Anthropic in early talks start using Microsoft’s Maia 200 AI chips

Anthropic is in advanced discussions to rent cloud servers powered by Microsoft’s custom-designed AI chips. The high-stakes negotiations arrive just months after Microsoft committed a blockbuster $5 billion investment to Anthropic, signaling a profound realignment in the tech industry as major AI developers aggressively secure alternative computing power to sustain the global expansion of generative AI.

Navigating the global compute crunch with custom silicon

The potential hardware deal specifically focuses on Microsoft’s newly unveiled, second-generation "Maia 200" AI accelerator. Built on Taiwan Semiconductor Manufacturing Co.’s (TSMC) elite 3-nanometer process technology, the Maia 200 chip was intentionally engineered by Microsoft to handle massive large language model (LLM) workloads natively within its Azure data centers.

If finalized, Anthropic will deploy the Maia 200 chips primarily to power the computationally intensive process of generating instant text and code answers for millions of users interacting with Claude.

The move is born out of raw necessity. Anthropic CEO Dario Amodei recently acknowledged that the startup is facing acute "difficulties with compute" due to exponential user demand. While Nvidia’s top-tier graphics processing units (GPUs) remain the undisputed gold standard for initial model training, Microsoft’s custom silicon offers specialized architectural advantages for running live AI systems.

The Maia 200 is heavily loaded with high-speed static random-access memory (SRAM), allowing it to process massive volumes of concurrent user prompts with extreme efficiency. Microsoft claims the custom chip delivers a 30% improvement in "tokens per dollar" compared to generic hardware fleets, offering Anthropic a vital weapon to slash its astronomical operating costs.

A multilateral infrastructure play

For Anthropic, the Microsoft chip discussions represent the latest leg of an aggressive multi-cloud diversification strategy designed to vacuum up computing capacity from every available geopolitical and corporate avenue. Over the past year, the Claude creator has built an incredibly complex infrastructure web, which now includes:

• Amazon Web Services: A 10-year, $100 billion infrastructure agreement utilizing AWS’s custom Trainium processors.

• Google Cloud: A multi-billion-dollar partnership leveraging Google's custom Tensor Processing Units (TPUs) alongside Broadcom silicon.

• SpaceX: A massive $1.25 billion monthly lease to consume the entire 300-megawatt capacity of Elon Musk's Colossus 1 data center.

By adding Microsoft's Maia 200 to its operational arsenal, Anthropic protects its software ecosystem against sudden supply chain bottlenecks or individual cloud provider outages.

Flipping the power dynamic with OpenAI

For Microsoft, securing Anthropic as a flagship customer for its custom silicon is an immense strategic and symbolic triumph. Historically, Microsoft’s cloud infrastructure strategy was tied almost exclusively to its close ally, OpenAI. However, as that partnership has loosened and OpenAI explores building its own independent hardware pipelines, Microsoft has aggressively diversified its Azure portfolio.

By investing $5 billion to integrate Anthropic’s Claude models directly into its corporate Copilot ecosystem, Microsoft has effectively reduced its reliance on ChatGPT. Successfully migrating a tier-one AI laboratory onto its internal Maia silicon would finally validate Microsoft's multi-billion-dollar investments in proprietary semiconductors, positioning the tech giant alongside Google and Amazon as a self-sufficient powerhouse capable of renting out proprietary, end-to-end AI infrastructure at a global scale.