Microsoft Maia 200: 3nm custom AI chip aimed at Azure Copilot — 30% more punch per cost claim

Azure wants fewer NVIDIA invoices in the Copilot bill. Microsoft’s Maia 200 — a 3-nanometer custom AI accelerator discussed in early-August infrastructure coverage — is pitched at roughly 30% higher performance than prior Maia silicon at comparable cost, headed for data centers in the western and central United States to run Microsoft 365 Copilot and Azure AI Foundry jobs.

Where it sits in the stack

Maia is Microsoft’s answer to Google TPU and Amazon Trainium: captive silicon for captive workloads. It does not kill GPU merchant buys overnight; it trims the most repetitive inference/training slices where a custom part wins on tokens-per-watt. CFO messaging still stresses reliability of Azure and Copilot even while AI capex stays huge.

Why it matters this week

Hyperscalers are racing unit volume (Google’s 12–15M TPU chatter) and software agents (Project Solara). Maia 200 is the chip half of Microsoft’s same bet: own more of the stack before margin evaporates to GPU vendors.

Watch Azure region capacity notes — that is where Maia stops being a slide and becomes a rack.