Meta–Broadcom MTIA Extension

Meta and Broadcom have extended their custom AI chip partnership through 2029, locking in Meta’s in-house MTIA accelerators for a multi-year data center buildout. (about.fb.com) Meta said the agreement covers chip design, advanced packaging, and networking, and includes an initial deployment commitment of more than 1 gigawatt of compute. Broadcom said that first phase is part of a broader multi-gigawatt rollout. (about.fb.com) (investors.broadcom.com) The companies said the roadmap includes multiple future generations of Meta Training and Inference Accelerator chips, or MTIA, with Broadcom calling the next design the industry’s first 2-nanometer AI compute accelerator. Meta said those chips are meant to run AI features across WhatsApp, Instagram, Threads, and its other apps. (investors.broadcom.com) (about.fb.com) An inference chip is built to answer a user’s prompt after a model has already been trained, like serving search results or chatbot replies at high volume. Meta said MTIA is optimized for inference and recommendation systems, where speed and power costs matter across billions of daily requests. (about.fb.com 1) (about.fb.com 2) That focus marks a different bet from the graphics processing units, or GPUs, that have dominated the AI boom. CNBC reported that hyperscalers including Meta are building narrower application-specific chips as an alternative to more expensive general-purpose AI hardware from Nvidia and AMD. (cnbc.com) Meta had already laid out that shift in March, when it unveiled four MTIA generations — 300, 400, 450, and 500 — on a roughly six-month cadence through 2027. The company said MTIA 300 is already in production, while the later chips are aimed mainly at generative AI inference. (about.fb.com) (cnbc.com) Meta said the jump from MTIA 300 to MTIA 500 would raise high-bandwidth memory throughput by 4.5 times and compute by 25 times. High-bandwidth memory is the chip’s fast local memory, and more of it helps feed large AI models without stalling the processor. (about.fb.com) (rcrtech.com) Broadcom’s role is not just the chip itself. Both companies said the partnership also relies on Broadcom’s Ethernet networking so Meta can connect large clusters of accelerators without bottlenecks inside expanding AI data centers. (investors.broadcom.com) (about.fb.com) Meta is still buying outside hardware as it builds MTIA. CNBC reported in March that the company’s chip push is meant to diversify supply and lower costs, not replace every Nvidia or AMD processor in its infrastructure. (cnbc.com) The new agreement gives Meta a longer runway to build AI systems with more of its own parts, while Broadcom becomes a deeper supplier across the chip, the package, and the network around it. The opening number in the deal is 1 gigawatt, but both companies are already framing that as the first phase. (about.fb.com) (investors.broadcom.com)