SNDC

TradingKey - As Claude developer Anthropic officially submitted its S-1 filing for an initial public offering to the U.S. Securities and Exchange Commission, its computing power supply chain, long hidden behind cloud service contracts, was systematically disclosed for the first time. Anthropic is not only a major AI model company, but has also become one of the world's most important AI chip buyers.

Based on estimates from public agreements and related disclosures, Anthropic has locked in approximately 11 GW of computing power. Among this, contracted capacity based on Nvidia GPUs is about 4.7 GW, accounting for roughly 42%; Google TPUs account for about 4.5 GW, nearly matching Nvidia in scale; Amazon Trainium accounts for about 1.2 GW; and AMD GPUs account for about 0.8 GW. However, the above figures are all third-party estimates and do not represent the actual computing power that Anthropic has taken delivery of, put into production, or is currently running.

Because some contracts adopt different metrics such as "maximum capacity," phased delivery, or capacity not yet operational, the sum of individual figures is slightly higher than 11 GW. Therefore, it is more appropriately viewed as an approximate supply structure rather than actual capacity online simultaneously.

Anthropic confirmed in its prospectus that Claude can run natively on four types of chips—AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs—and allocates different workloads based on cost, performance, and supply conditions through its proprietary scheduling system.

Nvidia Remains in Top Spot, But No Longer Holds Absolute Monopoly

Nvidia (NVDA) remains Anthropic's largest chip supplier. Of the approximately 4.7 GW of Nvidia computing power, a significant portion comes from the SpaceX AI partnership, with the rest distributed across cloud computing platforms such as Microsoft Azure, Nscale, and Lambda.

Nvidia's advantages stem primarily from its mature CUDA software ecosystem, GPU cluster scalability, and developer base. For Anthropic, which requires simultaneous large-model training and intensive inference, Nvidia's platform remains the easiest choice to deploy at scale.

However, the 42% share also indicates that Anthropic is not betting its computing needs entirely on a single supplier. Unlike some AI cloud computing companies that rely heavily on Nvidia GPUs, Anthropic is mitigating risks associated with supply constraints, price increases, and delivery delays through a multi-chip architecture.

This strategy has a dual impact on Nvidia. On the one hand, Anthropic will remain an important customer, with expanding demand for AI model training and inference expected to continue driving GPU procurement. On the other hand, the growing share of Google TPUs, AWS Trainium, and AMD GPUs means Nvidia faces rising substitution pressure within major AI labs.

Google TPU Scale Nears Nvidia as Custom Chips Emerge as Biggest Winner

Google parent company Alphabet (GOOGL) may be the most obvious beneficiary of Anthropic's compute diversification strategy. The Google TPU compute capacity locked in by Anthropic is approximately 4.5 GW, trailing Nvidia GPUs by only about 0.2 GW.

This means that TPUs are no longer just internal proprietary hardware for Google to train its Gemini models, but are gradually becoming a commercialized compute platform available to major external clients. Anthropic's adoption also provides crucial validation of TPUs' capability to handle frontier model training and inference workloads.

As Google TPUs are co-designed with Broadcom, this demand benefits not only Google Cloud but could also generate additional revenue for Broadcom from custom AI chips and networking products. Anthropic previously reached long-term partnerships with Google and Broadcom, with the relevant capacity expected to come online gradually starting in 2027.

Amazon (AMZN) has likewise entered Anthropic's core supply system through AWS Trainium. Although its current estimated scale of approximately 1.2 GW is lower than those of Nvidia and Google, Amazon is both a major investor in Anthropic and a key cloud service partner. If Trainium can demonstrate a cost advantage in Claude workloads, it will help AWS reduce its reliance on external GPUs and improve margins for AI cloud services.

AMD Makes Breakthrough as Anthropic Gains Stronger Bargaining Power

AMD (AMD) currently accounts for a relatively low share of Anthropic's compute portfolio, with an identified scale of approximately 0.8GW, but its strategic significance cannot be ignored. AMD and Anthropic previously announced a collaboration to deploy up to 2GW of Instinct MI450 series GPUs, with initial capacity expected to go live starting in the first half of 2027.

If AMD can successfully scale up the deployment, it will prove that the ROCm software ecosystem and Helios rack-level systems possess the capability to support frontier models. This will not only help AMD win more AI lab and cloud provider customers, but could also break Nvidia's high concentration in the general-purpose GPU market.

For Anthropic, introducing AMD may serve an even more important role in boosting its procurement bargaining power. When the same model can be scheduled across four chip architectures, the company can select the most economical hardware based on price, supply, and power consumption, reducing the risk of being locked into a single chip vendor.

Anthropic describes this capability as a "fungible unified compute pool." The value of this architecture lies not only in ensuring access to sufficient chips, but also in increasing the model capability and revenue generated for every dollar spent on compute.

Compute Diversification Still Fails to Relieve Anthropic's Financial Pressure

While a multi-chip layout mitigates supply chain risks, it fails to address the issue of overinvestment in computing power. Anthropic's S-1 filing shows that the company has signed numerous long-term cloud computing, equipment leasing, and data center contracts, a significant portion of which constitute fixed or minimum payment obligations that may still require payment even if actual usage falls short of expectations.

Relevant filings indicate that Anthropic's future minimum commitments for computing power and infrastructure payments reach hundreds of billions of dollars, with additional agreements to be settled based on actual delivered capacity. The company also warned that continuing to expand its computing infrastructure will consume substantial operating and investing cash flows for the foreseeable future.

Therefore, Anthropic's 11 GW computing power reserve is both a competitive advantage and a key financial risk for IPO investors to weigh. If Claude's revenue, customer usage, and inference demand continue to grow, this pre-locked computing capacity will form a high barrier to entry. However, if AI demand slows or chip efficiency improves rapidly, these long-term contracts could turn into a heavy fixed cost burden.

Find out more