OpenAI says it built a homegrown AI inference chip in nine months, with Broadcom (AVGO) support

OpenAI says it developed its first homegrown inference chip in about nine months, moving from initial design through tape-out, a manufacturing step that finalizes a chip layout for production. Inference chips are optimized for running already-trained AI models, typically in real-world applications where power efficiency and latency matter as much as raw performance.

The company also claims the new chip can perform up to 1.9 times more AI work per watt than the comparison systems it references. “Per watt” is a common efficiency yardstick in data centers because it ties directly to power and cooling costs for large-scale deployments.

Yahoo Finance reports that the rapid timeline and performance claims position OpenAI’s custom hardware as a counterweight to the assumption that top-tier AI inference is dominated by leading accelerators. The story frames the “investable surprise” less around the existence of custom silicon and more around the supply-chain and partner ecosystem that makes such a chip feasible on an accelerated schedule.

Broadcom is named as one of the contributors to the effort. While the reporting does not spell out the specific Broadcom components or engineering work involved, Broadcom is broadly known in the market for supplying networking and custom silicon used in data-center systems. The implication in the account is that Broadcom’s role helped bridge the gap between OpenAI’s chip design ambitions and a manufacturable, deployable compute stack.

The angle for Broadcom investors is that even when a hyperscaler builds its own chips, the broader platform still requires supporting hardware to move data and integrate systems. If OpenAI’s inference roadmap increasingly relies on custom silicon, suppliers of interconnect, system components, and other “plumbing” elements can still benefit, even if they are not the headline chip designer.

Sector context matters because the AI hardware market is in a phase shift. Training remains GPU-heavy, but inference is where the spending ramps as models move into production and where operators push for efficiency. The market has treated custom accelerators as one of the most direct ways for large AI users to reduce power cost per inference, making supplier relationships and system integration as important as the silicon itself.

What remains unclear from the published account is how Broadcom’s support breaks down in technical terms, whether it was tied to specific packaging, networking, or other platform elements, and how much of OpenAI’s broader inference systems will use the new chip versus alternative hardware. The reporting also does not provide details on volume commitments, timelines for deployment, or any commercial terms tied to Broadcom’s involvement.

Looking ahead, investors and engineers will likely focus on two questions: whether OpenAI can sustain a fast cadence for follow-on inference chips, and whether its efficiency and deployment claims translate into broader adoption across workloads. Separately, the market will watch whether custom-silicon announcements like this lead to more measurable changes in how suppliers like Broadcom are integrated into next-generation data-center systems.