lab in January, co-developed DCGAN and contributed to OpenAI's o1 system card. See the latest move in Meta's talent raid.
• SAM ALTMAN IS WARNING ABOUT POWER CONCENTRATION FROM INSIDE THE MOST VALUABLE LAB. He told Business Insider that AI could end up controlled by a few powerful companies if safety fears persuade people to trade liberty for protection. The warning will sharpen every open-weights argument next week. Read Altman's case.
  The Year Governments Got Serious
• SOUTH KOREA IS PUTTING A SPEED BUMP IN FRONT OF THE AI-CHIP TRADE. Regulators plan to cap individual exposure to leveraged chip ETFs and require a weeklong course before retail investors can buy them. The guardrail follows a boom in bets on Samsung and SK hynix. See the new investor rules.
• OPENAI ASKED CALIFORNIA FOR STRONGER FRONTIER-LAB RULES. The company wants SB 53 to add monitoring around frontier training after recent agent hacking incidents. A lab lobbying for tighter state law will complicate the usual industry-versus-regulator script. See OpenAI's requested changes.
  The AI Capex Tax
• GOVERNORS WHO RECRUITED DATA CENTERS ARE NOW CAMPAIGNING AGAINST THEM. The Wall Street Journal tracks the reversal as electricity, water and local approval become election issues. The next constraint on AI infrastructure may be political permission rather than chips. See the state-level backlash.
• TRUMP TOLD THOSE COMMUNITIES THEY ARE MAKING A MISTAKE. In a Sunday interview with Michael Cohen, the president said AI data centers could become bigger than oil and dismissed grid concerns because operators are building power plants. The Republican split is now explicit. See Trump's defense of the buildout.
  Auto Mode Everything
• GOOGLE AND MICROSOFT ARE RACING TO WIRE US SCHOOLS FOR AI. The New York Times reports that both companies are pushing classroom tools and training as they compete to shape which assistant students learn first. Ten experts in our stream shared the report. See the school distribution race.
• AI AGENTS ARE TURNING STARTUP FOUNDERS INTO 90-HOUR MANAGERS. The Wall Street Journal found founders supervising fleets of agents around the clock, driven by fear that a competitor will automate faster. More agent capacity is creating a new form of productivity FOMO. See the agent-management treadmill.
  Nvidia Reports Wednesday. Ask Who Gets Its Cash
  Nvidia reports fiscal second-quarter results Wednesday. The webcast starts at 2 p.m. Pacific.
  The actionable question is no longer limited to how many accelerators it can ship. Ask what Nvidia plans to own.
  The Perplexity talks turn capital allocation into strategy. A stake would put Nvidia closer to the search and answer layer that consumes its chips. Listen for whether management describes investments as financial exposure, ecosystem development or distribution. Those labels imply different ambitions.
  Then track three numbers: cash committed to strategic investments, buybacks and supply agreements. If strategic checks keep expanding while repurchases slow, Nvidia is using chip profits to build influence beyond hardware. If investment stays small beside buybacks, Perplexity is an option rather than a new operating model.
  The Q&A matters more than the prepared growth language. Analysts should ask whether Nvidia sought technology rights, commercial commitments, board access or hiring options before discussing equity. The Information says licensing and staff hires were considered before the investment route. Each path buys something different.
  By the end of the call, write down one sentence: Nvidia is primarily selling the AI boom, financing it, or trying to own more of its distribution. That answer will matter longer than the first earnings beat or miss.
  Key Takeaways
• Track the financing route as closely as the valuation. Equity, retail debt and supplier investment place risk with different people.
• Separate agent capacity from managerial leverage. A bigger fleet can add supervision faster than it adds useful output.
• Treat school access as distribution. The assistant students learn first can become the tool they expect at work.
• Put Nvidia's Wednesday Q&A on the calendar. Ask what rights and commercial commitments come with strategic investments.
  Found First
  Primary research AI Weekly surfaced before it reached the press cycle.
• ALL FIVE TESTED LLM MEMORY FRAMEWORKS LOST TO NO MEMORY AT ALL. Across long-horizon tasks, the best memory methods still finished more than ten percentage points below the no-memory baseline. Persistent memory can add retrieval noise faster than it adds useful context. See the paper and our early read.
• THE BEST CODING AGENT STAYED BELOW 50% ON SCIENTIFIC SOFTWARE. SWE-bench Science tested 119 real tasks across 20 scientific domains. Even Claude Code with Opus 5 failed more than half at pass@1. See where coding agents break.
  Worth Reading
• How Much of the Internet Is Written With AI?: Pew analyzed 490,000 English-language pages from Common Crawl. Ten percent of the July 2026 sample showed significant signs of AI authorship, rising above one-third among pages published after ChatGPT's launch. Detection is imperfect, but the time series makes the direction hard to dismiss.
  Wait, What?
• HARVARD BUSINESS SCHOOL IS SELLING ACCESS TO AI CLONES OF ITS PROFESSORS FOR $699. Its eight-week Foundry bootcamp uses HeyGen avatars to critique pitches and conduct mock board meetings between live sessions. One instructor called his own digital twin “creepy.” See the AI faculty.
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  This week's poll
  Where will next week's most useful AI signal appear?
  Where will next week's most useful AI signal appear?
  Back Wednesday.
  Alexis