Meta extends Broadcom deal
Meta reportedly extended its custom AI‑chip partnership with Broadcom through 2029, covering chip design, packaging and networking with an initial deployment commitment above 1 gigawatt. The extension supports Meta’s broader push to build out infrastructure for Llama and large-scale enterprise AI. (qz.com)
Meta and Broadcom have extended their custom artificial intelligence chip partnership through 2029, locking in a multi-year buildout of Meta’s in-house processors. (broadcom.com) The companies said on April 14 that the deal covers Meta Training and Inference Accelerator chips, with an initial deployment commitment above 1 gigawatt and plans for a broader multi-gigawatt rollout. (broadcom.com) Broadcom said the work spans chip design, packaging and Ethernet networking, and that the first deployment uses a 2-nanometer process for Meta’s next generation of artificial intelligence silicon. (broadcom.com; cnbc.com) Meta is building its own chips to handle training and inference, the two core jobs behind large language models and generative artificial intelligence features. Training teaches a model from large datasets; inference is the live step when the model answers a prompt. (cnbc.com) The push comes as Meta ramps spending on artificial intelligence infrastructure. In January, the company said 2026 capital expenditures would reach $115 billion to $135 billion, up from $72.22 billion in 2025. (usnews.com) Meta said the Broadcom-backed systems will support artificial intelligence features across WhatsApp, Instagram and Threads, alongside its broader effort to expand Llama models and enterprise artificial intelligence services. (broadcom.com; about.fb.com) The industry backdrop is a scramble to reduce dependence on Nvidia’s graphics processing units, which remain the standard hardware for many artificial intelligence workloads. Rivals including Google and Amazon have also built custom application-specific integrated circuits, or chips tailored for narrower tasks at lower cost. (cnbc.com) Meta first unveiled its custom silicon effort in 2023 and introduced four new Meta Training and Inference Accelerator versions in March 2026. Broadcom Chief Executive Hock Tan said in March that Meta’s roadmap would scale to multiple gigawatts in 2027 and later years. (cnbc.com) The announcement also reshapes the companies’ governance ties. Meta said Tan, who joined its board in 2024, will not stand for reelection and will instead remain an adviser on Meta’s custom silicon roadmap. (cnbc.com; fool.com) The immediate result is that Meta has committed real power capacity, not just research dollars, to its in-house chip program. The longer test is whether those chips can take a larger share of the work now handled by expensive outside processors. (broadcom.com; cnbc.com)
