Anthropic is in early-stage talks with Microsoft to use the software giant's custom Maia artificial intelligence (AI) chips to run inference for its Claude family of models, according to reports from The Information confirmed by multiple outlets on May 21, 2026. <cite index="1-1">Anthropic is in talks with Microsoft to adopt its Maia artificial intelligence chip, but the two companies have not signed a deal, a person familiar with the matter said.</cite>
Neither company has commented publicly. <cite index="1-18,1-19">Anthropic declined to comment. Microsoft did not immediately respond to a request for comment.</cite> Microsoft shares moved modestly on the news, with <cite index="6-21">shares of Microsoft climbing approximately 2% following reports that Anthropic is considering leasing servers equipped with Microsoft's proprietary Maia AI processors.</cite>
The Maia 200 chip
<cite index="7-18">Microsoft introduced Maia 200 on January 26 as an inference accelerator for Azure workloads with 216GB HBM3e memory and a 30% better performance-per-dollar claim versus the latest generation hardware already in its fleet.</cite> On Microsoft's April earnings call, <cite index="1-20,1-21">CEO Satya Nadella said Maia 200 "offers over 30% improved tokens per dollar, compared to the latest silicon in our fleet." He said the chips are now running in Microsoft data centers in Arizona and Iowa.</cite> <cite index="1-3">Microsoft announced the processor in January, but has not made it available to Azure customers.</cite>
The chip is fabricated on an advanced process node but is purpose-built for serving models rather than building them. <cite index="6-35,6-36">Microsoft introduced its second-generation Maia processor in January—designated Maia 200. Manufactured by TSMC utilizing 3-nanometer fabrication technology, it incorporates high-bandwidth memory, albeit an earlier iteration than what Nvidia plans to integrate into its forthcoming Vera Rubin processors.</cite> <cite index="6-39">The crucial limitation: Maia lacks the architecture required for training new AI models, functioning exclusively for inference operations.</cite>
Anthropic's compute diversification
The talks come as Anthropic seeks to expand its supply of accelerators amid surging usage of Claude. <cite index="1-14">Anthropic has had "difficulties with compute," Dario Amodei, the company's co-founder and CEO, said at an event earlier this month.</cite> <cite index="7-14">Anthropic's first-quarter demand reached an 80-fold annualized growth pace after the company had planned for only a 10-fold increase.</cite>
Anthropic has progressively diversified its compute beyond Nvidia graphics processing units (GPUs). <cite index="1-16,1-17">In April, Anthropic said it would use Amazon Web Services' custom Trainium chips in a 10-year arrangement worth more than $100 billion. Anthropic announced plans to use Google's tensor processing unit chips in October.</cite>
Microsoft and Anthropic deepened their commercial relationship late last year. <cite index="1-12,1-13">In November, Microsoft said it would invest $5 billion in Anthropic, while Anthropic committed to spending $30 billion on Azure. Anthropic also relies on cloud services from Amazon and Google.</cite>
Strategic stakes
For Microsoft, landing Anthropic would be the first major external deployment of Maia. <cite index="1-7">A deal would represent a win for Microsoft, which is behind cloud rivals Amazon and Google when it comes to supplying clients with special-purpose AI silicon.</cite> <cite index="9-4">Securing Anthropic as a client would mark a significant win for Microsoft, whose in-house chip development faced delays last year.</cite>
For Anthropic, <cite index="9-1">renting Maia chips provides additional options to run its models and the opportunity to tailor future generations of the chip to its specific needs.</cite> The talks remain preliminary, and there is no guarantee they will produce a binding agreement.