US AI Pulse: OpenAI Finally Gets Its Own Chip — Infrastructure Wars Begin

Silicon Valley doesn’t do cautious. OpenAI unveiled its first custom AI chip today — built with Broadcom — and it hit 342 points on Hacker News within hours. This is the kind of story that makes the US AI ecosystem feel like it’s accelerating again.

The story: OpenAI, the company that popularised the term “AI infrastructure” through sheer API usage, has finally built its own silicon. Built on Broadcom’s custom design platform, the chip is reportedly optimised for inference workloads — the compute that happens after a model is trained, when users are actually querying it.

What this means in the US context: The AI chip race just got a new participant.

Google has TPUs. Amazon has Trainium and Inferentia. Meta has MTIA. Microsoft has been quietly building custom silicon through its partnership with Qualcomm and others. OpenAI has been the outlier — heavily dependent on Nvidia’s H100 and B200 GPUs, paying Microsoft billions annually for Azure compute.

That’s a strategic vulnerability that Sam Altman has been vocal about. The custom chip is the fix.

Why it matters: Custom silicon gives OpenAI control over its cost curve. Nvidia’s GPUs are excellent but expensive, and OpenAI is essentially renting someone else’s compute infrastructure. Custom chips — even if they start modestly capable — let OpenAI optimise for its specific models and workloads, reducing cost per query over time.

The competitive angle: This is also a Microsoft calculation. OpenAI is Microsoft’s most important AI partner, but Microsoft also builds its own Maia chips. If OpenAI has its own silicon roadmap, does that change the Microsoft-OpenAI relationship? Microsoft’s Satya Nadella has been characteristically diplomatic about this — but the dynamics underneath are worth watching.

What to watch: Whether this chip can actually compete with Nvidia at scale. Broadcom designs chips well, but the real challenge is the software ecosystem — CUDA, cuDNN, and the entire PyTorch/TensorFlow optimisation stack that makes Nvidia hardware work. OpenAI will need to solve the software problem, not just the silicon one.

The bottom line: OpenAI joining the custom silicon club is a sign that the AI infrastructure wars have moved from “who has the most GPUs” to “who controls the whole stack.” That’s a different kind of moat — and a more durable one.