Meta–Broadcom Chip Deal
- Meta expanded its partnership with Broadcom to co-develop multiple generations of custom AI accelerator chips. - The alliance covers chip development through 2029, moving significant AI hardware design in-house. - That shift increases capital spending on infrastructure and raises the value of infrastructure and systems literacy in engineering roles. (simplywall.st)
Meta and Broadcom have widened their chip partnership through 2029 to build multiple generations of Meta’s in-house artificial intelligence processors. (about.fb.com) Meta said the deal covers its Meta Training and Inference Accelerator, or MTIA, the custom chips it uses to run recommendation systems and generative artificial intelligence across Facebook, Instagram, WhatsApp, and its other services. Broadcom said it will supply chip technology, advanced packaging, and Ethernet networking for the program. (about.fb.com) (broadcom.com) The companies said the first deployment commitment is more than 1 gigawatt of computing capacity, with a broader multi-gigawatt rollout planned after that. CNBC reported the announcement on April 14 and said Broadcom Chief Executive Hock Tan will leave Meta’s board as the partnership expands. (broadcom.com) (cnbc.com) An artificial intelligence accelerator is a chip built for a narrow set of jobs, the way a forklift handles warehouse loads more efficiently than a family car. Meta said it uses a “portfolio approach” to match different chips to different workloads instead of relying on one processor for everything. (about.fb.com) That approach puts more of Meta’s hardware design inside the company at the same time its infrastructure budget is swelling. In January, Meta told investors it expects 2026 capital expenditures of $115 billion to $135 billion, up from $72.22 billion in 2025. (investor.atmeta.com) Meta is also building out the physical footprint to use that hardware. In February, the company said a new campus in Lebanon, Indiana, will be its 27th U.S. data center and 31st worldwide. (datacenters.atmeta.com) Broadcom framed the deal as more than chip design. Its announcement said Meta’s MTIA systems will use Broadcom’s Ethernet-based rack-scale interconnects, the links that move data between servers inside a data center. (broadcom.com) Reuters reported the agreement as part of a broader push by Meta to use custom silicon for both training and inference, the two main phases of artificial intelligence work: building models and then running them for users. That gives Meta another route to add capacity beyond buying standard chips from outside suppliers. (money.usnews.com) The thread running through the announcement is control: control over chip design, control over network architecture, and control over how fast new data centers can come online. Meta and Broadcom have now tied that buildout to a timetable that runs through 2029. (about.fb.com) (broadcom.com)
