Anthropic has built one of the fastest-growing businesses in technology. The problem is that running it currently costs considerably more than the business makes.
The company behind Claude generated approximately $4.6 billion in revenue in 2025, up from just $386 million the previous year. That is extraordinary growth for a company founded in 2021 and competing directly with some of the richest corporations on Earth.
Then comes the other number.
Anthropic lost just over $8 billion from operations during the same year.
Those figures come from the confidential IPO prospectus Anthropic submitted to the US Securities and Exchange Commission and subsequently reviewed by Reuters. Anthropic confirmed in June that it had confidentially submitted a draft S-1, giving the company the option to proceed with an initial public offering once the SEC review process is complete.
The numbers emerging from that document provide something the AI industry has rarely offered investors: a detailed view of what it actually costs to operate a frontier AI laboratory at enormous scale.
The answer is uncomfortable.
Anthropic spent $7.33 billion on compute and infrastructure during 2025 alone, substantially more than its entire revenue. Its total operating expenses reached $12.65 billion. And while the company is preparing to ask public investors to value its future, it is simultaneously locking in staggering amounts of computing capacity to make sure that future can exist.
One deal captures the scale perfectly.
Anthropic has committed approximately $11.6 billion over seven years to Akamai for dedicated cloud capacity and related services. The agreement can potentially expand by another $9 billion, taking the relationship toward $20 billion.
Even that is only one piece of a much larger infrastructure strategy.
According to Anthropic’s confidential prospectus as reported by Reuters, the company expects at least $518 billion in infrastructure commitments over roughly the next decade, with about 80% of those commitments either non-cancelable or requiring payment regardless of actual usage.
This is what the frontier AI business increasingly looks like beneath the chatbot.
The industry is still competing over intelligence.
It is also becoming a competition over who can afford enough machines to produce it.
Table of contents
- Anthropic is preparing Wall Street for the real economics of Claude
- $4.6 billion in revenue, $8 billion in operating losses
- Why Anthropic also reported a $42 billion net loss
- Compute already costs Anthropic more than its revenue
- The $11.6 billion Akamai deal changes the scale again
- The surprising part of the Akamai deal: CPUs
- $518 billion is the number investors cannot ignore
- Anthropic is betting that compute, not demand, becomes the bottleneck
- Anthropic depends on the companies it is trying to compete with
- The IPO will force a different conversation about AI
- The AI race is becoming an infrastructure race
- Anthropic IPO: key questions
Anthropic is preparing Wall Street for the real economics of Claude
Anthropic announced on June 1 that it had confidentially submitted a draft registration statement on Form S-1 to the SEC for a proposed IPO. That does not guarantee that the company will list, nor have the final number of shares or offering price been established, but it formally opened a process that could turn one of the world’s leading private AI laboratories into a publicly traded company.
Because the submission is confidential, the prospectus itself is not yet publicly available through the SEC in the way a conventional public S-1 would be. The detailed financial figures reported in late September come from journalists, principally Reuters, who obtained and reviewed the document.
That distinction matters. So does what the numbers reveal.
Private AI companies have spent the past several years announcing funding rounds, model benchmarks, enterprise customers and enormous infrastructure agreements. Those disclosures tell us plenty about how much capital is flowing into AI, but much less about the underlying economics of the companies consuming it.
An IPO changes the nature of the conversation.
Public investors eventually want to know how much revenue comes in, how much cash goes out, what produces the losses, how concentrated the customers are, what contractual obligations exist and, ultimately, how the business is supposed to generate returns on the capital being poured into it.
Anthropic’s numbers provide an unusually sharp answer to the first half of that equation.
The second half remains a bet on the future.
$4.6 billion in revenue, $8 billion in operating losses
The growth is difficult to overstate.
Anthropic generated approximately $386 million in revenue in 2024. One year later, according to figures from the confidential prospectus reported by Reuters, revenue had reached approximately $4.59 billion.
In twelve months, sales increased almost twelvefold.
For most technology companies, that would be the number dominating every conversation.
At Anthropic, it has competition.
The company’s operating loss widened from approximately $2.98 billion in 2024 to $8.06 billion in 2025, while total operating expenses climbed to roughly $12.65 billion.
Anthropic therefore spent considerably more running the company than it generated from customers even as Claude adoption exploded.
That does not necessarily mean the business model is broken. Frontier AI companies are investing aggressively in an industry whose ultimate size remains uncertain, and the economics can change dramatically as model efficiency, hardware utilization, pricing and demand evolve.
But the numbers illustrate the magnitude of what public investors are being asked to underwrite.
Anthropic is not a conventional software company temporarily sacrificing modest profits to acquire customers faster.
It is building a computational system whose appetite for infrastructure is measured in billions of dollars today and potentially hundreds of billions tomorrow.
Why Anthropic also reported a $42 billion net loss
There is an even more spectacular number in Anthropic’s 2025 accounts: a net loss of approximately $42 billion.
That figure deserves attention, but it also requires explanation.
Roughly $34 billion of the loss resulted from a non-cash accounting charge largely associated with the increasing fair value of financing instruments that could ultimately convert into Anthropic shares. As the estimated value of the company increased, the accounting value of those liabilities increased with it.
In other words, Anthropic did not burn $42 billion in cash operating Claude during 2025.
For understanding the underlying economics of the business, the approximately $8 billion operating loss is considerably more informative.
The $42 billion figure nevertheless demonstrates how strange the finances of enormously valuable private technology companies can become before an IPO. Financing structures designed during successive fundraising rounds can produce accounting effects that look extraordinary when translated onto a conventional income statement.
Once Anthropic becomes public, assuming the IPO proceeds, some of those distortions should become easier to separate from the operating story.
The operating story is dramatic enough without them.
Compute already costs Anthropic more than its revenue
Anthropic spent approximately $7.33 billion on compute and infrastructure in 2025, according to the prospectus figures reported by Reuters.
That expense was roughly three times the previous year’s level and represented about 58% of the company’s total operating expenses.
It was also approximately 1.6 times Anthropic’s entire 2025 revenue.
This is where comparisons between frontier AI companies and traditional software businesses begin to break down.
Software became one of the world’s most attractive business models partly because distributing another copy of a successful product could cost almost nothing. Development was expensive, but reproduction was cheap. Once Microsoft had built Windows or Adobe had built Photoshop, selling another license did not require constructing another factory.
Generative AI has different physics.
Every prompt requires computation. More users create more inference workloads. Larger models require vast clusters of accelerators to train. AI agents that spend minutes or hours completing complex tasks can consume far more computing resources than a conventional search query or software interaction.
Efficiency improvements matter enormously, but growing demand can absorb those improvements just as quickly.
Claude therefore exists simultaneously as software and as an industrial workload.
The interface is a conversation.
Behind it sits a machine park.
The $11.6 billion Akamai deal changes the scale again
On September 24, Akamai announced the largest contract in its history.
Anthropic has committed approximately $11.6 billion over seven years for dedicated cloud computing capacity and managed support services. The agreement expands a relationship established earlier this year and is subject to delivery, service-availability and termination conditions described in Akamai’s regulatory filing.
The relationship can become considerably larger.
Akamai says Anthropic can expand its commitment by up to another $9 billion, potentially taking the total toward $20 billion.
The financial engineering around the deal is almost as remarkable as the headline number. Akamai issued Anthropic a warrant that could eventually represent approximately 5% of Akamai’s outstanding common stock on an as-converted basis. Around 2% is associated with the initial commitment, while additional portions can vest as Anthropic purchases more services.
Akamai, meanwhile, expects to spend approximately $5.5 billion in capital expenditures to provide the required infrastructure.
This is not an AI company clicking “upgrade” on a cloud dashboard.
One company’s demand for computing power is large enough to reshape another public company’s capital allocation for years.
The surprising part of the Akamai deal: CPUs
Most discussions of the AI infrastructure boom revolve around GPUs.
Nvidia’s accelerators have become the defining hardware of the generative AI era because training and inference for large neural networks require enormous amounts of parallel computation.
The Akamai agreement highlights another layer of the infrastructure stack.
Akamai says the capacity will support Anthropic’s accelerating CPU workload demands through its distributed cloud infrastructure.
That detail is important because modern AI services require far more than the accelerators running neural networks. CPUs still handle vast amounts of orchestration, application logic, preprocessing, networking and general-purpose computation surrounding AI workloads.
The rise of AI agents could increase those requirements further.
An assistant answering a question is one kind of workload. An agent browsing websites, running code, manipulating files, interacting with applications and coordinating multiple operations over extended periods creates a much broader computational footprint.
The future of AI infrastructure therefore cannot be reduced to buying more GPUs.
As models become embedded into increasingly complicated software systems, the supporting machinery around them expands too.
Anthropic’s Akamai contract is evidence of that expansion.
$518 billion is the number investors cannot ignore
The Akamai commitment is enormous in isolation.
Placed beside Anthropic’s broader infrastructure plans, it becomes almost modest.
According to Reuters’ review of the confidential prospectus, Anthropic expects at least $518 billion in cloud, computing and infrastructure obligations over roughly the coming decade through agreements with multiple partners.
About 80% of those commitments are reportedly non-cancelable or require Anthropic to pay regardless of how much capacity it ultimately uses.
That is a remarkable strategic commitment for a company that generated $4.6 billion in revenue in 2025.
It does not mean Anthropic must write a $518 billion check tomorrow, and long-term infrastructure commitments should not be confused with a single year’s expenses. Revenue is also growing extremely quickly, while the company’s economics could look radically different several years from now.
But the scale tells us what Anthropic believes will be necessary to remain competitive.
Frontier AI is no longer a research competition in which the best laboratory wins because it has the smartest scientists and the cleverest model architecture.
Those things still matter.
Increasingly, they have to be attached to an industrial supply chain capable of delivering extraordinary quantities of computation.
Anthropic is betting that compute, not demand, becomes the bottleneck
Perhaps the most revealing claim reported from Anthropic’s prospectus concerns the company’s explanation for committing so aggressively to infrastructure.
Anthropic believes future AI demand could exceed the available computing capacity required to serve it, leaving growth constrained principally by access to compute.
That is an extraordinarily bullish assumption.
It means Anthropic is effectively betting that the dangerous mistake would not be reserving too many machines.
It would be reserving too few.
If that thesis is correct, the enormous infrastructure commitments begin to resemble strategic capacity reservations in an industry where the scarce resource determines market share. Anthropic would rather carry the financial burden of securing compute early than discover later that customer demand exists but cannot be served.
If the thesis is wrong, the same agreements look very different.
Unused computing capacity is expensive. Hardware depreciates quickly. New generations of accelerators can make previous generations economically less attractive. Model architectures can become more efficient. Competitors can push prices down. Demand can grow more slowly than expected.
That asymmetry is one of the central risks embedded in the AI infrastructure boom.
The industry is committing capital today based on assumptions about computational demand many years into the future.
Anthropic is making that wager on a colossal scale.
Anthropic depends on the companies it is trying to compete with
There is another complication buried inside the economics.
Anthropic competes with some of its most important investors and infrastructure providers.
Amazon and Google have both invested heavily in Anthropic while supplying cloud capacity and distributing Claude to customers. At the same time, Amazon develops its own AI products and Google develops Gemini.
According to Reuters’ review of Anthropic’s confidential filing, approximately 47% of Anthropic’s 2025 sales were routed through Amazon and Google cloud marketplaces.
That creates an unusual network of mutual dependence.
Anthropic needs enormous amounts of infrastructure and access to enterprise customers. The cloud giants want demand for their computing platforms and exposure to successful frontier models. Each can benefit from the other’s growth even while competing elsewhere in the AI stack.
Akamai adds another major infrastructure relationship while reducing, at least at the margin, Anthropic’s dependence on the hyperscalers for certain workloads.
This interconnected structure makes the AI economy look less like a collection of isolated companies and more like an industrial ecosystem in which competitors simultaneously finance, supply, distribute and challenge one another.
Following the money becomes almost as complicated as following the models.
The IPO will force a different conversation about AI
Anthropic going public would matter far beyond the company itself.
The largest frontier AI laboratories have developed mostly behind private-market walls. Investors have been able to assign enormous valuations based on funding rounds without requiring the same continuous financial disclosure expected from public companies.
An Anthropic IPO changes that.
Quarterly revenue becomes public.
Operating expenses become public.
Infrastructure commitments become public.
Margins become public.
Cash burn becomes public.
Investors will be able to watch the economics of frontier AI evolve quarter by quarter rather than reconstructing them from funding announcements and leaked projections.
That could make Anthropic one of the most important financial experiments in the technology industry.
Claude does not merely have to become more capable.
Anthropic eventually has to demonstrate that each generation of intelligence can be sold for more than the infrastructure required to produce and operate it — or that scale, efficiency and new products can eventually bend that relationship decisively in its favor.
The company’s extraordinary revenue growth suggests there is genuine demand.
Its extraordinary spending shows how expensive satisfying that demand currently is.
Wall Street will eventually have to decide what the gap between those two numbers is worth.
The AI race is becoming an infrastructure race
For much of the generative AI boom, competition was easy to visualize.
OpenAI released a model. Anthropic responded with Claude. Google released Gemini. Benchmarks moved. Context windows expanded. Coding scores improved. Prices fell. Users argued about which model was smartest.
That competition still exists, but underneath it another race has been accelerating.
Who has the chips?
Who has the data centers?
Who has the electricity?
Who has the networking capacity?
Who can reserve enough compute years in advance?
And, eventually, who can pay for all of it?
Anthropic’s IPO numbers make that second race impossible to ignore.
A company can grow revenue almost twelvefold in a year and still produce an operating loss larger than its entire sales. It can spend more on compute than customers pay it. It can simultaneously commit $11.6 billion to one infrastructure provider while planning hundreds of billions of dollars in broader computing obligations.
None of this proves that Anthropic’s strategy will fail.
The opposite outcome is entirely plausible. If Claude and future agentic systems become deeply embedded across businesses and software, today’s extraordinary infrastructure commitments could eventually look like the foundations of an enormous platform.
But that future now has a visible price tag.
For years, the central question surrounding frontier AI was how intelligent these systems could become.
Anthropic’s prospectus introduces another question that public markets will have to answer.
How expensive can intelligence become before the economics finally have to catch up?
Anthropic IPO: key questions
Is Anthropic definitely going public?
No. Anthropic announced in June 2026 that it had confidentially submitted a draft S-1 registration statement to the SEC for a proposed IPO. That gives the company the option to proceed after the SEC review process, but the timing, number of shares and offering price have not been finalized publicly.
How much revenue did Anthropic make in 2025?
According to its confidential IPO prospectus as reviewed by Reuters, Anthropic generated approximately $4.6 billion in revenue in 2025, up from roughly $386 million in 2024.
How much money did Anthropic lose?
Anthropic reported an operating loss of just over $8 billion in 2025. Its net loss was much larger, at approximately $42 billion, but roughly $34 billion of that resulted from a non-cash accounting charge largely associated with financing instruments that could convert into shares.
How much did Anthropic spend on compute?
Compute and infrastructure expenses reached approximately $7.33 billion in 2025, representing about 58% of Anthropic’s $12.65 billion in operating expenses.
What is Anthropic’s deal with Akamai?
Anthropic has committed approximately $11.6 billion over seven years for dedicated Akamai cloud computing capacity and related managed services, subject to contractual delivery, availability and termination provisions. The relationship can potentially expand by another $9 billion.
Why is Anthropic buying CPU capacity from Akamai?
Akamai says the infrastructure will support Anthropic’s growing CPU workloads. AI services require substantial general-purpose computing alongside specialized accelerators, particularly as increasingly complex AI agents execute code, interact with software and coordinate multiple tasks.
Is Anthropic really planning to spend $518 billion?
According to the confidential prospectus reviewed by Reuters, Anthropic expects at least $518 billion in infrastructure obligations over roughly the coming decade across multiple partners. The figure represents long-term commitments rather than immediate spending, and about 80% is reportedly non-cancelable or payable regardless of usage.
Is Anthropic profitable?
Not on the 2025 figures disclosed from its confidential prospectus. Despite extremely rapid revenue growth, Anthropic recorded an operating loss exceeding $8 billion.
Is Anthropic stock a good investment?
This article does not make an investment recommendation. If Anthropic completes an IPO, its eventual valuation, financial performance, infrastructure obligations, competitive position and offering terms would all need to be considered. Past growth does not guarantee future returns.
Sources: