The money behind the AI buildout is starting to loop back on itself. Nvidia — the company selling the picks and shovels — is increasingly the one financing the miners, and this week the credit market flinched. At the same time, the industry's oldest argument, open versus closed, hardened from a philosophical debate into a public lobbying fight. Here's what moved and why it matters.
Nvidia is now financing the customers who buy its chips — and the credit market noticed
Nvidia's credit-default swaps jumped the most on record after roughly $750B of stacked AI commitments revived fears of "circular financing". The centerpiece is a reported ~$250B backstop that lets OpenAI lease SoftBank's 10-gigawatt campus in Piketon, Ohio — a former uranium-enrichment site — with a separate ~$350B tranche financing the chips that go inside. When the largest chip seller underwrites the demand for its own chips, revenue and credit risk stop being independent variables, and a wobble at the customers becomes a wobble on Nvidia's own book. Equity holders are cheering the bookings; the credit market, as Axios notes, is pricing the loop.
A $5 billion check into a lab with no product and no revenue
The clearest illustration of that loop is Nvidia's direct stake in Safe Superintelligence, the roughly two-year-old lab founded by former OpenAI chief scientist Ilya Sutskever. SSI is valued near $32B despite shipping nothing and booking no revenue, and Nvidia's $5B comes bundled with priority access to its next-generation "Vera Rubin" compute. This is Nvidia backing a frontier bet rather than merely selling into one — its equity stakes are quietly becoming a second business line, one whose returns depend on which labs win the race rather than on how many chips ship.
Open versus closed goes public
The industry's deepest divide stopped being subtext this week. Nvidia, Microsoft, Meta, IBM, Hugging Face, a16z, Dell, Mozilla and the Linux Foundation co-signed a letter, "Open Weights and American AI Leadership," urging Washington not to restrict open-weight models or limit distillation, CNBC reported. Just as telling was who stayed off it: OpenAI and Anthropic, the two frontier labs whose businesses rest on proprietary weights. A company's signature — or its absence — now reads as a statement about where it wants value to accrue in the stack, and whether it defends a closed moat or commoditizes the layer below its rivals.
The frontier discards its own recent past
OpenAI pushed through its largest single model deprecation to date, retiring 18 legacy models — o1, o1-pro, o3-mini, o4-mini, Codex and various deep-research snapshots — from its API. For developers it means another forced migration; for everyone else it's a measure of how fast the frontier throws away last generation's work. Building on a single vendor's fast-moving catalog carries a recurring tax, and it makes the abstraction layers that insulate against churn look more valuable each cycle.
Record money for the anti-Nvidia bet
Not all the capital is flowing toward Nvidia's orbit. Etched, the startup building a transformer-specialized ASIC called "Sohu," raised $300M at a $10.3B valuation — the highest valuation Sequoia has ever led a Series C into — with a16z, Jane Street and SK Hynix joining. The pitch is a hedge against GPU dominance: purpose-built silicon that trades flexibility for raw transformer throughput. The risk is architectural lock-in if the transformer turns out not to be the last word in model design.
The money-view is that risk and capital both routed through one company this cycle, and for the first time the credit market started pricing the danger of Nvidia financing its own demand even as equity investors cheered the deals. Watch three things next: Nvidia's disclosures on the guarantees and off-balance-sheet exposure behind that $750B, whether OpenAI or Anthropic answers the open-weights letter, and Etched's first Sohu benchmarks — the earliest real test of whether the anti-Nvidia bet has silicon to back the valuation. Signal, not advice.