Anthropic is reportedly in early talks to rent server space powered by Microsoft’s custom-designed AI processors. According to a report from The Information, the creator of the Claude chatbot is looking to secure a broader supply of computing infrastructure to meet the rising global demand for its enterprise AI services.
The report claims that by rent Microsoft's Maia-powered servers, Anthropic is looking to diversify its backend hardware and access to Microsoft's custom silicon will give the AI startup an alternative to run and train its Claude models without being vulnerable to the industry-wide shortages and high premium costs associated with Nvidia hardware.
Anthropic’s $200 billion deal with Google
The report arrives just weeks after Anthropic signed a $200 billion infrastructure agreement with Google to utilise its cloud systems and custom Tensor Processing Units (TPUs) over 5 years. Moreover, if the talks with Microsoft materialize, it will provide a major validation boost for Microsoft's independent, in-house chip manufacturing efforts.
Reducing dependence on Nvidia chips
This potential partnership highlights a broader race among tech giants to reduce their financial dependency on Nvidia, which currently dominates the market for advanced AI processors. Major cloud providers, including Google and Amazon, are spent billions of dollars designing proprietary silicon optimised specifically for large language models (LLMS).
For example, Google has heavily scaled its proprietary Tensor Processing Unit (TPU) ecosystem, and has opened doors for the other companies to rent their TPUs. Meanwhile, Amazon continues to develop its Trainium and Inferentia lines, and Microsoft entered the arena with its custom Maia AI accelerator chips. ChatGPT-maker OpenAI is also developing custom AI chips with Broadcom.
It is to be noted that Anthropic remains deeply aligned with Amazon that has invested billions of dollars into the AI startup. The report, however, highlights that notes that Anthropic's discussions with Microsoft executives are still in the early stages.