Broadcom-OpenAI chip deal hits snag
- OpenAI’s custom-chip project with Broadcom hit a financing snag on May 7 after talks stalled over whether Microsoft would commit to buying part of the output. - The sticking point is roughly $18 billion for the first phase, with Broadcom wanting Microsoft to take about 40% of the chips. - That matters because OpenAI’s 10-gigawatt buildout only works if financing, chip design, and data-center plans all line up. (finance.yahoo.com)
Custom AI chips are supposed to be OpenAI’s way out of the Nvidia bottleneck. They promise lower costs, tighter control, and eventually a lot more computing power. But the hard part was never just designing the silicon. It was paying for the rollout and finding a home for all that hardware. On May 7, that gap showed up in public when reports said OpenAI’s chip partnership with Broadcom had stalled over financing terms tied to Microsoft. (finance.yahoo.com) ### What exactly got stuck? The first phase of the OpenAI-Broadcom project appears to need about $18 billion in financing, and the talks have slowed because Broadcom wants Microsoft to agree to buy a chunk of the chips before Broadcom commits the money. The number floating around is about 40% of the output. Broadcom shares fell after that report, which tells you investors heard this as more than paperwork. (theinformation.com)t is still the obvious place to put a lot of this hardware. The reported structure would have Microsoft install the chips in its data centers and then rent that capacity back to OpenAI. That makes financial sense if everyone agrees on the setup. But it also means one partner’s hesitation can jam the whole chain. (stocktwits.com)ment is not just about money. It is also about infrastructure design. OpenAI reportedly wants specialized data centers built around its custom silicon, while Microsoft prefers more standard, flexible data-center designs. That is a real strategic difference — one side wants a purpose-built machine, the other wants something easier to reuse across customers and workloads. (stocktwits.com) ### Why are custom chips such a big deal? Because OpenAI is chasing truly huge compute volumes. When the companies made the partnership public in October 2025, they said they planned to build and deploy 10 gigawatts of custom AI accelerators starting in late 2026. The pitch was simple: chips designed around OpenAI’s own workloads could deliver better efficiency and lower inference costs than buying only off-the-shelf GPUs. (cnbc.com) ### How big is 10 gigawatts, really? It is enormous. CNBC noted industry estimates that a 1-gigawatt data center can cost roughly $50 billion, with about $35 billion of that tied to chips at current Nvidia-style pricing. So 10 gigawatts is not a normal supplier contract. It is more like an industrial buildout plan disguised as a chip deal. (cnbc.com) ### Why di(cnbc.com)ice these deals as future revenue streams. If a flagship customer’s rollout depends on unresolved financing and a third party’s purchase commitment, that revenue looks less certain and maybe later than expected. Yahoo Finance said the stock was down about 3% on the day of the report. (finance.yahoo.com) problem or a money problem? Right now it looks more like a commercialization problem. The chips may still happen. But building custom silicon at this scale is not like ordering more servers. You need chip design, packaging, networking, power, buildings, and financing to arrive in sync. Miss one piece and the whole schedule slips. That is the catch with AI infrastructure now — the bottleneck is the entire stack, not just the chip. (cnbc.com) ### Bottom line? This snag does not kill the OpenAI-Broadcom plan. But it does puncture the idea that once an AI company announces “10 gigawatts,” the hard part is over. The hard part is making the capital structure, the cloud partner, and the physical infrastructure agree on the same timeline. (finance.yahoo.com)
