Hugging Face is having the kind of week that’s hard to summarize in one sentence, so here are two: a state attorney general just subpoenaed OpenAI over the hack that hit the platform in July, and separately, Hugging Face is reportedly shopping itself for $13 billion — nearly triple its 2023 valuation. Alabama’s AG says the state is investigating OpenAI’s “complete lack of oversight and adequate safeguards” after an unreleased, guardrail-free cybersecurity model escaped its test environment, got online, and spent a weekend loose inside Hugging Face’s infrastructure. That’s an awkward backdrop for a company trying to convince bidders it’s worth three times what it was two years ago, but it also isn’t stopping them: open-model hosting has become infrastructure nobody wants to build twice, hack or no hack.
Infrastructure, and who actually needs the expensive kind, is also the story at Anthropic. Despite an annualized revenue run rate north of $65 billion heading into an IPO that could rival SpaceX’s, Anthropic’s flagship Fable 5 model accounts for just 11% of the company’s spend on Ramp’s tracked corporate cards — most customers are routing everyday work to cheaper models, Anthropic’s own older ones or open-weight alternatives from Chinese labs, and saving the frontier model for the tasks that actually need it. “Most people don’t need to operate at the frontier,” an Accel partner told the FT, which is either a mundane observation about diminishing returns or a preview of margin compression across the entire industry, depending on how bullish your IPO deck needs to be.
Nvidia knows the feeling. The stock is on a seventh straight day of declines heading into Wednesday’s earnings call, with investors nervous about the same demand questions dogging Anthropic — even as the company keeps shipping. Groq 3 LPX, the inference chip born of Nvidia’s $20 billion Groq acquisition, hit full production this week, promising 3,400 tokens per second for the kind of long, multi-step agent workloads that are supposed to justify all this spending. It landed amid Hot Chips 2026 at Stanford, where Intel showed off its own answer: a 256-core Diamond Rapids Xeon, a 480GB Crescent Island inference GPU, and an edge chip called Wildcat Lake. The conference used to be a nerdy sideshow for chip architects; this year it was basically a trade show for the agentic-AI economy, which tells you where the money — real and hoped-for — is actually going.
Meanwhile Microsoft is getting a reminder that “the AI economy” is really just a few very large bets. A Bloomberg analysis this week flagged how much of Microsoft’s AI compute revenue rides on OpenAI, TikTok and Meta specifically — a concentration risk that cuts both ways if any one of those relationships wobbles. Apple, for its part, is reshuffling rather than betting big, cutting jobs across Vision Pro, Siri and its Intelligent Systems Experiences group to free up headcount for its AI rebuild — the second such trim this month.
On the safety side, xAI is suing a photographer already facing over 200 charges for using Grok to generate sexual images of his own clients, arguing he — not the tool — bears responsibility for what it produced. And a University of Toronto study published in PLOS Digital Health found that of 1,357 AI medical devices the FDA has cleared, exactly three have been tested on whether they actually improve a patient’s health. “Substantial equivalence” to an existing device is the bar; proof of benefit isn’t required. Two very different stories, same underlying pattern: the tools are shipping faster than anyone’s checking what they do once they’re out in the world.
