Meta’s gigawatt chip deal

Meta has agreed to deploy more than one gigawatt of custom artificial intelligence chips with Broadcom, locking in several future processor generations through 2029. (cnbc.com) The April 14 agreement expands an existing partnership around Meta Training and Inference Accelerator chips, or MTIA, Meta’s in-house processors for artificial intelligence work across Facebook, Instagram, WhatsApp and its other apps. Broadcom said the first phase exceeds 1 gigawatt and is part of a larger multi-gigawatt rollout. (about.fb.com) Broadcom said it will supply technology for the MTIA program as well as Ethernet-based rack-scale interconnects, the networking links that tie large groups of chips together inside artificial intelligence data centers. Reuters reported the expanded tie-up covers several generations of processors. (broadcom.com) MTIA is Meta’s custom silicon family, built to run specific artificial intelligence jobs more efficiently than buying only general-purpose graphics processors from outside vendors. Meta said in March that it was developing and deploying four new MTIA generations within two years. (about.fb.com) Those chips are aimed first at inference, the step where a trained model produces recommendations, rankings, ads, images or other outputs for users in real time. Meta said it already deploys hundreds of thousands of MTIA chips for inference across organic content and ads on its apps. (about.fb.com) Meta’s current roadmap runs on a fast cadence. CNBC reported on March 11 that MTIA 300 had already been deployed, while MTIA 400, MTIA 450 and MTIA 500 were scheduled to follow at roughly six-month intervals, with the later chips aimed at more generative artificial intelligence inference tasks. (cnbc.com) The deal fits a broader pattern in which large cloud and app companies are designing more of their own chips while leaning on Broadcom for custom design, packaging and networking. On April 6, Broadcom said it had also expanded artificial intelligence chip work with Google and added a 3.5-gigawatt capacity arrangement tied to Anthropic. (cnbc.com) Broadcom has been telling investors that this market is getting much larger. In March, Chief Executive Hock Tan said Broadcom could generate artificial intelligence chip revenue in 2027 “significantly in excess of $100 billion,” and analysts cited nearly 10 gigawatts of capacity across six customers. (cnbc.com) Meta is still buying outside capacity while it builds its own. On April 9, CNBC reported that Meta committed another $21 billion to CoreWeave cloud infrastructure from 2027 through 2032, on top of an earlier $14.2 billion arrangement centered on Nvidia graphics processors. (cnbc.com) The announcement also reshaped the board relationship between the two companies. Meta said Broadcom Chief Executive Hock Tan decided not to stand for re-election to Meta’s board after notifying the company on April 8, and Broadcom said he will instead advise Meta on custom chip strategy. (sec.gov) For Meta, the new commitment puts a concrete power number on its in-house chip push: more than one gigawatt in the first phase, with Broadcom and Meta now planning for multiple generations after that. (reuters.com)