Meta and Broadcom on custom AI silicon

Meta and Broadcom said on April 14 they will co-develop several generations of Meta’s custom artificial intelligence chips through 2029. (about.fb.com) The deal starts with more than 1 gigawatt of Meta Training and Inference Accelerator capacity and is framed as the first phase of a multi-gigawatt rollout for Meta’s artificial intelligence data centers. (broadcom.com) A custom chip is a processor built for one company’s own workloads, rather than a general-purpose graphics processing unit sold to many buyers. Meta said its Meta Training and Inference Accelerator chips are tuned for ranking, recommendations, ads, and generative artificial intelligence inference across Facebook, Instagram, WhatsApp, and Threads. (about.fb.com) Inference is the step where a trained model answers a prompt, ranks a post, or generates a reply, and it has become a major cost center as Meta pushes artificial intelligence features to billions of users. Meta said it already deploys hundreds of thousands of Meta Training and Inference Accelerator chips for inference on organic content and ads. (about.fb.com) Meta first introduced the Meta Training and Inference Accelerator program in 2023, and on March 11, 2026, it laid out four new chip generations to be deployed within two years. The first of those, Meta Training and Inference Accelerator 300, is already in production for ranking and recommendation training, while Meta Training and Inference Accelerator 400, 450, and 500 are aimed mainly at generative artificial intelligence inference into 2027. (about.fb.com) Broadcom said the partnership also covers packaging and Ethernet networking, the links that connect large groups of chips so they can work as one system inside a data center. The company said Meta’s initial deployment will use Broadcom’s advanced Ethernet technology to reduce bottlenecks across expanding compute clusters. (broadcom.com) The arrangement does not replace outside chip suppliers. Meta said it is taking a portfolio approach by sourcing silicon from several industry companies while keeping its own Meta Training and Inference Accelerator line at the center of its infrastructure plan. (about.fb.com) Other cloud giants have followed a similar path. CNBC reported that Google began shipping Tensor Processing Units in 2015 and Amazon announced its first custom chip in 2018, as large operators looked for alternatives to expensive, supply-constrained graphics processing units from Nvidia and Advanced Micro Devices. (cnbc.com) Broadcom has become a central supplier for that shift because it helps hyperscalers design application-specific integrated circuits, or chips built to one customer’s specifications. In its 2025 annual report, Broadcom said those custom products include accelerators for hyperscalers, frontier-model companies, and system integrators. (investors.broadcom.com) The announcement also changes the governance tie between the two companies. Reuters and CNBC reported that Broadcom Chief Executive Hock Tan will leave Meta’s board and take an advisory role on Meta’s custom chip strategy. (finance.yahoo.com) (cnbc.com) Meta’s message is that it still needs huge amounts of compute, but it wants more control over how that compute is built, connected, and deployed. Broadcom’s message is that the custom-chip business is moving from one-off projects to multi-year infrastructure programs measured in gigawatts. (about.fb.com) (broadcom.com)