Big tech pushes custom chips
Several large firms are pressing to reduce reliance on Nvidia by building custom AI silicon — Meta and Broadcom have deepened a partnership aimed at AI‑infrastructure independence. Nvidia still projects strong scale: CEO Jensen Huang said the company has the supply‑chain capacity to support growth “at a trillion dollars in scale” and argued the U.S. should continue selling chips to China. The shift frames the AI competition around who controls hardware as much as software. (markets.financialcontent.com) (dwarkesh.com)
Meta and Broadcom expanded their custom-chip partnership on April 14, extending Meta’s in-house artificial intelligence silicon plans through 2029. (about.fb.com) Meta said the deal covers “multiple generations” of its Meta Training and Inference Accelerator chips, which it uses for both training systems and the inference work that serves artificial intelligence features inside its apps. Broadcom said the initial deployment commitment exceeds one gigawatt and is the first phase of a larger multi-gigawatt rollout. (about.fb.com) (investors.broadcom.com) A custom chip is a processor built for one company’s own workloads rather than sold as a general-purpose product. Meta said its Meta Training and Inference Accelerator family is meant to lower the cost and power use of ranking, recommendation and generative artificial intelligence jobs in its data centers. (about.fb.com) Meta started the chip program in 2023 and said in March that it plans to develop and deploy four new generations within two years. In April 2024, the company said those chips would sit alongside graphics processing units, not replace them outright. (about.fb.com 1) (about.fb.com 2) The push comes as the largest artificial intelligence buyers try to control more of the hardware stack inside their own facilities. Reuters reported on April 14 that the Meta-Broadcom agreement extends support for Meta’s training and inference accelerators through 2029. (money.usnews.com) Nvidia is still arguing that scale favors its model. In an interview published April 15, Chief Executive Jensen Huang said, “If our next several years are a trillion dollars in scale, we have the supply chain to do it,” and he said the United States should keep selling chips to China. (dwarkesh.com) Meta’s announcement also tied the chip deal to networking gear, saying Broadcom will help supply the technology that links large clusters of processors inside Meta’s artificial intelligence data centers. Reuters and Yahoo Finance both described the arrangement as covering Broadcom’s Ethernet technology as well as Meta’s accelerators. (money.usnews.com) (finance.yahoo.com) The companies paired the announcement with a governance change. CNBC reported that Broadcom Chief Executive Hock Tan agreed to leave Meta’s board as the partnership was extended, and said the design agreement runs through 2029. (cnbc.com) The contest is no longer only about who has the best model. It is also about who can afford, design and connect enough chips to run artificial intelligence services at the scale of billions of users. (about.fb.com) (dwarkesh.com)
