Meta Unveils 4 AI Chips, Buys Millions of Nvidia GPUs
Get the Finance newsletter
Daily finance — markets, central banks, M&A, the prints that move money. Free.
- Meta unveiled four new AI chips (MTIA 300, 400, 450, 500), with the 400/450/500 rolling out beginning in early 2027 for generative AI inference, using a modular chiplet design that iterates every six months rather than the typical two-year cadence.
- Broadcom supplies SerDes components, packaging, and manufacturing support for Meta's custom XPUs and counts Meta as one of its five major XPU customers, with CEO Hock Tan arguing workload-specific silicon outperforms general-purpose GPUs running software kernels.
- Meta simultaneously signed a multiyear deal with Nvidia to deploy millions of Nvidia Blackwell and Rubin chips plus CPUs and SpectrumX ethernet switches, suggesting homegrown chips serve legacy recommendation workloads while Nvidia handles frontier LLM training and inference.
- Nvidia acquired inference chip startup Groq's intellectual property and engineering talent for $20 billion late last year, a signal it sees emerging demand for non-GPU silicon as the industry pivots from training to inference.
- AI compute demand is still growing exponentially, so new inference-focused chips are likely incremental rather than displacing Nvidia's training-focused GPUs—creating a rising-tide dynamic in which both custom XPUs and Nvidia GPUs expand.
- The MTIA 300 is already in use for Meta's core ranking and recommendation workloads, while the 450 doubles the 400's high-bandwidth memory and the 500 pushes HBM 50% higher still—a tiered roadmap for escalating generative AI inference demands.
Why it matters: Meta is building custom chips while also signing multiyear deals for millions of Nvidia GPUs—contradicting the narrative that custom XPU silicon displaces Nvidia. The $20B Groq acquisition shows Nvidia hedging its own GPU-centric thesis, and Meta's parallel procurement strategy suggests inference-specific chips will carve out complementary niches rather than dethrone Nvidia's training and frontier-AI dominance.
