Hyperscalers pick custom silicon
Multiple market pieces report that major AI buyers are deepening bets on application‑specific silicon—Google’s TPU strategy was highlighted as a tailwind for suppliers like Marvell, and Broadcom reportedly won a multi‑generation Meta AI silicon deal. ( ) Analysts framed Anthropic, Alphabet and Broadcom moves as competitive pressure on general‑purpose GPU incumbents. (fool.com)
The biggest artificial intelligence buyers are putting more work onto custom chips instead of relying only on standard graphics processors. (blog.google) Google said on April 9, 2025 that its seventh-generation Tensor Processing Unit, called Ironwood, was built specifically for inference, the stage when a trained model answers prompts. Google said Ironwood offers five times the peak compute capacity and six times the high-bandwidth memory of the prior generation. (blog.google) Meta said on April 14, 2026 that it expanded its partnership with Broadcom to co-develop multiple generations of Meta Training and Inference Accelerator chips. Broadcom said the agreement sets a “multi-generation blueprint” for future versions of Meta’s internal accelerator portfolio. (about.fb.com; broadcom.com) A custom chip is a processor built for one buyer’s workloads, the way a delivery fleet might use vans designed for packages instead of every driver buying the same pickup truck. Meta said its Meta Training and Inference Accelerator chips are meant to match “different accelerators for different workloads” across ranking, recommendations, and generative artificial intelligence. (about.fb.com; about.fb.com) That shift has been building for months across the largest cloud and model companies. Anthropic said on April 6, 2026 that it trains and runs Claude on Amazon Web Services Trainium, Google Tensor Processing Units, and NVIDIA graphics processors, rather than a single hardware stack. (anthropic.com) Amazon Web Services has been making the same pitch with its own silicon. Amazon says its Trainium family is built for training and inference, and Anthropic said in November 2024 that Amazon Web Services became its primary cloud and training partner in an expanded deal that included a new $4 billion Amazon investment. (aws.amazon.com; anthropic.com) Meta’s internal roadmap shows how fast that effort is moving. The company said in March 2026 that it plans four new generations of custom chips within two years, a pace it described as faster than typical chip cycles, to support recommendations and generative artificial intelligence workloads. (about.fb.com) None of this means the big buyers are abandoning general-purpose graphics processors. Anthropic said last week that NVIDIA chips remain part of its mix, and Meta said on February 17, 2026 that it also signed a long-term infrastructure partnership with NVIDIA. (anthropic.com; about.fb.com) The immediate change is that the largest customers now want several kinds of accelerators inside the same data center, with custom silicon handling more of the repetitive, high-volume work. That leaves Broadcom, Google, Amazon, and Meta pushing a portfolio model in which cost, power use, and workload type decide which chip gets the job. (broadcom.com; about.fb.com; anthropic.com)
