OpenAI sits on $122 billion in cash with no public listing deadline, a combination that lets its finance chief treat the world’s most valuable chipmaker as a vendor rather than a partner. When Sarah Friar told CNBC that diversifying supply is simply “good CFO risk mitigation,” she described a negotiating posture that compresses Nvidia’s premium long before it dents revenue.
The numbers behind the posture
Broadcom fell 7.95 percent while AMD barely moved after the interview, a split that reflects how the market prices each company’s exposure to OpenAI’s procurement shift. Friar confirmed the startup still buys “enormous quantities” of Nvidia silicon and has locked in roughly 12 gigawatts of Nvidia compute through 2030. The volume remains; the leverage has moved.
Training and inference are becoming separate markets
Friar drew a distinction that the industry has blurred for years. Training a frontier model is episodic and rewards general-purpose accelerators. Inference runs on every user prompt thereafter, and its economics favor silicon built around a single model’s operations. She cited a chip, referred to as “jalapeno”, designed specifically for that inferencing workload, saying it is “set up exactly for our models” and therefore more efficient.
What the terms reveal
The commitment to 12 gigawatts through 2030 is a floor, not a ceiling. What changed is the ability to source the marginal gigawatt elsewhere. Nvidia still owns the training workload that ends once; purpose-built inference chips capture every query that follows. For a buyer with $122 billion and no IPO clock, that distinction is the only one that matters.