Broadcom shares fall on OpenAI snag

Custom AI chips are supposed to do one thing for companies like OpenAI — cut dependence on Nvidia and lock in cheaper, more tailored computing at huge scale. But that only works if somebody is willing to pay for the factories, servers, and power behind them. That’s where this story turns. On May 8, Broadcom shares fell after a report said OpenAI’s custom-chip project with Broadcom had run into a financing snag around its first phase. (theinformation.com) ### What actually got stuck? The reported problem is not the chip design itself. It’s the money needed to get the first phase moving. The Information said OpenAI and Broadcom are discussing terms for Broadcom to finance initial production, but the price tag for phase one is about $18 billion and the talks have stalled around who takes that risk. (theinformation.com) ### Why is Microsoft in the middle of this? Because Microsoft is still the obvious home for a lot of OpenAI’s compute. The report said Broadcom would finance phase one only if Microsoft agreed to buy roughly 40% of the chips, install them in its data centers, and effectively rent the capacity back to OpenAI. Basically, Broadcom does not want to bankroll a giant buildout without a big committed customer on the other side. (theinformation.com) ### Why does that hit Broadcom’s stock? Because investors had treated custom AI chips as one of Broadcom’s cleanest next legs of growth. If a flagship project with OpenAI slows down over financing, the market starts asking a harder question — how many of these “obvious” AI infrastructure wins are really const(theinformation.com)n the session. (theinformation.com) ### Wasn’t this deal already announced? Yes — last fall OpenAI and Broadcom said they would work together on custom AI chips, with deployment starting in the second half of 2026. At the time, the partnership looked like another sign that the biggest AI labs were moving beyond renting Nvi(theinformation.com)erent things. (economictimes.indiatimes.com) ### Why is the number so large? Because this is not just a chip order. It’s a compute-and-power buildout. The reporting around the project tied it to enough future chip capacity to consume 10 gigawatts of power before 2030. That is utility-scale demand. The first phase alone is described as requiring 1.3 gigawatts, which helps explain why an $18 billion opening bill is even on the table. (theinformation.com) ### How does energy get pulled into this? AI data centers are making electricity a strategic bottleneck. Microsoft is now reportedly weighing whether to delay or abandon its 2030 goal of matching all hourly electricity use with renewable purchases, because the AI buildout is getting so large and so power-hung(theinformation.com)r angle — AI expansion is now colliding with financing and energy constraints at once. (msn.com) ### So what matters next? Watch whether Microsoft commits to the chips, and on what terms. If it does, the project looks more like a delayed capital-allocation fight than a broken strategy. If it doesn’t, this becomes a warning that the AI boom’s next bottleneck is not model demand or chip design. It’s who is willing to underwrite tens of billions of dollars before the economics are fully proven.