Microsoft negotiates running Anthropic Claude on Maia 200 chips via Azure

Microsoft is in early negotiations with Anthropic to run Claude workloads on Maia 200 chips, the custom AI processor developed internally by Microsoft for Azure use. If the deal materializes, it will be the first case of a third-party model running on Microsoft's proprietary chip at scale — an important signal about the company's direction in AI infrastructure.

The Maia 200 was announced in 2023 and began being used internally by Microsoft to train and run inference on the Phi model family. Expanding it to host Claude would be critical external validation: it means the chip does not serve only proprietary models — it can compete with NVIDIA H100 for third-party models.

What drives Microsoft in this direction

Microsoft invested over $13 billion in OpenAI. Anthropic's Claude is a direct competitor to ChatGPT. At first glance, the negotiation seems contradictory.

But the logic is industrial: Microsoft wants to be the best place to run AI — any AI. If Azure is where companies choose to host Claude, Microsoft wins the infrastructure layer regardless of which model wins the quality war. This is the same logic as AWS, which hosts both Claude and GPT-4 without declared preference.

Additionally, having Claude run on Maia 200 (instead of NVIDIA GPUs) reduces Microsoft's dependence on NVIDIA supply — which has been a chokepoint for the entire industry.

What changes for Azure users

For those already using Claude via Azure AI Studio: the change would be transparent at the API layer — the model would respond the same way, but would run on Microsoft hardware instead of NVIDIA. The potential benefit is lower inference cost, since Microsoft controls the chip margin.

For those evaluating where to host AI workloads: the possibility of running Claude on Maia 200 expands cost optimization options in Azure.

The larger implication

The custom silicon trend is accelerating. Apple, Google (TPU), Amazon (Trainium/Inferentia), Microsoft (Maia) and Qualcomm are all developing specialized AI chips. NVIDIA still dominates, but hardware diversification is reducing dependence on a single vendor — which is structurally good for compute buyers.

For 10Dobro, which uses Claude as the orchestrator of 26 systems, any move that reduces the cost of Claude inference via Azure is positive in the medium term.