Meta strikes Broadcom chip deal

Meta and Broadcom have expanded their chip partnership through 2029, with Meta committing an initial deployment of more than 1 gigawatt of custom artificial intelligence hardware. (about.fb.com) The companies said on April 14 that Broadcom will help Meta design multiple generations of Meta Training and Inference Accelerator chips, along with the packaging and networking needed to run them in artificial intelligence data centers. (broadcom.com) Meta said the first phase is more than 1 gigawatt, with a larger multi-gigawatt rollout planned over time, and Broadcom said the agreement now runs through 2029. (about.fb.com) (broadcom.com) The chips are meant for inference and recommendation work, which is the stage when a trained model answers prompts or ranks what users see in products like Instagram, WhatsApp, and Threads. Meta said its Meta Training and Inference Accelerator program is built to handle those jobs at lower total cost than using the same hardware for every task. (about.fb.com) (economictimes.indiatimes.com) That puts Meta in the same camp as Google and Amazon, which have spent years building in-house chips to reduce dependence on Nvidia’s and Advanced Micro Devices’ general-purpose graphics processors. CNBC reported that hyperscalers are pursuing application-specific integrated circuits because they can be cheaper and more efficient for narrower workloads. (cnbc.com) Broadcom’s role goes beyond the chip itself. The company said its Ethernet technology will connect Meta’s artificial intelligence clusters, and its XPU platform will be used to co-develop the custom accelerators across several generations. (broadcom.com) Broadcom also said the Meta Training and Inference Accelerator chips in this program are slated to use a 2 nanometer manufacturing process. CNBC reported Broadcom shares rose about 3% in extended trading after the announcement, while Meta shares were little changed. (broadcom.com) (cnbc.com) The deal came with a governance change. Meta said Broadcom Chief Executive Hock Tan will leave Meta’s board and move into an advisory role focused on Meta’s custom silicon roadmap; CNBC reported Tan joined the board in 2024 and told Meta last week he would not seek reelection. (about.fb.com) (cnbc.com) Meta has been building toward this for several years. CNBC said the company first unveiled its custom silicon in 2023, and Reuters reported that the first Meta Training and Inference Accelerator chip, the MTIA 300, is already running ranking and recommendation systems, with three more chips due by 2027. (cnbc.com) (economictimes.indiatimes.com) For Meta, the immediate result is simple: more of its artificial intelligence workload will run on hardware it helped design itself, and Broadcom will be one of the main companies building the plumbing behind that shift. (about.fb.com) (broadcom.com)