Anthropic, the developer of the Claude chatbot, uses four chip families: Amazon Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs. This strategy allows the company to distribute computing workloads across different platforms and could, over time, reduce major AI developers' dependence on Nvidia—and, with it, the chipmaker's ability to dictate accelerator prices.

Anthropic detailed the capabilities of its computing infrastructure in IPO registration documents filed on Monday. The company describes its computing resources as a unified pool of hardware from several strategic partners. Specialized software routes workloads to the chips that can handle them most efficiently.

Analysts say this approach could give Anthropic greater flexibility in choosing suppliers and strengthen its position in price negotiations. However, moving away from Nvidia quickly will not be easy: switching between hardware platforms requires substantial engineering resources, and the company still spends billions of dollars on Nvidia computing capacity, Business Insider reports.

Four Platforms Instead of Dependence on a Single Supplier

In its IPO documents, Anthropic emphasizes that building a multi-chip infrastructure took several years of work and substantial investment.

The company uses Nvidia accelerators, Amazon's Trainium chips, Google's TPUs, and AMD hardware. This allows it to select the right platform based on the nature of the computing workload and the availability of resources.

For Nvidia, the potential problem is not just that customers may buy fewer of its chips. The shift in their bargaining power is far more important.

“The main risk for Nvidia may be a reduction in its pricing power. A customer that can distribute workloads across four platforms negotiates very differently from one that cannot,” said Patrick Moorhead, founder and CEO of Moor Insights & Strategy.

A multi-chip strategy also reduces exposure to supply fluctuations. According to IDC analyst Ashish Nadkarni, using several types of accelerators makes Anthropic less vulnerable to disruptions in Nvidia chip supplies.

Why Replacing Nvidia Is Harder Than It Looks

Despite the potential benefits, diversification does not mean Anthropic can seamlessly move all its computing workloads to alternative platforms.

Switching between different types of chips is particularly challenging when training AI models, Moorhead notes. It takes more than simply installing different hardware: software must be adapted and computing processes optimized for the characteristics of each platform.

Such capabilities are primarily available to the largest AI companies, which have substantial financial and engineering resources. For smaller developers, the cost of supporting multiple hardware platforms may be difficult to justify.

Gil Luria, head of technology research at D.A. Davidson, believes other major players, including OpenAI, may also use several types of accelerators. For smaller companies, however, the costs and complexity of such a strategy often outweigh the benefits.

Moreover, Anthropic continues to spend billions of dollars on Nvidia computing capacity. The existence of alternatives does not, in itself, eliminate Nvidia's competitive advantages in hardware performance, networking technology, and its software ecosystem, stressed Brad Gastwirth, head of research and market intelligence at Circular Technology.

In the near term, therefore, a multi-chip strategy is more likely to strengthen the bargaining position of Nvidia's largest customers than to threaten its dominance.

The Next Threat: Computing Capacity Becoming a Commodity

Nvidia also faces a broader risk: as AI infrastructure expands, computing capacity could become less scarce. If the supply of accelerators and servers begins to exceed demand, the cost of renting computing resources could fall sharply.

Citadel CEO Ken Griffin has pointed to this possibility. He compared the GPU market to the airline industry, where the marginal cost of serving an additional customer is relatively low compared with the cost of building the infrastructure.

“GPUs, like airlines, have very low marginal operating costs. Once excess capacity emerges, prices fall rapidly,” Griffin said in an interview with Bloomberg.

For Nvidia, this would mean pressure not only from competing chipmakers but also from customers with more choices and access to cheaper computing resources.

For now, Anthropic is not abandoning Nvidia accelerators, and its multi-chip infrastructure remains the product of years of investment. But the IPO shows that, for the largest AI companies, the ability to operate across different hardware platforms is becoming a competitive advantage in its own right. If this approach becomes widespread, Nvidia could retain its performance lead while gradually losing some of its influence over pricing and business terms.