Start with the number: $250 billion. That’s roughly what Nvidia is discussing guaranteeing so OpenAI can lease a 10-gigawatt data center that SoftBank’s energy arm is building on the site of a former uranium enrichment plant in Piketon, Ohio. It doesn’t even cover the chips — that’s a separate $350 billion conversation. All in, the campus could run past $500 billion. For OpenAI, it’s the first step toward owning infrastructure instead of renting it from Microsoft, Amazon, and Oracle. For Nvidia, it’s a demand guarantee dressed up as a favor. Neither company is really taking on risk here so much as making sure someone, somewhere, is contractually obligated to eat it.
It was that kind of week: everyone finding new walls to build around AI, and a few finding out the old ones don’t hold.
Samsung and Broadcom signed their own $200 billion pact — five years of HBM4 memory and 2-nanometer foundry capacity, aimed squarely at loosening Broadcom’s dependence on TSMC and giving Samsung’s badly lagging foundry business a marquee customer. Two chip megadeals in three days is less a headline than a weather pattern: money is now arriving in the AI supply chain in units too large to describe any other way.
Meanwhile the wall meant to keep the good chips out of China turned out to have a door in it. Commerce’s Jeffrey Kessler told Congress that H200 shipments remain “trivial” despite $10 billion in approved licenses — but lawmakers from both parties spent the hearing on a worse detail: Nvidia’s more advanced Blackwell chips appear to have been reaching Chinese firms for nearly a year through a licensing loophole the department only quietly closed in May. Ten Chinese companies, per the Pentagon’s own count, already had approved H200 access. The export-control wall isn’t being stormed. It’s just got gaps nobody photographed until now.
And the wall around frontier intelligence itself came down on its own schedule. Moonshot released Kimi K3’s full open weights a day early — 2.8 trillion parameters, free to download, the largest open-weight model anyone has shipped. You’ll need eight-plus 80GB GPUs just to load it, so “free” and “accessible” remain different words. But the frontier gap that mattered a year ago — you paid an API bill or you didn’t have a frontier model — is closing from the open-weight side faster than most labs priced in.
None of this week’s news was really about intelligence. It was about scale outrunning the structures built to hold it: financing too large for any one balance sheet, chips too valuable to keep fully contained, and now weights too big to keep proprietary even when someone tries. The containers keep getting bigger. So does what leaks out of them.
