Anthropic is reportedly in early discussions to rent server capacity powered by Microsoft’s Maia chips, according to recent reports from The Information and CNBC. The talks come shortly after Anthropic agreed to use Google’s cloud infrastructure and Tensor Processing Units under a large multi-year agreement. They also follow a separate Microsoft agreement announced in November. Under that deal, Microsoft said it would invest US$5 billion in Anthropic, while Anthropic committed to spending US$30 billion on Azure.
Why is Anthropic talking to Microsoft about Maia chips?
Anthropic needs more compute capacity and has had difficulties with compute, so it is exploring Microsoft’s custom Maia chips. The company’s CEO, Dario Amodei, said earlier this month that the company has had “difficulties with compute.” This statement directly explains why Anthropic is actively seeking additional infrastructure from multiple cloud providers.
The reported discussions remain at an early stage. CNBC said Anthropic has not closed a deal with Microsoft over the use of Maia. However, the fact that talks are happening at all signals that Anthropic is serious about diversifying its hardware options beyond Nvidia GPUs, which it has historically relied on heavily to train and run generative AI models.
Anthropic’s compute needs are enormous. The company is best known for Claude, its family of AI models. Training and running models like Claude requires vast amounts of processing power, and the company has been securing capacity from multiple vendors to ensure it has enough.
What other cloud providers does Anthropic use?
Anthropic also uses Google Cloud with TPUs and Amazon Web Services with Trainium chips. The company has infrastructure and investment relationships with Amazon, Google, and Microsoft. Amazon has invested billions of dollars in the company, while Google and Microsoft also have financial and cloud relationships with the AI company.
Anthropic’s reported US$200 billion infrastructure agreement with Google gives Anthropic access to Google Cloud systems and Google’s custom Tensor Processing Units over five years. Anthropic also said in April that it had signed an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity. That capacity is expected to come online starting in 2027.
Google has long used TPUs for AI workloads across its own services and cloud customers. Through Google Cloud, those chips are also available to external customers. This gives Anthropic a substantial amount of dedicated compute power for training and inference.
In addition, Anthropic has expanded its use of Amazon Web Services infrastructure. In April, Anthropic said it would use AWS’s custom Trainium chips under a 10-year arrangement worth more than US$100 billion.
What is Microsoft’s Maia 200 chip?
Maia 200 is Microsoft’s second-generation AI accelerator designed for inference, offering 30% better performance per dollar. Microsoft has been developing Maia as part of its wider custom silicon programme. Chief executive Satya Nadella said on Microsoft’s April earnings call that Maia 200 offers more than 30% better tokens per dollar compared with the latest silicon in Microsoft’s fleet.
Microsoft said Maia 200 is already running in production in its US Central region near Des Moines, Iowa. Deployment in US West 3, near Phoenix, Arizona, is expected to follow. This means the chip is not just a theoretical design — it is actually deployed and serving real workloads.
The Maia 200 is designed specifically for AI inference, which is the process of running a trained model to make predictions or generate responses. This is different from training, which requires different hardware characteristics. For Anthropic, which runs Claude for millions of users, inference capacity is just as critical as training capacity.
How much compute capacity has Anthropic secured from AWS?
Anthropic secured up to 5GW of new capacity under a 10-year, US$100 billion agreement with AWS. The commitment covers Graviton and Trainium chips, including future generations of Amazon’s custom silicon. This is a massive amount of compute power, measured in gigawatts rather than the more typical server-level metrics.
To put 5GW in perspective, a typical large data center might consume 50-100 megawatts. Anthropic’s commitment represents the equivalent of 50 to 100 large data centers worth of power, all dedicated to training and running Claude. This scale underscores how compute-intensive modern AI models are.
The agreement with AWS complements Anthropic’s deals with Google and Microsoft. By securing capacity from all three major cloud providers, Anthropic is hedging against any single vendor’s supply constraints or pricing changes. This multi-cloud strategy also gives the company flexibility to optimize costs across different hardware platforms.
What did Anthropic’s CEO say about compute difficulties?
CEO Dario Amodei said the company has had “difficulties with compute,” which explains its multiple cloud deals. This statement was made earlier this month and provides direct insight into why Anthropic is pursuing such an aggressive multi-cloud strategy.
Compute difficulties can take many forms in the AI industry. They can mean not having enough GPUs or accelerators to train models quickly. They can mean high costs that eat into margins. They can mean long wait times for hardware allocations from cloud providers. For a company like Anthropic, which is racing to improve its models and deploy them at scale, any delay in compute availability can be costly.
Anthropic has historically relied heavily on Nvidia GPUs to train and run generative AI models. Nvidia remains a major supplier of advanced AI processors. However, demand for Nvidia’s hardware has outstripped supply for years, creating a bottleneck for AI companies. By diversifying to custom chips from Microsoft, Google, and Amazon, Anthropic is trying to bypass that bottleneck.
How do custom chips compare to Nvidia GPUs?
Google, Amazon Web Services, and Microsoft have each developed custom AI chips for cloud infrastructure. Google offers its TPU platform through Google Cloud. Amazon Web Services has Trainium and Inferentia, while Microsoft has Maia. These chips are designed specifically for AI workloads, whereas Nvidia GPUs were originally designed for graphics rendering.
Custom chips can offer better performance per dollar for specific tasks. Microsoft claims Maia 200 delivers 30% better performance per dollar than the latest-generation hardware in its fleet. This kind of efficiency gain is significant when operating at the scale that Anthropic requires.
However, custom chips also have limitations. They are less flexible than general-purpose GPUs. They may not support all the same software frameworks and libraries. Companies that use them often need to invest in custom software optimizations to get the best performance. For Anthropic, which already works with multiple hardware platforms, this may be a manageable trade-off.
What does this mean for the cloud computing market?
The cloud computing market is seeing a shift toward custom silicon. AWS, Google, and Microsoft are all developing their own chips rather than relying solely on Nvidia. This trend gives cloud providers more control over their supply chains and margins.
You may also enjoy reading: Keychron M5 Review: 5 Clever Details in Unusual Gaming Mouse.
For customers like Anthropic, having multiple hardware options creates competition. Cloud providers must offer competitive pricing and performance to win business. This can lead to better deals for customers over time.
OpenAI has also been reported to be working with Broadcom on custom AI chips. This suggests that the entire AI industry is moving toward custom hardware solutions. The era of relying solely on Nvidia for AI compute may be coming to an end.
What are the risks of relying on multiple cloud providers?
Managing infrastructure across multiple clouds adds complexity. Anthropic must maintain relationships with three different cloud providers, each with its own pricing models, APIs, and support structures. This requires significant engineering and operational overhead.
There is also the risk of vendor lock-in at the chip level. If Anthropic optimizes its models for Maia, TPU, and Trainium architectures, it may become dependent on those specific hardware platforms. Switching to a different chip later could require substantial rework.
On the other hand, having multiple providers reduces the risk of a single point of failure. If one provider experiences an outage or capacity shortage, Anthropic can shift workloads to another. This redundancy is valuable for a company that needs to keep Claude running reliably for its users.
What is the timeline for a potential Maia deal?
The talks between Anthropic and Microsoft remain at an early stage. No deal has been closed yet. The servers under discussion would use Maia, Microsoft’s in-house AI accelerator. However, the processor is not yet available through Azure in a general sense.
Microsoft said Maia 200 is already running in production in its US Central region near Des Moines, Iowa. Deployment in US West 3, near Phoenix, Arizona, is expected to follow. This means the hardware exists and is operational, but it may not be widely available yet.
A deal could take months or even years to finalize. Anthropic would need to negotiate pricing, capacity commitments, and technical integration details. The company would also need to ensure that its software stack works well with Maia chips.
How does this affect Nvidia?
Nvidia remains a major supplier of advanced AI processors. The company’s GPUs are still the most widely used hardware for AI training and inference. However, the rise of custom chips from cloud providers is a long-term threat to Nvidia’s dominance.
If Anthropic and other AI companies shift significant workloads to custom chips, Nvidia could lose market share. However, this shift will take time. Custom chips require software optimization and ecosystem development that Nvidia has already invested heavily in.
For now, Nvidia’s position remains strong. The company continues to release new generations of GPUs with improved performance. But the trend toward diversification is clear, and Nvidia will need to compete on price and performance to retain its customers.
Frequently Asked Questions
How will Anthropic benefit from using Microsoft’s Maia chips?
Anthropic could gain access to additional compute capacity at a lower cost per token. Microsoft claims Maia 200 delivers 30% better performance per dollar than other hardware in its fleet. This could help Anthropic reduce its infrastructure costs while running Claude at scale.
What is the difference between Maia chips and Nvidia GPUs for AI workloads?
Maia chips are custom-designed specifically for AI inference, while Nvidia GPUs are general-purpose processors originally built for graphics. Maia may offer better efficiency for inference tasks, but Nvidia GPUs are more flexible and support a wider range of software frameworks. Maia chips are less widely available and require more custom integration.
Is it risky for Anthropic to rely on multiple cloud providers for compute?
Yes, managing infrastructure across multiple clouds adds complexity and operational overhead. However, it also reduces the risk of a single point of failure. If one provider experiences an outage or capacity shortage, Anthropic can shift workloads to another. This diversification is a common strategy for large-scale AI companies.
