AI startup Anthropic is reportedly in talks with Microsoft to rent servers powered by the tech giant's in-house Maia AI chips for handling complex workloads. According to a report by The Information, the discussions are still at an early stage and may not necessarily result in a final agreement. The partnership would allow Anthropic to secure additional computing power for its Claude models using Microsoft's custom-built AI hardware. This development aligns with Microsoft's broader strategy alongside Google and Amazon to reduce dependence on Nvidia's semiconductors in the AI infrastructure stack, as companies increasingly prioritize inference capabilities over model training. The discussions come as Microsoft continues to face delays in getting its custom Maia chips into the market, making landing Anthropic as a customer particularly significant for the Azure business.
Microsoft shares rose 2% Thursday following the report of discussions with Anthropic, demonstrating strong investor confidence in the potential partnership. The stock movement reflects market optimism about Microsoft's ability to compete with rivals such as Alphabet and Amazon in the AI infrastructure space by offering an alternative to Nvidia's dominant AI hardware. The deal would allow Microsoft to compete with cloud rivals Google and Amazon by offering an alternative to Nvidia's hardware, which currently dominates the market and eats into cloud profit margins. For Anthropic, renting Maia chips provides additional options to run its models and the opportunity to tailor future generations of the chip to its specific needs.
The potential partnership would involve Microsoft's custom-built Maia 200 chip, launched in January 2026, which is built using Taiwan Semiconductor Manufacturing Co.'s (TSMC) 3-nanometer process technology. Each tray connects four Maia accelerators with direct, non-switched links, keeping high-bandwidth communication within the system to improve inference efficiency. The chips rely on Ethernet connections, a different networking approach from Nvidia's InfiniBand-based systems. Microsoft CEO Satya Nadella stated that the Maia 200 delivers 30% better performance per dollar compared with current systems, though the chips are designed to run existing AI models faster than Nvidia's hardware rather than train new models. If Microsoft successfully lands Anthropic as a customer, it could open the door for other companies seeking alternative options for meeting the sky-high demand for AI inference, as agentic models continue to consume trillions of tokens.
The discussions highlight Anthropic's urgent need for compute capacity, as the company has become 'a victim of its own success' in rolling out products that can be quickly adopted by enterprise clients. This rapid, wide-scale adoption has revealed significant compute constraints, forcing Anthropic to seek compute capacity from multiple providers including CoreWeave, Amazon, Google, Broadcom, and even xAI. Amazon and Google have both experienced hot demand for their custom inference chips, demonstrating the broader market shift toward specialized AI hardware. For Anthropic, the move could provide greater flexibility in running its AI systems and help shape future chip designs suited to its requirements, potentially reducing its dependence on multiple providers and improving operational efficiency.
