Last fall, OpenAI and Broadcom grandly announced a partnership to design custom AI chips, aiming to break free from Nvidia’s grip. But now the deal is stuck on one question: who pays? According to an internal memo obtained by The Information, the two companies are locked in a tug-of-war over the first $18 billion in manufacturing costs. Broadcom’s core condition for fronting the money: Microsoft must commit to buying roughly 40 percent of the chips.
The chip project, codenamed Nexus, is breathtaking in scope — its ultimate goal is to deploy chips powering 10 gigawatts of compute (with production costs alone exceeding $180 billion). To manage risk, Broadcom wants deep-pocketed Microsoft to buy that initial 40 percent of chips, install them in its own data centers, and then lease them back to OpenAI — essentially serving as a financial guarantee. If Microsoft doesn’t bite, OpenAI has to find another buyer on its own. So far, Microsoft has only reserved rack space in its data centers, not made any purchase commitment. The OpenAI memo calls this a “persistent systemic risk” to the project.
OpenAI wanted to build its way to independence, but now it’s tied even tighter. Sachin Katti, the executive in charge of the project, warned internally that making Microsoft’s procurement pledge a prerequisite for financing would “severely limit long-term growth” and make the deal financially unattractive. Still, the company decided to push ahead for the strategic value. On top of that, Broadcom’s tight allocation of capacity at TSMC means the first chip — codenamed Jalapeno, a inference accelerator — won’t hit volume production until 2027. As a concession, Broadcom recently dropped its demand that OpenAI match its investment dollar-for-dollar and agreed to shoulder more upfront costs.