Anthropic arrived in the capital markets with US$ 518 billion in future cloud, compute and infrastructure commitments, according to the IPO prospectus seen by Reuters. The figure highlights an important shift in the economics of frontier AI: competing for the most advanced models is requiring infrastructure contracts on a scale normally associated with the largest technology companies.
For Anthropic, the challenge is particularly relevant because these commitments are advancing much faster than its current revenue. In 2025, the company generated nearly US$ 4.6 billion, recorded an operating loss of US$ 8.06 billion and spent US$ 7.33 billion on compute and infrastructure. The question the IPO now exposes is how far a company still built around AI models can sustain a cost structure similar to that of a hyperscaler.
Compute is no longer a cost that follows demand
The most important difference revealed by the prospectus lies less in the absolute size of the US$ 518 billion than in the nature of these commitments.
According to Reuters, Anthropic expects to spend at least US$ 111.1 billion with Google, US$ 110 billion with Amazon and US$ 31.4 billion with Microsoft under seven- to ten-year contracts. There are also approximately US$ 161.2 billion in lease obligations related to Broadcom. In many of these agreements, reducing consumption does not eliminate the payment.
This changes the traditional economics of a software company. A SaaS company typically tries to keep a significant portion of its costs variable. If demand slows, infrastructure and capacity acquisition can be adjusted.
Anthropic is making the opposite move. To secure chips, energy and data centers ahead of competitors, it is turning an increasing share of compute into a financial obligation contracted in advance.
The agreement announced with Amazon in April already showed this direction. Anthropic committed more than US$ 100 billion over ten years to secure up to 5 gigawatts of capacity in AWS technologies, including several generations of Trainium.
The logic is operationally defensible. If compute is scarce, waiting for demand to appear before contracting infrastructure can mean failing to serve it. The risk is that the company has to get it right years in advance on how much of that capacity it will be able to monetize.
The comparison with hyperscalers has an important limit
US$ 518 billion brings the scale of Anthropic's commitments closer to that normally associated with the largest technology infrastructure operators. But a direct comparison with the annual capex of Amazon, Microsoft or Alphabet would be misleading. Anthropic's figure covers several years and mixes cloud contracts, computing capacity and leases.
The structural difference, however, is relevant.
In 2025, Amazon made US$ 128.3 billion in cash capex, mainly in technology infrastructure and AWS expansion. Alphabet invested US$ 91.4 billion in capex, while Microsoft added US$ 64.6 billion in property and equipment in its fiscal year 2025.
These companies support the infrastructure with businesses that generate hundreds of billions of dollars in revenue and can distribute the capacity across thousands of products and customers.
Anthropic still depends mostly on a single product family and the AI model market. The prospectus also states that almost a quarter of 2025 revenue came from just two customers and warns that many of the largest buyers are not locked into long-term contracts.
This is the main economic mismatch: a significant portion of future costs is contracted for many years, while part of the revenue can disappear much more quickly.
2026 growth changes the interpretation of the US$ 518 billion
Looking only at the US$ 4.6 billion booked in 2025, however, also distorts the picture.
Anthropic's growth rate in 2026 was enough to rapidly change the relationship between revenue and infrastructure. In April, the company itself said it had surpassed US$ 30 billion in annualized revenue, versus approximately US$ 9 billion at the end of 2025.
At the end of July, annualized revenue had already surpassed US$ 65 billion, according to a source familiar with the figures cited by Reuters.
Annualized revenue is not equivalent to revenue actually recognized over a full year, especially in a business growing at this speed. Still, the jump shows that the infrastructure commitments were signed against a much larger business base than the one reflected in the 2025 financial statements.
The question, therefore, is not simply whether US$ 518 billion is excessive. It is whether Anthropic will be able to maintain enough growth and utilization to make these contracts represent productive capacity, not idle capacity.
This mechanism is already starting to appear in the numbers disclosed to investors. In August, Reuters reported that the company projected approximately US$ 190 billion to US$ 200 billion in revenue in 2028.
The same report pointed to an expectation of at least US$ 10.9 billion in revenue in the second quarter of 2026 and the company's first quarterly operating profit, of US$ 559 million.
These projections still need to be confirmed by subsequent results. But they make clear the financial assumption behind the expansion: Anthropic is contracting infrastructure for a company much larger than it is today.
Diversifying suppliers reduces one risk and creates another
The company is also avoiding dependence on a single architecture. Claude runs on AWS Trainium, Google TPUs and Nvidia GPUs, while new agreements add Microsoft, Broadcom and other suppliers to the compute map.
This diversity can reduce exposure to scarcity of a given chip or provider. Financially, however, it does not eliminate the minimum commitments.
The prospectus itself highlights an additional tension. Amazon, Google and Microsoft can simultaneously be infrastructure suppliers, distributors, investors, customers and competitors of Anthropic.
The company acknowledges that the interests of these companies may not remain aligned with its own.
The relationship with Akamai shows how this model spreads through the chain. An agreement registered with the SEC this month commits Anthropic to approximately US$ 11.6 billion over seven years for dedicated cloud capacity. The contract is large enough to have been classified by Akamai itself as material to its business.
Anthropic, therefore, is not just buying servers. It is creating commitments that help suppliers justify their own infrastructure expansion.
The IPO will have to prove that contracted capacity turns into margin
The sustainability of this model cannot be answered by the size of the contracts alone. It will depend on the relationship among three variables: revenue growth, infrastructure utilization and reduction in the cost per unit of intelligence delivered.
If demand continues to grow rapidly and improvements in chips, models and software allow more tokens to be processed per dollar, committing capacity in advance can secure supply in a tight market and improve unit economics over time.
If growth slows, competing models cut prices or technical advances reduce the amount of compute needed faster than expected, the same contracts could prevent costs from falling as fast as revenue.
That is precisely why the US$ 518 billion do not automatically turn Anthropic into a hyperscaler. They show that a frontier lab is taking on part of a hyperscaler's cost structure.
The difference lies in the balance sheet that supports this infrastructure. Amazon, Alphabet and Microsoft entered the AI race carrying large profitable businesses and cash flows capable of financing data centers for years. Anthropic is trying to build capacity first and grow revenue until it catches up.
The public IPO document should allow a better assessment of this bet. The most important data point will not be just the valuation sought by the company, but the annual schedule of these commitments, the evolution of margins, customer concentration and how much of the contracted capacity is actually being used.
These are the numbers that will show whether Anthropic has found sustainable economics for frontier AI companies or whether the race for compute has placed an infrastructure structure on a software company that its cash generation still needs to reach.
