Microsoft and Meta have both sharply reduced employee use of Anthropic's Claude. Microsoft cut its projected internal Claude spend by more than a third from over $1 billion per year, and slashed per-employee AI budgets in its 60,000-person cloud and AI group from $100,000 to about $10,000 per month. Meta's Claude Code user base fell from approximately 60,000 to 30,000 employees, with staff shifting to Meta's in-house MetaCode and Muse Code tools.
Microsoft has also announced that GitHub Copilot will use a model router to direct routine coding tasks away from Claude Haiku 4.5 to its own MAI-Code-1.1-Flash, a 137 billion total parameter model with only 6.8 billion active parameters, 3-bit quantization, and a 256,000-token context window. Microsoft claims the model matches full-precision performance on SWE-Bench Verified and Terminal-Bench 2.1.
Satya Nadella announced the Maia chip family will become the primary inference system for Microsoft products starting in 2027, scaled to multigigawatt deployments. Maia 200 is already live in the US, Europe, and Australia running Copilot and Teams workloads at up to 1.7x more tokens per watt than leading GPU systems. The Maia 300, built on TSMC's 2nm process with eight HBM4 memory stacks, delivers 40% more tokens per megawatt than leading GPU systems, and a full 96-chip rack produces up to 7x the token output of a Maia 200 rack. Microsoft is reportedly negotiating with TSMC for over 300,000 Maia 300 units for 2027 delivery, though no purchase order has been signed. NVIDIA fell about 1.3% on the Maia 300 announcement while Microsoft shares rose.
An AI newsletter author said Microsoft previously provided over $2 billion in ARR to Anthropic and added they are not worried about Anthropic's revenue numbers. MilkRoadAI characterized the Maia expansion as the reason for the Claude spending cuts, stating that each token moved off Claude and onto Maia is money that stays with Microsoft.