Meta and Broadcom expand custom AI silicon work

Meta and Broadcom said on April 14 they are extending their custom chip partnership through 2029 to build several generations of Meta’s in-house artificial intelligence processors. (about.fb.com) A custom artificial intelligence chip is a processor built for one company’s own workloads, instead of a general-purpose chip sold to many customers. Meta said Broadcom will help co-develop multiple generations of Meta Training and Inference Accelerator, or MTIA, the chip family Meta uses for its own artificial intelligence systems. (about.fb.com) Broadcom said the agreement covers technology for MTIA chips, advanced packaging, and Ethernet-based rack-scale interconnects, which are the links that move data between servers inside large data centers. Broadcom said the plan runs through 2029 and starts with a deployment commitment of more than 1 gigawatt of Meta custom silicon. (investors.broadcom.com) Meta has been building MTIA for several years as an alternative for specific jobs that do not need the flexibility of merchant graphics processing units, or graphics chips sold off the shelf. In April 2024, Meta said its second-generation MTIA was designed for deep learning recommendation models that rank content and ads across its apps. (about.fb.com, engineering.fb.com) The company widened that effort in March 2026, when it said it was developing and deploying four new generations of MTIA chips within two years for ranking, recommendations, and generative artificial intelligence workloads. Meta said that pace is faster than a typical chip cycle. (about.fb.com) Meta is not replacing outside suppliers with one internal chip program. In February 2026, Meta said it had also signed a long-term infrastructure agreement with Advanced Micro Devices and described its strategy as a mix of partner hardware and its own MTIA silicon. (about.fb.com) That mixed approach reflects the scale of Meta’s artificial intelligence build-out. In April 2024, Meta said it was designing data centers around clusters of 24,576 graphics processing units and using a new network fabric to connect them for training large models and serving recommendations. (about.fb.com) Broadcom framed the new deal as part of that larger build-out, saying Meta’s custom silicon rollout is the foundation for a sustained multi-gigawatt deployment in state-of-the-art artificial intelligence data centers. Reuters reported on April 14 that the expanded deal is aimed at producing several generations of processors as Meta races to add computing capacity for artificial intelligence features across its apps. (investors.broadcom.com, msn.com) The immediate result is not a consumer product launch but a longer supply and design commitment: Meta gets a chip roadmap tuned to its own software, and Broadcom gets a central role in one of the biggest disclosed in-house artificial intelligence silicon rollouts so far. (about.fb.com, investors.broadcom.com)