AI Growth Hits Reality | AI Field Notes #10

AnalysisBeijing's National Development and Reform Commission ordered Meta on Monday to unwind its $2 billion acquisition of Manus, the agentic AI startup that had moved to Singapore last summer to dodge exactly this scenario. Manus founders Xiao Hong and Ji Yichao were already barred from leaving China during the probe. The deal is essentially complete on paper: Manus staff joined Meta in December, the website now reads as part of Meta. Beijing does not care. Analysts at Omdia and BDA China called it the first real use of China's 2020 foreign investment security review, and a message to every Chinese founder eyeing a Singapore flip. The timing, weeks before a planned Trump and Xi summit, is not subtle. AI talent is now treated as strategic export inventory, on both sides.

AnalysisA Wall Street Journal report Monday night said OpenAI has been missing its own monthly revenue and weekly active user targets, and that finance chief Sarah Friar has privately told colleagues the company may struggle to fund its compute commitments if growth does not accelerate. ChatGPT slipped past an internal goal of one billion weekly active users by year end, with Anthropic's Claude eating into coding and enterprise revenue and Google's Gemini taking consumer share. The numbers attached to this matter: a $250 billion Azure spend pledge, roughly $600 billion in lifetime compute commitments, and a planned IPO near a $1 trillion valuation. Oracle dropped about 5%, CoreWeave fell 7%, and SoftBank, OpenAI's largest investor, sank around 10% in Tokyo. OpenAI called the framing ridiculous. The market did not.

AnalysisResearchers disclosed CVE-2026-25874 in Hugging Face's LeRobot, an open source robotics platform with about 24,000 GitHub stars. CVSS 9.3. The fix is unwritten, scheduled for version 0.6.0. The cause is the kind of mistake every Python developer is taught to avoid: the async inference PolicyServer calls pickle.loads() on data arriving over an unauthenticated gRPC port without TLS. Anyone who can reach the network gets remote code execution on machines that typically run with GPU access, robot hardware, and elevated privileges. The maintainers acknowledged in January that the code needed to be almost entirely refactored, calling LeRobot a research tool. It is now downloaded for production. Hugging Face also created safetensors specifically because pickle is unsafe, which makes this a self-inflicted joke in the security community.

AnalysisOn Sunday Meta signed a capacity reservation with Overview Energy, a Boston-area startup, for up to one gigawatt of space-based solar power, plus a separate deal with Noon Energy for 1 GW and 100 GWh of 100 hour battery storage. Overview's plan is to put satellites in geosynchronous orbit, collect sunlight 24/7, and beam it as low intensity near infrared light to existing terrestrial solar farms, which then feed the grid at night. Initial orbital demonstration is targeted for 2028, commercial delivery 2030. Neither technology has ever run at scale. Meta has signed up to 30 GW of new energy across 28 states and is now one of the largest corporate buyers of nuclear power in the US. The honest read: hyperscalers no longer believe the existing grid can power their AI plans, so they are funding science projects to get ahead of it.

AnalysisOn Monday roughly 600 Google employees sent CEO Sundar Pichai an open letter pleading with him to refuse a classified AI deal with the Department of Defense, citing risks of mass surveillance and lethal autonomous weapons. By Tuesday, The Information reported Google had signed it anyway, with language that lets the Pentagon use Gemini for any lawful government purpose. Google says it kept guardrails against domestic mass surveillance and autonomous weapons without human oversight. The contract reportedly does not give Google any right to control or veto how DoD applies the model. The pattern is now clean: OpenAI and xAI signed similar deals, Anthropic refused and was branded a supply chain risk and sued the government. The 2018 Project Maven walkout that made Google withdraw is officially historical.

AnalysisMicrosoft released the Visual Studio 2026 April update on Monday, putting cloud agent integration directly into the IDE: a developer can now start a remote Copilot coding agent session inside the editor, hand off a task, and let it run on remote infrastructure while continuing other work. The update also ships custom user-level agents that travel between projects via .agent.md files, repository-discovered agent skills, and a fix that finally stops Copilot completions from fighting IntelliSense (the language-aware autocomplete) for the same dropdown. C++ code editing tools for Copilot agent mode reached general availability. Buried in the announcement: GitHub Copilot moves to usage based billing on June 1. The shape of daily work for a Visual Studio user has changed. The IDE is now the dispatcher for autonomous work, not just an editor with autocomplete.

AnalysisSereact, the Stuttgart based robotics software company, closed a $110 million Series B led by Headline on Monday, more than four times its January 2025 round. The pitch is a model called Cortex 2.0, a vision language action model (a system that turns what a robot sees and what a human says into physical movement) bolted onto a world model that lets the robot simulate possible outcomes of an action before performing it. The numbers Sereact attaches: 200 systems deployed across Europe, more than one billion production picks completed, and one in 53,000 picks needing remote human help. Customers include BMW, Daimler Truck, Mercedes-Benz, PepsiCo, and Austrian Post. The new round funds a Boston office and pushes Sereact into the humanoid market alongside Figure AI, Physical Intelligence, and Skild. The company sells software only, designed to run on whatever metal the warehouse already owns.

AnalysisOpening statements began Tuesday in Oakland federal court for Elon Musk's lawsuit against OpenAI, Sam Altman, Greg Brockman, and Microsoft. Musk wants $150 billion in damages routed to OpenAI's nonprofit foundation, Altman and Brockman removed as officers, and the for-profit conversion reversed. His lawyer Steven Molo told the jury the defendants stole a charity. OpenAI's lead counsel William Savitt countered that Musk is the one who wanted to merge OpenAI with Tesla and own more than half of it, then quit when he could not. On the stand, Musk warned about a Terminator outcome and said AI could kill us all. The trial is scheduled to run three weeks. The timing matters: it lands the same week OpenAI restructured its Microsoft partnership and saw a Wall Street Journal report on missed revenue, both of which feed the IPO calculus a Musk verdict could detonate.

AnalysisTuesday's market reaction to the OpenAI revenue report was the cleanest stress test yet of the AI capex thesis. The Nasdaq closed down 0.9%. Oracle, which signed a $300 billion compute partnership with OpenAI last year, fell about 5%. CoreWeave dropped 7%. SoftBank, OpenAI's largest investor, sank near 10% in Tokyo. Bloom Energy, Seagate, AMD, Broadcom, and Nvidia all sold off, with Nvidia the worst performing of the Magnificent Seven. The Wall Street Journal report did not contain a single new financial number that informed investors did not already roughly know. What it did was give the market permission to ask, out loud, whether $600 billion in projected 2026 AI infrastructure spending across the hyperscalers can be supported by current revenue. Big Tech earnings hit Wednesday and Thursday. Their capex guidance becomes the only line of the call that matters.