Anthropic May Use Microsoft AI Chips Over Nvidia
Anthropic is reportedly considering Microsoft’s Maia AI chips for Claude workloads, signalling a broader shift in the AI infrastructure race.
The AI chip race is no longer centred around a single company. Reports suggest that Anthropic is exploring the use of Microsoft’s in-house AI chips to support parts of its Claude AI models, highlighting a broader shift in how major AI firms are sourcing computing power.
According to reports cited by The Hindu, Anthropic is in discussions to access Azure cloud servers powered by Microsoft’s Maia AI accelerators. The arrangement would reportedly complement the company’s existing Nvidia-based infrastructure rather than replace it.
Why is computing becoming the AI industry’s biggest asset
Access to computing infrastructure has become one of the most critical competitive advantages in artificial intelligence. Training and running large AI systems requires enormous amounts of specialised hardware capable of processing data at high speed and scale.
For years, NVIDIA has dominated that space through its GPUs, which power many of the world’s leading AI models. But growing demand, rising costs, and supply limitations are now pushing AI companies to diversify their infrastructure strategies.
Instead of buying physical hardware directly, Anthropic would reportedly rent access to Maia-powered systems through Microsoft Azure. The focus is expected to be on inference workloads, the stage where trained AI models respond to user prompts in real time.
Inference has become increasingly important as generative AI tools scale to millions of users. Unlike training, which happens periodically, inference runs constantly and requires reliable, cost-efficient throughput.
What Microsoft’s Maia chips bring to the table
Microsoft introduced Maia as part of its broader push to develop custom AI infrastructure for Azure cloud services. Unlike general-purpose processors, Maia chips are designed specifically for AI workloads and integrated closely with Microsoft’s networking, cooling, and software systems.
The company aims to improve efficiency while lowering operational costs at cloud scale. Many modern AI accelerators, including Maia, also use high-bandwidth memory. This is a faster memory technology that enables processors to move large volumes of data more efficiently than traditional server memory.
That matters because AI models continuously process massive datasets while generating responses. Faster memory access can improve overall system performance and reduce delays during inference tasks.
While Nvidia’s GPUs remain widely regarded as the industry benchmark for advanced AI training, alternative chips are becoming more attractive for production-level workloads where scalability and efficiency matter heavily.
The race to break Nvidia’s AI hardware grip
Anthropic’s reported discussions with Microsoft reflect a wider industry trend. Major cloud providers, including Microsoft, Google, and Amazon, are investing heavily in custom silicon to reduce dependence on Nvidia and gain greater control over infrastructure economics.
At the same time, AI developers increasingly want flexibility across multiple hardware platforms rather than relying entirely on one supplier. For Anthropic, adding Maia-backed infrastructure could improve service reliability, create more predictable operating costs, and reduce exposure to supply constraints in the AI hardware market.
Microsoft’s AI chief says AI could automate most desk jobs within 18 months
What to watch next
The biggest questions now are whether the talks evolve into a formal agreement and how Anthropic’s workloads perform on Maia compared to Nvidia-powered systems in real-world deployment. Software support will also be critical, including optimisation tools, monitoring systems, and compatibility across AI frameworks.
Even if Anthropic expands onto Microsoft’s custom chips, Nvidia is expected to remain central to advanced AI training for the foreseeable future. But the discussions highlight a growing reality in AI: companies increasingly want infrastructure flexibility, not dependence on a single hardware ecosystem.