Anthropic in early talks to run Claude on Microsoft's Maia 200 chip

CNBC reported on 21 May 2026 that Anthropic and Microsoft are in early discussions about running Claude inference on Microsoft's Maia 200 custom AI chip via Azure. The report — citing sources familiar with the talks — was clear that no agreement has been reached and discussions remain at an early stage. The day before, Anthropic signed a deal to pay xAI $1.25 billion per month for compute on Colossus (TechCrunch, 20 May). Two diversification moves in 48 hours is a signal, not a coincidence.

Per Microsoft, Maia 200 launched in January 2026 with TSMC 3nm and 140+ billion transistors, 216GB HBM3e at 7 TB/s, and >10 petaFLOPS in FP4. Microsoft claims >30% better performance (tokens) per dollar than its prior fleet — all vendor figures, not yet independently benchmarked. The chip is inference-only (FP4/FP8); training still needs Nvidia. It already runs OpenAI GPT-5.2 and Microsoft 365 Copilot in production — it is not a prototype.

Watch for a pilot announcement — a bounded Claude workload on Maia 200 — as the signal that talks have moved to execution. If that happens, Claude becomes the first external frontier model to validate Microsoft's custom silicon at scale, a significant proof point for the chip's commercial viability beyond Microsoft's own products. The broader question: if both Anthropic and OpenAI serve inference from Maia 200, Nvidia's lock on the serving layer weakens materially.