OpenAI and Broadcom partner on AI hardware (techcrunch.com)

OpenAI announced a partnership with Broadcom to deploy 10 gigawatts of custom AI accelerator hardware—racked systems that will be installed in OpenAI and partner data centers between 2026 and 2029. The companies say OpenAI will “design its own chips and systems” to bake model and product learnings into hardware, though financial terms were not disclosed; the Financial Times has estimated the program could cost somewhere in the hundreds of billions. The deal follows a flurry of recent infrastructure pacts for OpenAI, including multi-gigawatt purchases from AMD and a letter of intent and investment tied to Nvidia hardware. For the AI/ML community, the partnership signals deeper vertical integration and a new phase of accelerator co‑design: custom ASICs or SoCs optimized for transformer workloads, memory bandwidth, interconnects and power efficiency could materially improve training throughput and inference cost per token. Broadcom’s background in switches, networking and silicon suggests the work may span compute, interconnect and rack-level integration—important for scaling models across thousands of nodes. The scale and cost also underscore an escalating arms race for data‑center capacity and supplier relationships, with implications for model performance, operational economics, and industry concentration among a few hyperscalers and chip vendors.