Four things crossed my feed this weekend. Two of them involve Nvidia writing checks that redefine who controls AI compute. One is the largest open-weight model release in history. And one is the White House quietly locking in a policy framework that every frontier lab will feel.
Nvidia Backstops OpenAI’s $500 Billion Data Center
The Wall Street Journal reported late Saturday that Nvidia is negotiating a roughly $250 billion financial guarantee to help OpenAI lease a 10-gigawatt data center campus in Piketon, Ohio. The facility sits on the site of a former uranium enrichment plant. Total build cost: north of $500 billion, with separate chip financing discussions that could add another $350 billion on top.
The builder read: This is Nvidia becoming an infrastructure bank. OpenAI gets to control its own compute instead of renting from Microsoft, Amazon, and Oracle. Nvidia gets guaranteed chip demand for years. Both sides lock in a dependency that makes this deal almost impossible to unwind once the concrete is poured. If you are building on OpenAI’s APIs, the reliability story just got more interesting: a company that controls its own 10-gigawatt campus is a different kind of provider than one renting capacity across three clouds. Negotiations are early and could still collapse, per Bloomberg’s reporting, but the signal is loud.
Nvidia and SK Group Lock In $500 Billion in AI Infrastructure
On Friday, Nvidia and South Korea’s SK Group signed letters of intent worth over $500 billion, covering AI data centers, HBM4 memory co-development, and next-generation computing infrastructure. SK Telecom will build a 2-gigawatt AI data center on Nvidia’s Vera Rubin platform, targeting first-phase operations in 2027. SK Hynix, already Nvidia’s primary HBM supplier, will deepen a long-term collaboration on HBM4, as CNBC reported.
The builder read: Memory supply is the bottleneck nobody talks about. Training runs are not compute-limited anymore; they are memory-limited. Nvidia locking down SK Hynix as its HBM supplier is a supply-chain moat, not a marketing exercise. If you are planning any serious self-hosted inference for 2027 and beyond, the cost of HBM is going to be shaped by deals like this one, not by spot pricing. A 2-gigawatt facility built around Vera Rubin also tells you where Nvidia’s next-generation hardware is going before the product announcement.
Kimi K3 Open Weights Go Live: 2.8 Trillion Parameters, Free to Download
Moonshot AI dropped the open weights for Kimi K3 on Saturday evening, a day ahead of the July 27 target. At 2.8 trillion parameters, it is the largest open-weight model ever released. The full download is roughly 1.4 terabytes using MXFP4 quantization. Only 16 of 896 mixture-of-experts layers fire per token (about 50 billion active parameters), so per-token compute resembles a mid-size model despite the headline parameter count.
The benchmarks are real: K3 leads the Frontend Code Arena at 1,679 (beating Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618), posts the strongest open-weight GPQA Diamond score at 93.5%, and ranks first on Program Bench, SWE Marathon, and BrowseComp. Together AI and Modal both announced day-zero hosted access.
The builder read: This is the story for anyone running inference at scale. A 2.8T model that activates 50B per token and beats the frontier on coding is exactly the kind of thing that makes self-hosting competitive again, if you can serve 1.4 terabytes of weights. The catch is real: you need a multi-GPU cluster just to load it, and the hallucination benchmarks have not been independently verified yet. But for coding workflows, internal tooling, and any use case where you cannot send data to a Chinese API endpoint, the self-hosting option changes the math. Together AI and Modal handling day-zero hosting means you do not need the cluster yourself to start testing.
White House Nears 30-Day Review Framework for Frontier AI Models
The White House is finalizing a voluntary framework with OpenAI, Anthropic, and Google that would give federal agencies up to 30 days to review new frontier models before public release. An announcement is expected before August 1, according to Eastern Herald’s reporting. The framework traces back to Executive Order 14409, signed June 2, which directed agencies to establish pre-deployment review processes for frontier AI. The benchmarks used to trigger the review are classified. Meta is notably absent from the deal.
The builder read: Voluntary, frontier-only, and the threshold is still unsettled. That sounds toothless until you remember that the labs signing on are the ones with the most to lose from a competitor shipping without review. The 30-day window creates a coordination mechanism: if you are OpenAI and you know Anthropic has to wait 30 days, you are less likely to race to ship. Meta staying out means open-weight releases are not subject to this gate, which makes the Kimi K3 story above even more interesting. For builders, the practical impact is minimal today. For the industry’s structure in 2027, this is the seed of something bigger.
Three of these four stories involve Nvidia. The company is not just selling chips anymore; it is financing the data centers, locking down the memory supply, and shaping which models get built on which hardware. That is the real headline this weekend.