Meta expands custom‑chip tie with Broadcom

Meta and Broadcom have extended their custom artificial intelligence chip partnership through 2029, with an initial deployment commitment of more than 1 gigawatt for Meta’s in-house Meta Training and Inference Accelerator systems. (broadcom.com) Broadcom said on April 14 that it will supply technology for multiple generations of Meta Training and Inference Accelerator chips and for Ethernet-based rack-scale interconnects, the networking gear that links large groups of servers inside data centers. (broadcom.com) Meta’s custom chips are built for inference, the step where a trained model produces answers, rankings, or recommendations for users, rather than for the heaviest model-training work. Meta said in April 2024 that its next-generation Meta Training and Inference Accelerator delivered three times better performance than the first version on recommendation models. (engineering.fb.com) Meta has already been using Meta Training and Inference Accelerator hardware in production. The company said in September 2025 that the chip was deployed at scale in its data centers, primarily for advertising workloads, and had produced efficiency gains over vendor silicon. (engineering.fb.com) The new agreement pushes that effort from a chip project into a data-center buildout. Broadcom said the first phase alone exceeds 1 gigawatt, and described it as the opening stage of a sustained multi-gigawatt rollout for Meta’s artificial intelligence infrastructure. (broadcom.com) Meta is still buying large volumes of graphics processing units from outside suppliers while it develops its own silicon. In an April 2 engineering post, Meta said its infrastructure runs on a mix of NVIDIA graphics processing units, Advanced Micro Devices graphics processing units, Meta Training and Inference Accelerator chips, and central processing units. (engineering.fb.com) That mix explains why Broadcom’s role now includes networking as well as chip support. Broadcom said the expanded deal covers rack-scale interconnects for “Meta Training and Inference Accelerator grids,” which are clusters of custom chips tied together to handle larger artificial intelligence workloads. (broadcom.com) The announcement also reshapes the companies’ governance ties. CNBC reported that Broadcom Chief Executive Hock Tan will leave Meta’s board after two years and shift to an advisory role on Meta’s custom-chip strategy under the expanded agreement. (cnbc.com) For Meta, the deal locks in a longer runway for chips tailored to its own apps, from ad ranking to recommendation systems and generative artificial intelligence services. For Broadcom, it secures a marquee customer for custom silicon and data-center networking through the end of the decade. (broadcom.com)