Anthropic in Talks to Use Microsoft Maia 200 Chips for Claude

Anthropic is in talks to adopt Microsoft's Maia 200 custom AI chip to power its Claude models, CNBC reported Thursday, citing a person familiar with the matter. No deal has been signed, but the discussions — first reported by The Information — underscore how aggressively Anthropic is diversifying its silicon supply chain.

The Maia 200, unveiled by Microsoft in January 2026, is designed specifically for AI inference — running existing models faster and cheaper, not training new ones from scratch. Microsoft CEO Satya Nadella told investors in April that the chip delivers "over 30% improved tokens per dollar, compared to the latest silicon in our fleet," CNBC reported. The chips are currently running in Microsoft data centers in Arizona and Iowa but aren't yet available to external Azure customers.

The Maia talks are the latest in a rapid‑fire series of chip and compute deals for Anthropic. In April, the company signed a 10‑year, $100 billion‑plus deal to use Amazon Web Services' custom Trainium chips. In October 2025, it announced plans to use Google's tensor processing units (TPUs). And on Wednesday, SpaceX's IPO filing revealed Anthropic is paying $1.25 billion per month for data center access through 2029. It still relies on NVIDIA GPUs as its historical backbone for training.

That's five separate compute vendors — NVIDIA, AWS Trainium, Google TPUs, SpaceX/xAI compute, and potentially Microsoft Maia. No other AI lab comes close to this level of multi‑vendor diversification.

The answer is straightforward: Anthropic can't get enough compute from any single source. CEO Dario Amodei publicly acknowledged "difficulties with compute" in early May, CNBC reported, as demand for Claude and the Claude Code developer tool outstripped available capacity.

Anthropic's revenue is surging — projected to hit at least $10.9 billion this quarter, per fundraising materials reported by Reuters — and every dollar of that growth requires more inference capacity. The company is approaching its first profitable quarter, but profitability depends on keeping inference costs down. That's where chips like Maia 200, designed to improve tokens‑per‑dollar economics, become critical.

For Microsoft, landing Anthropic as a Maia customer would be a major competitive win. The company has trailed Amazon (Trainium) and Google (TPUs) in offering custom AI silicon to cloud clients, Quartz noted. Microsoft's custom chip development faced delays last year, and the Maia 200 is still not available externally — a deal with Anthropic would be its first marquee external silicon customer.

The relationship between the two companies is already deep. Microsoft invested $5 billion in Anthropic in November 2025, and Anthropic committed to spending $30 billion on Azure cloud services. But Anthropic also uses Amazon and Google cloud infrastructure, maintaining a deliberate multi‑cloud strategy.

Microsoft shares rose about 2% on the news, Investing.com reported, reflecting investor optimism that the Maia chip could finally compete in the AI silicon market.

One important detail: the Maia 200 is designed to run existing models, not to train new ones, The Information reported. That means even if Anthropic signs a deal, it would still need NVIDIA GPUs or other training‑capable hardware for developing future Claude models.

For inference workloads — the day‑to‑day serving of Claude to millions of users and developers — Maia 200's cost advantage could be significant. Nadella's claim of 30% better tokens per dollar suggests that running Claude inference on Maia would be meaningfully cheaper than on NVIDIA hardware. At Anthropic's scale, the improved tokens‑per‑dollar efficiency could translate to billions of dollars in annual savings.

Anthropic's silicon strategy is a blueprint for where the industry is heading. The days of single‑vendor NVIDIA dependence are giving way to a landscape where AI labs mix and match chips based on workload, cost, and availability. Inference goes to Maia or Trainium. Training stays on NVIDIA or TPUs. Overflow capacity gets rented from whoever has it — even competitors like xAI.

For builders, this diversification is good news. More chip vendors competing for AI workloads means downward pressure on inference costs, which eventually shows up in lower API prices. The Maia 200 talks are one more data point suggesting the NVIDIA monopoly on AI silicon is cracking — slowly, but in enough places to matter.