Anthropic is in talks with Microsoft about using servers with Azure Maia, Microsoft’s proprietary AI chip. According to The Information, the negotiations are still in the early stages. Maia 200 is designed to run AI inference as efficiently as possible. Microsoft is one of Anthropic’s largest customers, having committed $30 billion in Azure spending.
The chip in question is the Maia 200, which Microsoft developed specifically for inference. In other words: running existing AI models, not training new ones. Microsoft has already realized cost savings with this chip through Copilot tools, which support both OpenAI and Anthropic models. We recently discussed this AI processor with Andrew Wall, General Manager of Azure Maia at Microsoft.
Read more: Microsoft’s Azure Maia chief on the complex future of AI compute
Strong investment relationship
A substantial financial relationship already exists between Microsoft and Anthropic. Late last year, Microsoft announced plans to invest up to $5 billion in Anthropic. In November, a strategic partnership with Nvidia was also announced, under which Microsoft and Nvidia will jointly invest up to $15 billion in Anthropic. In turn, Anthropic has committed to $30 billion in Azure spending.
Part of that collaboration involves the use of Claude models for Microsoft Copilot, with a value of at least $500 million for that deal. Microsoft has also allocated additional existing Nvidia server capacity to Anthropic and is building new server clusters specifically for the AI company.
Proprietary chips as an alternative to Nvidia
Following the example of Google’s TPUs and Amazon’s Trainium, Microsoft aims to build its own chip ecosystem. Maia is currently focusing on inference and reducing the costs of running models. It does not yet have an AI model that competes with the strongest LLMs, but Microsoft AI (MAI) would like to achieve this eventually. That team was established in March 2024. In a conversation with Wall, we already suggested that those insights might help improve Maia, but it now appears that advice can also come from another angle.
Anthropic expects to be able to provide input into the design process for the next generation of Maia chips. Meanwhile, Anthropic’s infrastructure choices are more diverse: the company signed a multi-year contract with CoreWeave for Nvidia hardware and also committed 200 billion to Google Cloud. It also runs on AWS Trainium chips.