OpenAI's Broadcom deal shows how the AI startup is diversifying its chip suppliers, including Nvidia chips for training and Broadcom chips for inference
How about... everybody? — OpenAI alone has proposed a total of 26 gigawatts of new AI infrastructure *in just the next 4 years*. Total cost: Around a *trillion* dollars. — (gift link) — www.wsj.com/tech/ai/open... … George Pearkes / @peark.es : A good round-up of how OpenAI is approaching custom silicon, as well as some scaling of how much compute they'll need. By @mims.bsky.social
Context & Ripple Effects
OpenAI’s supplier strategy follows its 10GW custom-chip deployment agreement with Broadcom and reporting that the co-developed chips could cost materially less than Nvidia alternatives. The company is now framing training and inference as separate procurement problems rather than a single accelerator purchase.
The scale of proposed infrastructure makes chip availability, cost and workload fit central operating constraints. That turns the earlier Broadcom arrangement from a one-off supply deal into part of a broader multi-vendor compute strategy.
First-order effects
- OpenAI can assign Nvidia hardware to model training while directing inference workloads toward Broadcom-designed chips, reducing its reliance on a single supplier across its compute estate.
- Broadcom gains a defined high-volume use case for its custom silicon, while Nvidia remains positioned in OpenAI’s training stack.
Second-order effects
- Nvidia and other accelerator vendors face greater pressure to compete on workload-specific economics, availability and software fit, not only general-purpose training performance.
- OpenAI’s infrastructure partners must accommodate a more heterogeneous fleet, increasing the importance of deployment, networking and operations built around both training and inference hardware.
Third-order effects
- If large AI operators continue splitting training from inference procurement, custom silicon could become a more durable counterweight to merchant GPUs for steady, high-volume serving workloads.
- The pattern points to AI infrastructure being planned as a long-lived, multi-supplier capacity system, with supply assurance and unit economics increasingly shaping model deployment decisions.
The trend: AI labs are moving from GPU-centric buying toward workload-specific, multi-vendor chip portfolios as inference capacity becomes an infrastructure business of its own.