Meta pairs with Broadcom on chips

Meta and Broadcom said on April 14 that they had expanded a chip partnership through 2029 to co-develop multiple generations of Meta’s custom artificial intelligence processors. The companies said the agreement covers Meta’s MTIA, or Meta Training and Inference Accelerator, chips and includes Broadcom’s work on chip design, advanced packaging and Ethernet networking. Meta said the initial commitment exceeds 1 gigawatt and described that as the first phase of a sustained multi-gigawatt rollout. Reuters reported the deal extends an existing tie-up as Meta races to add computing capacity for AI features across its apps. ### Why did Meta put Broadcom at the center of this buildout? Meta said Broadcom will help co-develop “multiple generations” of MTIA chips, not a single part, which makes the arrangement a roadmap as much as a product announcement. Meta said the chips are intended to support ranking, recommendations and generative AI workloads across its services, while Broadcom said the work will run over the next three years and extend through 2029. (about.fb.com) Broadcom said the partnership is built on its XPU platform and will use its Ethernet technologies for scale-up and scale-out networking inside Meta’s AI clusters. Mark Zuckerberg said the companies were working across “chip design, packaging, and networking” to build the computing foundation Meta needs, while Broadcom Chief Executive Hock Tan called the first MTIA deployment the start of a multi-generation roadmap. (about.fb.com) ### Why does the 1-gigawatt figure matter? Reuters said Meta’s initial commitment of more than 1 gigawatt of computing capacity is roughly enough to power 750,000 U.S. homes on average. Meta and Broadcom both described that figure as only the first phase, with later deployments expected to move into multiple gigawatts. (about.fb.com) Broadcom had already been framing AI infrastructure around “gigawatt-scale clusters” in product materials released in March. That language places the Meta project in a category where power delivery, network fabric and packaging capacity matter alongside the chip design itself. That is an inference from Broadcom’s description of its AI infrastructure portfolio and the companies’ statements about Meta’s rollout. (finance.yahoo.com) ### What exactly is Meta building with MTIA? Meta said in March that it was developing and deploying four new generations of MTIA chips within two years, a faster cadence than typical chip cycles. Reuters reported the first chip in the current roadmap, the MTIA 300, already powers Meta’s ranking and recommendation systems, with three more due through 2027 and later generations aimed at inference workloads. (investors.broadcom.com) Meta has said MTIA is part of a portfolio approach in which different accelerators are matched to different workloads. In practice, that means Meta is continuing to buy outside processors while also pushing more internal silicon into recommendation and inference tasks where it believes custom hardware can improve performance and total cost of ownership. Meta said that directly in its April and March statements. (about.fb.com) ### How does this reach beyond Meta and Broadcom? TSMC said on May 14 that it was rapidly expanding CoWoS and SoIC advanced packaging capacity as AI demand drives construction of 18 new fabs and advanced packaging facilities worldwide. Reuters separately reported on May 14 that TSMC now expects the global semiconductor market to exceed $1.5 trillion by 2030 as AI drives growth. (about.fb.com) Meta’s own spending plans show the scale of the buildout behind those supply-chain moves. Reuters reported on January 28 that Meta expects 2026 capital expenditures of $115 billion to $135 billion, driven largely by infrastructure costs tied to AI. Taken together, the company’s capex guidance, the Broadcom partnership and TSMC’s packaging expansion show that custom-chip competition is pulling in foundries, packaging lines and data-center infrastructure at the same time. The final sentence is an inference drawn from those disclosed plans and statements. (digitimes.com) ### What changes next? Broadcom said the companies will collaborate over the next three years on next generations of AI accelerator chips, with the agreement running through 2029. Meta said the next step is continued deployment of MTIA generations across its AI infrastructure, while Reuters reported Hock Tan will leave Meta’s board and move to an advisory role on the company’s custom-chip strategy. (broadcom.com) (usnews.com)