Wall Street just agreed to treat GPUs like mortgage bonds and OpenAI is racing Anthropic on privacy promises, but the evidence behind AI's biggest claims — in medicine, in diplomacy — keeps lagging behind the cash.
Nvidia recruits Wall Street to turn GPUs into a $500bn asset class
- Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are working with Nvidia to arrange $500bn in financing for compute, The Verge reported.
- CEO Jensen Huang told CNBC that chips are now "revenue-generating assets" that are "long-lived" and "fungible" — a reversal from his warning last year that old Hopper chips would soon be worthless.
Why it matters: If GPUs become loan collateral the way mortgages once did, a slowdown in AI demand won't just dent tech balance sheets — it could ripple through credit markets that have never priced this kind of asset before.
Marvell hands Google a stake worth up to $12.2bn to lock in a custom chip deal
- Marvell granted Alphabet a stock warrant worth as much as $12.2bn as part of an expanded custom silicon and data-centre networking partnership, Reuters reported.
- Marvell shares jumped more than 12% in premarket trading on the news, according to TradingView.
Why it matters: Paying suppliers in equity rather than cash is becoming a way for hyperscalers to guarantee capacity — and it ties Marvell's fortunes directly to Google's AI infrastructure bets.
OpenAI dangles zero data retention to outflank Anthropic on privacy
- OpenAI said in a blog post that it does not retain prompts or responses for eligible API customers, and previewed "Private Safety Processing" to spot misuse patterns without exposing content to staff.
- TechCrunch reported the move directly counters Anthropic's July policy of retaining conversation data for 30 days on its "Mythos-class" models, which has unsettled some enterprise customers.
Why it matters: Privacy terms, not just model capability, are becoming a real competitive front — enterprises handling sensitive data now have genuine leverage to negotiate.
Washington and Beijing still haven't nailed down Xi's AI talks — weeks before his US visit
- With Chinese President Xi Jinping's US visit approaching, the venue, participants and agenda for a hoped-for AI dialogue remain unsettled, the South China Morning Post reported, citing two sources familiar with the discussions.
- US officials haven't decided which government agency should lead the talks, nor whether technical experts or private companies should take part.
Why it matters: Without agreed ground rules, companies operating across both markets face growing regulatory whiplash — and a missed dialogue window could harden divergent AI standards for years.
Only 3 of 1,357 FDA-cleared AI medical devices have been tested for real patient benefit
- A study in PLOS Digital Health found just 34 of 1,357 FDA-authorised AI medical devices were evaluated in registered clinical trials — and only three tested whether they actually improved patients' health.
- FDA clearance generally requires only "substantial equivalence" to an existing device, not proof of benefit.
Why it matters: Hospitals worldwide adopting FDA-cleared tools may be assuming a clinical evidence base that mostly doesn't exist — a gap regulators outside the US will need to close themselves.
The Hexalink view
Notice the shape of today's news: money is moving with total confidence — half a trillion dollars to turn chips into bonds, twelve billion in equity to lock down a chip supplier — while the underlying proof keeps lagging. Only three medical AI devices out of well over a thousand have ever been checked for whether they help patients. Two governments can't agree on who should even be in the room to discuss AI safety. The industry is very good at pricing confidence and very bad at pricing evidence.
For a technology leader, the lesson is to stop assuming clearance or certification means proof. If deploying clinical AI, ask for outcome data, not just approval status. If negotiating enterprise AI contracts, treat data-retention terms as a live variable. And watch the Xi visit closely — a stalled dialogue now previews the compliance headaches ahead.
We'll be back tomorrow with the next briefing — until then, our thinking lives on the Hexalink blog.
Sources
- Nvidia's $500bn compute-as-asset-class push — The Verge
- Marvell's $12.2bn Google chip deal — Reuters
- OpenAI's zero data retention push — OpenAI News
- Xi's AI talks still unsettled — South China Morning Post
- FDA-cleared medical AI rarely tested for benefit — News-Medical.net