Anthropic’s Reported IPO Prospectus Puts Its AI Infrastructure Bill at $518 Billion

Anthropic expects to incur at least $518 billion in cloud, computing and infrastructure obligations over roughly the next decade, according to a confidential IPO prospectus reviewed by Reuters. About 80% of the reported total is non-cancelable or must be paid even if Anthropic uses less capacity than it reserved.

The disclosures, reported on September 29, provide the most detailed view yet of the fixed costs behind Anthropic’s rapid expansion. They cover relationships with Google, Amazon, Microsoft, Broadcom, xAI and AMD, although the financial structures and cancellation rights vary substantially from one agreement to another.

The prospectus remains confidential and has not been published through the Securities and Exchange Commission. That means outside investors cannot yet examine the complete filing, payment schedules or contract definitions for themselves. The financial figures and specific contract terms in this article are therefore attributed to Reuters, which reviewed the document, while Anthropic and its partners have publicly confirmed the broader infrastructure agreements.

The central issue is not simply the size of the headline number. Anthropic is securing years of compute supply before it has equally durable commitments from many of its customers. If demand keeps accelerating, the contracts could give the Claude developer a major capacity advantage. If growth slows, model efficiency improves faster than expected or AI pricing falls sharply, the same agreements could become an expensive constraint.

What the $518 billion figure includes

The reported total is not a bill due at once, nor is it a conventional construction budget. It combines minimum cloud purchases, dedicated computing capacity, equipment leases and expected spending under agreements that extend as far as 2036.

The obligations associated with Broadcom, Google, Amazon and Microsoft add up to approximately $413.7 billion. Those four arrangements account for nearly 80% of the overall figure and represent the most rigid portion of the reported package.

The remaining agreements should not be interpreted in the same way. The xAI capacity is described as mostly cancelable, while the AMD figure represents expected capacity spending rather than a simple unconditional payment obligation. Adding every line together is useful for illustrating scale, but it does not mean Anthropic is guaranteed to pay every dollar under all circumstances.

There is also an unresolved naming discrepancy in the public record. Reuters’ account of the prospectus identifies xAI as the counterparty for up to $84.5 billion in Nvidia-based capacity. Anthropic’s May announcement, however, described its access to the Colossus 1 data center as a partnership with SpaceX. Publicly available materials do not explain whether the difference reflects corporate structuring, a revised agreement or separate but related arrangements.

Anthropic is locking in supply before equivalent customer demand

▶Play video

Anthropic’s rationale is straightforward: advanced-model development and deployment increasingly depend on access to scarce computing capacity. Chips alone are insufficient. A frontier AI company also needs data-center buildings, high-speed networking, cooling equipment, electrical connections and software capable of coordinating enormous clusters.

Anthropic has publicly described compute availability as a constraint on Claude’s growth. In April, it announced a commitment of more than $100 billion to Amazon Web Services over 10 years for up to five gigawatts of capacity. The same month, it expanded a Google and Broadcom partnership for multiple gigawatts of TPU capacity expected to begin coming online in 2027.

The company has also committed to large deployments on Microsoft Azure, Nvidia systems and AMD’s planned MI450 platform. That diversified approach can reduce dependence on a single accelerator architecture and allow Anthropic to assign workloads to different systems based on cost, availability and performance.

Yet the prospectus reportedly identifies a mismatch between Anthropic’s supply contracts and its customer relationships. Nearly one-quarter of its 2025 revenue came from two unnamed customers, while many large customers were not bound by long-term agreements and could reduce or stop their spending.

Anthropic is therefore assuming substantial long-term obligations without an equivalent guarantee that customer demand will persist on the same schedule. This is a familiar risk in capital-intensive industries, but it is unusual for a software-oriented company seeking a technology valuation.

The bet could work if demand continues to exceed available capacity. Reserving infrastructure early would let Anthropic serve customers that less-prepared competitors might have to turn away. The danger is that revenue, model usage and infrastructure costs may not move together. A downturn in demand could arrive well before Anthropic’s payment obligations expire.

Explosive growth has not yet produced operating profits

The financial figures in the reported prospectus help explain why Anthropic is willing to make such large commitments. Revenue reached almost $4.6 billion in 2025, roughly 12 times the previous year’s total. Compute and infrastructure expenses rose approximately threefold to $7.33 billion, accounting for more than half of the company’s $12.65 billion in total operating expenses.

Anthropic recorded an operating loss of more than $8 billion for the year. Its net loss approached $42 billion, but that larger figure requires context: roughly $34 billion came from an accounting charge tied to the rising estimated value of financing instruments that could eventually convert into Anthropic shares. It was not equivalent to $34 billion in cash spent on operations.

The distinction makes the net-loss headline less useful as a measure of day-to-day cash consumption. The operating loss nevertheless shows that the underlying business was spending substantially more than it generated in 2025.

Anthropic reportedly held $20.28 billion in cash, cash equivalents and short-term investments at the end of that year. On February 12, 2026, it announced a $30 billion funding round at a $380 billion post-money valuation, giving the company a much larger capital base before the heaviest infrastructure payments arrive.

Revenue also continued to accelerate in 2026. Anthropic said its annualized revenue run rate had exceeded $30 billion in April. Axios later reported, citing figures seen by Bloomberg, that the run rate passed $65 billion at the end of July and that preliminary second-quarter revenue exceeded $11.5 billion.

A run rate is not the same as audited annual revenue. It extrapolates recent performance and can change quickly. Even so, the increase helps explain why Anthropic is planning infrastructure at a scale that would have appeared disconnected from its business only a year earlier.

The key question for investors is whether revenue can grow faster than the cost of serving that demand. High usage is not automatically high-margin usage when every additional request requires expensive computing resources. Improvements in model efficiency, hardware utilization and pricing will be as important as customer growth.

Cloud providers are suppliers, investors and competitors

▶Play video

Anthropic’s infrastructure network is notable because many counterparties sit on both sides of its business. Amazon, Google and Microsoft provide computing services and distribution, while also investing in Anthropic or offering competing AI products.

Microsoft and Nvidia announced in November 2025 that they could invest up to $5 billion and $10 billion, respectively, in Anthropic. In the same announcement, Anthropic committed to purchase $30 billion in Azure capacity. Amazon said in April that it was making an additional $5 billion investment, with the possibility of investing up to $20 billion more, while Anthropic expanded its long-term AWS commitment.

AMD’s July agreement follows a similar pattern. AMD committed to a potential equity investment of up to $5 billion as Anthropic agreed to deploy as much as two gigawatts of AMD hardware.

These relationships help finance Anthropic while creating large customers for the suppliers’ infrastructure. They also complicate the economics. A cloud provider can earn revenue when Claude usage grows, benefit from the value of its Anthropic investment and distribute Claude to its own customers. At the same time, it may promote an internally developed model or another partner’s system.

The prospectus reportedly warns that the incentives of Anthropic’s cloud providers, investors, distributors and competitors may not always align with its own. That risk becomes more significant when the contracts last seven to 10 years and involve minimum spending requirements.

For the infrastructure companies, the commitments offer unusually long visibility into AI demand. Broadcom, AMD, Nvidia and the cloud platforms can plan equipment, networking and data-center capacity around large anchor customers. But they must also manage the risk of rapid hardware obsolescence, construction delays and shortages of electrical power.

Why the disclosure matters to the wider AI race

Anthropic’s reported obligations show that frontier AI is becoming less like a conventional software market and more like a capital-intensive infrastructure industry. Model developers increasingly compete not only on research and product quality, but also on their ability to finance and reserve chips, data centers and power years in advance.

The $518 billion figure is comparable to the original $500 billion target announced for the Stargate infrastructure initiative associated with OpenAI, Oracle, SoftBank and MGX. The structures are different, however. Stargate was presented as a shared development program involving several financial and technology partners. Anthropic’s figure represents obligations and expected spending attached to a single AI developer across multiple contracts.

That difference makes Anthropic’s disclosure particularly relevant to prospective shareholders. It connects the infrastructure race directly to one company’s future payment requirements, rather than describing only what a consortium hopes to invest.

The obligations may also reinforce advantages held by the largest AI laboratories. A company capable of signing multibillion-dollar minimums can secure scarce supply and negotiate earlier access to new hardware. Smaller rivals may be left purchasing capacity at higher prices or waiting for it to become available.

At the same time, efficiency could change the competitive equation. If new model architectures perform the same work using fewer chips, the value of owning or reserving the largest cluster may decline. If increasingly capable models require still more computation, Anthropic’s early reservations could become a strategic asset.

What investors and Claude customers should watch

Anthropic reportedly filed confidentially for an IPO in June. Reuters has reported that a public-market debut could be delayed until after the November 3, 2026, U.S. midterm elections and could seek a valuation above $2 trillion. No public prospectus, final timetable or price range has been released.

When an amended filing becomes public, several disclosures will be especially important:

• Annual payment schedule: Investors need to know when the largest minimum payments and leases begin, rather than relying on a decade-long aggregate.
• Cancellation and shortfall provisions: The practical risk depends on which commitments can be reduced and how much Anthropic must pay when usage falls below contracted levels.
• Infrastructure utilization: High utilization could turn reserved capacity into revenue; low utilization would leave Anthropic carrying expensive idle resources.
• Gross margin: Revenue growth must be weighed against the cost of training models and serving customer requests.
• Customer concentration and contract length: Anthropic will be in a stronger position if major customers accept terms that more closely match its long obligations to suppliers.
• Financing structure: Investors will want to distinguish cash purchases, cloud minimums, equipment leases and obligations supported by partners or special-purpose financing.
• Counterparty dependencies: The final filing should clarify how the Broadcom, Google, xAI or SpaceX, and other capacity agreements relate to one another.

Claude customers may see direct benefits if the buildout arrives on schedule. Additional capacity can support higher usage limits, better availability and larger enterprise deployments. Anthropic has already linked new infrastructure agreements to increases in Claude Code and API limits.

The prospectus nevertheless turns compute into Anthropic’s defining financial risk. The company is betting that securing extraordinary capacity now will allow it to satisfy future demand and spread fixed costs across a much larger revenue base. Its prospective IPO will test whether public investors believe that rapid AI adoption can become durable profit before the industry’s enormous infrastructure bill comes due.