OpenAI signs $10B chip deal with Broadcom to mass-produce custom AI inference chips on TSMC 3nm
Sep 5, 2025
Key Points
- OpenAI signs $10 billion deal with Broadcom to design and mass-produce custom AI inference chips on TSMC's 3nm process, targeting production in 2026 to address GPU shortages constraining model deployment.
- Richard Ho, ex-Google TPU lead, heads a 40-person team building chips optimized for inference rather than training, signaling OpenAI's shift from raw capability gains to reducing costs and improving margins.
- The deal reflects AI infrastructure becoming the binding constraint on progress: NVIDIA GPU lead times stretch months, making custom silicon a competitive hedge alongside OpenAI's $30 billion Oracle data center commitment.
Summary
OpenAI has signed a $10 billion deal with Broadcom to design and mass-produce custom AI inference chips, addressing what CEO Sam Altman has called a critical GPU shortage limiting the company's ability to roll out new ChatGPT versions. The arrangement, initially disclosed via Broadcom's earnings call mentioning a fourth major AI developer customer with a one-time $10 billion order, represents a significant vertical expansion into chip manufacturing for OpenAI.
Richard Ho, formerly of Google's TPU team, is leading the effort with approximately 40 engineers. Broadcom will provide design and intellectual property while TSMC will fabricate the chips on its 3nm node, with target mass production in 2026. The chips are designed for inference only—running trained models rather than training new ones—which aligns with OpenAI's stated priority to reduce inference costs and improve margins rather than push further model capability gains.
The timing reflects genuine infrastructure constraints. Altman publicly cited GPU shortages in February when explaining delays and costs around GPT-4.5, stating the company was "out of GPUs" despite adding hundreds of thousands of GPUs online. The custom chip strategy complements OpenAI's broader infrastructure buildout: a $30 billion annual data center deal with Oracle signed earlier this year and the company's Stargate project, a data center construction venture. Access to off-the-shelf NVIDIA GPUs requires months of lead time, making custom silicon a logical hedge against sustained demand from hyperscalers.
The strategic calculus is straightforward—inference economics matter more than raw model capability at this stage. The hosts note that GPT-4 and GPT-4.5 already deliver sufficient quality for most users; the constraint is speed and cost, not performance gains. Broadcom's stock surged 11 to 14 percent on the news, reflecting investor confidence in the company's position as a critical infrastructure supplier to large AI developers. The deal underscores how AI training capacity and cost have become the binding constraint on progress for frontier labs, making chip supply and custom silicon central to competitive positioning.