Anthropic eyes Microsoft Maia chips amid massive compute expansion

Anthropic is reportedly in early discussions to rent server capacity running on Microsoft’s in-house Maia AI accelerator chips. No deal has closed yet, according to CNBC, but the talks alone are worth examining in the context of everything else Anthropic has committed to over the past several months.

The numbers involved across Anthropic’s existing compute relationships are genuinely difficult to put in perspective. In April, the company signed a 10-year arrangement with Amazon Web Services covering Trainium and Graviton chips, valued at more than $100 billion and securing up to 5 gigawatts of capacity for training and running Claude. That same month, Anthropic also reached an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected to come online from 2027, as part of a broader $200 billion infrastructure relationship with Google covering five years of cloud access and custom chip usage.

A separate Microsoft agreement announced in November saw the company commit $5 billion in investment while Anthropic pledged $30 billion in Azure spending, alongside contracted compute capacity of up to 1 gigawatt.

A closed Maia deal would put Anthropic in the unusual position of running custom chips from every major US cloud provider at once. AWS supplies Trainium, Google brings its TPUs, and Microsoft would add Maia to the mix, all sitting alongside the Nvidia GPU infrastructure Anthropic already depends on heavily.

Microsoft’s Maia 200, its second-generation in-house AI chip, targets inference workloads specifically. Satya Nadella said on Microsoft’s April earnings call that the chip delivers more than 30% better performance per dollar compared with the latest hardware currently running in Microsoft’s fleet. The chip is already in production in the US Central region near Des Moines, with deployment near Phoenix expected to follow.

Anthropic CEO Dario Amodei acknowledged earlier this month that the company has faced difficulties with compute availability. That admission, paired with the scale of commitments it has made across multiple providers, reflects how acutely the AI industry is feeling the gap between model ambition and available hardware.

For a company running one of the most capable AI model families in production, spreading compute risk across multiple chip architectures and providers is less a strategic luxury and more an operational necessity at this point.