Meta and Broadcom are cranking their AI relationship up several notches, unveiling an expanded, multiyear deal to co-develop Meta's custom AI accelerators and bulk up the social giant's AI compute footprint. The agreement starts with more than 1 gigawatt of MTIA (Meta Training and Inference Accelerator) compute and sketches out a long-term plan for sustained, multi-gigawatt rollouts. As part of the arrangement, Broadcom chief executive Hock Tan will leave Meta's board and shift into an advisory role tied to the partnership.
Meta says the collaboration spans chip design, advanced packaging and networking, and that Broadcom's XPU platform will sit at the heart of future MTIA generations, according to Meta Newsroom. The company describes the initial phase as exceeding 1GW and casts the effort as part of its broader push to bring "personal superintelligence" features across its apps. Meta is positioning the deal as an acceleration of its in-house MTIA work rather than a pivot away from its other silicon suppliers.
Broadcom, for its part, is calling the pact a multigeneration strategic partnership running through 2029, with the company set to supply high-radix Ethernet switches, optical connectivity, PCIe switches and SerDes to power MTIA clusters, according to Bar Chart. Its release describes deep co-design work across logic, memory and high-speed I/O and labels the first 1-plus-gigawatt deployment as just the opening phase of a longer rollout.
What MTIA Is
Meta has laid out a four-generation MTIA roadmap that targets chips optimized for inference and recommendation workloads on a faster-than-usual cadence, according to Tom's Hardware. Coverage of the MTIA family highlights modular chiplets, a big emphasis on higher HBM bandwidth for inference and a roughly six-month release rhythm, Tom's Hardware reported. That mix of packaging, memory and interconnect requirements helps explain why Meta and Broadcom are leaning into tightly coupled co-development instead of a simple supplier relationship.
Silicon Valley Impact
Because both companies concentrate much of their engineering in the Bay Area, the partnership is poised to push even more design, packaging and systems work through local contractors and suppliers. The deal drew attention from the local business press, including the Silicon Valley Business Journal. Earlier coverage of Meta's planned $10 billion data campus in Lebanon has already shown how gigawatt-scale compute demand can ripple out into jobs and infrastructure, as the project dubbed $10B Lebanon data campus detailed. In the near term, local firms specializing in advanced packaging, optics and thermal engineering could be among the main beneficiaries.
Where This Fits In The AI Race
The Broadcom deal slots neatly into a broader AI strategy that favors multi-vendor, vertically integrated stacks. Across the industry, hyperscalers are locking in multi-gigawatt agreements to secure long-term compute capacity and reduce dependence on any single chipmaker. Meta recently rolled out an expanded agreement with AMD that could see up to 6 gigawatts of Instinct GPUs deployed, according to AMD. Analysts say that approach, which blends in-house accelerators, GPUs from multiple vendors and custom networking, is aimed at squeezing more efficiency and resilience out of massive inference workloads.
Board Shift And Legal Notes
Broadcom confirmed that Hock Tan will step off Meta's board and move into an advisory role focused on the MTIA silicon roadmap, according to Bar Chart. Both companies also included the usual forward-looking statements that flag risks and uncertainties tied to execution. Corporate-governance watchers will be keeping an eye out for any SEC filings or proxy materials that formally document the change and spell out any related compensation or conflict-of-interest details.
Over the coming quarters, expect a drip of technical roadmaps, engineering milestones and supply-chain updates as Meta and Broadcom move from design to packaging to full-scale deployment. What is clear from the announcement is that custom accelerators, advanced packaging and high-bandwidth networking are set to anchor the next generation of commercial AI infrastructure, with a hefty slice of that work rooted in Silicon Valley.
