Anthropic tries Microsoft's Maia chips: OpenAI's main competitor diversifies hardware

According to reports from The Information and CXO Digitalpulse, Anthropic is in early-stage talks with Microsoft about using its own AI accelerators from the Maia family. Specifically, these are the Maia 200 chips, the second generation of custom silicon that Microsoft is developing for its Azure cloud platform.

At first glance, this is a surprising move. Anthropic is one of Amazon's largest partners — the company has invested 8 billion dollars in it — and has also received a 2 billion dollar investment from Google. Both companies run their own cloud infrastructure and are also developing their own AI chips (AWS Trainium and Google TPU). Yet Anthropic is turning to Microsoft, whose Azure is a direct competitor to both AWS and Google Cloud.

The reason is pragmatic: hunger for computing power. Claude, like competing ChatGPT or Gemini, requires enormous amounts of computational resources for both training and operation (inference). Demand for Nvidia chips — the current king of AI hardware — significantly outstrips supply and prices are soaring to astronomical heights. Each Nvidia H100 costs 30–40 thousand dollars and delivery times stretch to months.

What are Microsoft Maia chips and why they matter

Microsoft introduced the first generation of its own AI chip — Maia 100 — in November 2023. It was historically the first custom silicon the company designed for its Azure cloud infrastructure. The chip was manufactured using a 5nm process in TSMC factories and contained 105 billion transistors. Microsoft collaborated with OpenAI from the start, which tested it on its models including GPT-3.5 Turbo.

The second generation — Maia 200 — was officially announced in 2025. According to The Verge, the Maia 200 is optimized primarily for inference — i.e., running already-trained models — where it is faster and more energy-efficient. In training new models, however, it still lags behind Nvidia. This makes sense: running models (when Claude responds to users) makes up the vast majority of computational load, while training happens periodically.

For Microsoft, a potential collaboration with Anthropic is strategically crucial. The company has invested billions of dollars in AI chips and needs someone other than just OpenAI to use them — otherwise, its investment would be hard to recoup. Anthropic is the ideal reference customer in this regard: it has top-tier models, a huge user base, and its workload would provide Microsoft with valuable data for further chip development.

The end of Nvidia's monopoly? A more diverse AI chip market is on the horizon

Nvidia today controls roughly 80–90% of the AI accelerator market. Its GPUs like the H100 and newer B200 are the de facto standard for training large language models. This monopoly, however, creates vulnerability: anyone without access to Nvidia cannot build cutting-edge AI.

Anthropic is far from the only one looking for alternatives:

• Google has been developing its own TPU (Tensor Processing Unit) since 2015 — currently in its sixth generation. It uses them for Gemini and for Google Cloud customers.
• Amazon has invested in Trainium and Inferentia chips, which it offers within AWS as a cheaper alternative to Nvidia.
• Meta is developing its own MTIA (Meta Training and Inference Accelerator) for its Llama models.
• OpenAI is testing Microsoft's Maia chips and is reportedly also considering developing its own silicon.

The addition of Anthropic to the list of companies diversifying their hardware is a significant signal. If the Claude model were to successfully run on Maia chips, it would open the door for other AI companies that have so far relied exclusively on Nvidia.

For end users, this diversification could mean lower prices for AI services. Today, model operators pay astronomical sums for computing power, which are reflected in API access prices and subscription fees. Greater competition in the hardware market would push prices down — both for enterprise solutions and consumer products like Claude Pro ($20/month) or ChatGPT Plus.

Anthropic and its infrastructure hunger

Hardware diversification at Anthropic also makes sense when looking at its overall infrastructure strategy. The company recently signed a massive contract with SpaceX for computing capacity worth 15 billion dollars annually, under which it will use Elon Musk's Colossus supercomputer. At the same time, according to The Information, Anthropic is "steadily increasing its use of Azure."

In other words: Anthropic is not betting on a single card. It uses a combination of its own solutions, cloud platforms (AWS, Google Cloud, Azure), and external supercomputers. This multi-cloud and multi-hardware approach is fairly unique in the AI industry — most companies stick to a single platform.

What it means for Europe and the Czech Republic

For European and Czech companies, this trend is important for two reasons. First, Microsoft Azure has a strong presence in Europe — if Maia chips prove their effectiveness, they could also appear in European Azure data centers, including the North Europe (Ireland) and West Europe (Netherlands) regions. This would allow European companies to run AI models on modern hardware while complying with GDPR.

Second, it's about strategic independence. The European Union, through the European Chips Act, is investing billions of euros in building its own chip manufacturing capacity. If it turns out that quality AI accelerators can be made outside of Nvidia, it strengthens the argument for European investments in the semiconductor industry.

For the average user in the Czech Republic, nothing changes immediately — Claude remains available in Czech via the web interface and API, whether it runs on chips from Nvidia, Microsoft, or anyone else. What matters, however, is the long-term impact: a more diverse hardware market = lower costs = more accessible AI services for everyone.