Anthropic is preparing to enter the public markets with a story that is almost impossible to fit into a traditional technology-company template.

The artificial-intelligence startup has produced explosive revenue growth, built one of the world's best-known frontier AI platforms and accumulated enormous financial backing from some of the biggest technology companies on the planet.

But the company's newly revealed IPO prospectus also contains numbers that show just how expensive the AI arms race has become.

Anthropic generated nearly $4.6 billion in revenue in 2025, a roughly twelvefold increase from the previous year. Yet it recorded a net loss of about $42 billion, including a roughly $34 billion accounting charge related to the estimated value of financing instruments that could eventually convert into Anthropic shares. On an operating basis, excluding those liability-related writedowns, the company lost more than $8 billion.

The headline loss is therefore more complicated than the $42 billion figure initially suggests.

Most of that loss was not cash Anthropic spent operating its AI business.

But the underlying operating loss is still enormous.

And then comes the number that may be even more important for investors:

$518 billion.

Anthropic's prospectus and related disclosures show that the company expects to commit at least $518 billion over the coming decade to cloud computing, chips, infrastructure and related AI capacity. Around 80% of those commitments are described as non-cancelable or requiring payment regardless of how much computing capacity Anthropic ultimately uses.

That puts Anthropic's prospective IPO directly at the center of a much bigger question about the economics of artificial intelligence.

How much money will the industry have to spend before advanced AI becomes consistently profitable?

Anthropic's growth story is extraordinary.

Revenue jumped twelvefold in 2025 to nearly $4.6 billion. The company has continued expanding Claude, its family of AI models, across consumer, enterprise and developer markets. Its growing customer base includes businesses that increasingly rely on AI for software development, research, analysis and other knowledge-intensive tasks.

But AI models require enormous amounts of computing power.

Anthropic spent roughly $7.33 billion on computing and infrastructure in 2025, about three times its spending the previous year. That represented more than half of its total operating expenses of roughly $12.65 billion.

In practical terms, Anthropic spent more on computing and infrastructure than it generated in revenue during the year.

That is the uncomfortable financial reality hiding behind the spectacular growth rates.

The economics of frontier AI are fundamentally different from many software businesses.

Traditional software companies can often add millions of users without a proportional increase in physical infrastructure costs.

Frontier AI does not work that way.

Every new customer creates additional inference demand.

Every more capable model requires additional training.

Every new feature can require more compute.

And when AI agents become capable of performing tasks continuously rather than answering occasional questions, the amount of processing required can increase even further.

Anthropic is betting that revenue will ultimately grow much faster than those costs.

The prospectus reveals just how aggressively Anthropic is attempting to lock in computing capacity before competitors consume it.

The company has agreements involving some of the largest technology and semiconductor companies in the world.

Anthropic expects to spend at least $111.1 billion with Google's cloud business, $110 billion with Amazon, and $31.4 billion with Microsoft under long-term infrastructure arrangements. It also disclosed approximately $161.2 billion of Broadcom-related equipment lease obligations, most of which are non-cancelable.

The agreements have different structures and timelines, but the message is unmistakable.

Anthropic sees computing availability itself as one of the biggest constraints on future AI development.

The company told investors that demand for advanced AI could exceed available supply and said future growth may be “limited principally by the availability of compute.”

That is a remarkable statement from an AI company preparing to sell shares to the public.

It suggests Anthropic does not view demand as the primary problem.

The problem, in its view, is securing enough processors and infrastructure to satisfy that demand.

That belief is also pushing Anthropic to change the way it builds its business.

Historically, the company relied heavily on cloud providers.

Now it says it is increasingly moving toward directly controlled infrastructure, dedicated data centers and leased or owned computing capacity.

The reason is strategic.

Cloud dependence provides flexibility, but it can also create vulnerabilities.

Anthropic's biggest cloud providers — Amazon, Google and Microsoft — are not neutral suppliers.

They are simultaneously investors, distributors and competitors developing their own AI models.

Anthropic explicitly warned investors that those relationships create incentives that may not be fully aligned with the company's interests. It also said its business could be adversely affected if third-party computing capacity were curtailed, repriced or terminated.

That creates a fascinating business model.

Anthropic needs these companies to build and distribute its AI.

But those same companies are developing competing products.

The race for compute is therefore also a race for strategic independence.

Anthropic's infrastructure strategy is not limited to Google, Amazon and Microsoft.

The prospectus also disclosed agreements with Elon Musk's xAI that could generate as much as $84.5 billion in spending on Nvidia-based computing capacity through 2029, although those commitments are largely cancelable with 90 days' notice.

Anthropic also has a relationship with AMD. The semiconductor company has agreed to buy up to $5 billion of Anthropic stock and supply computing capacity expected to exceed $20 billion.

The arrangement effectively links three major parts of the AI ecosystem.

Anthropic needs compute.

Chipmakers need huge AI customers.

Cloud and infrastructure companies need long-term demand to justify enormous capital investments.

Those relationships are becoming increasingly intertwined.

The prospectus also points toward an extraordinary potential valuation.

Anthropic could seek a public-market valuation above $2 trillion, according to Reuters reporting, more than double the company's estimated value of roughly $965 billion in May. The eventual IPO is expected after the November U.S. midterm elections, although the precise timing remains subject to change.

That would place Anthropic among the most valuable technology companies ever to approach the public market.

And that creates a difficult comparison.

A company producing less than $5 billion in recognized annual revenue would be seeking a valuation measured in trillions of dollars.

Investors would therefore be valuing not the company Anthropic is today, but the company it hopes to become.

That makes the IPO a referendum on the future economics of frontier AI.

Anthropic's growth also comes with customer-concentration concerns.

The prospectus says nearly a quarter of the company's 2025 revenue came from just two customers. It also warns that many major customers do not have long-term contracts and could reduce or stop spending.

That matters because AI spending is still relatively young.

Companies are experimenting with different models and providers.

A major customer can potentially change vendors if another system becomes cheaper, more capable or better integrated into its workflow.

Anthropic therefore has to keep releasing competitive models while simultaneously persuading customers that Claude is sufficiently valuable to justify growing AI budgets.

That creates a difficult cycle.

The company must spend heavily to remain at the frontier.

Its customers must spend heavily to use the resulting systems.

And investors must believe the eventual economic value will be large enough to cover all of it.

Anthropic's IPO filing contains an unusually extensive discussion of AI safety.

Around 80 of the 261 pages in the main body of its prospectus are devoted to risk factors — nearly twice the 48 pages used to describe its business.

The company warns that advanced AI systems could potentially display self-preserving behavior, resist shutdown, conceal or manipulate information and behave in ways resembling blackmail.

It also warns that some capabilities may not become visible until models are deployed, making safety evaluation difficult.

This is extraordinary because Anthropic is not merely acknowledging the ordinary risks associated with a software product.

It is telling potential investors that the technology itself could create catastrophic or existential risks if improperly developed or deployed.

The company nevertheless argues that reliable, trustworthy and secure AI systems will ultimately be rewarded by the market.

That creates another tension.

Anthropic says AI development must move quickly enough to compete.

At the same time, its own prospectus says safety work consumes scarce computing resources and talent.

The company reported that about 6% of its AI research computing capacity went toward safety work during a sample week in July.

Anthropic's IPO will therefore test several assumptions simultaneously.

Can AI revenue scale faster than computing costs?

Can long-term contracts secure enough infrastructure without creating crushing obligations?

Can customers remain loyal when competition from OpenAI, Google, Meta and xAI is intensifying?

Can Anthropic maintain leadership while spending billions on models, chips and data centers?

And perhaps the most basic question:

Can frontier AI become a durable, profitable business rather than an industry permanently dependent on enormous external financing?

Those questions make Anthropic's prospectus much more important than an ordinary IPO filing.

The company has already demonstrated that businesses are willing to pay billions for AI capabilities.

What remains unproven is whether those capabilities can ultimately produce returns large enough to justify the extraordinary infrastructure spending required to build them.

Anthropic's numbers tell the story in three figures:

$4.6 billion in 2025 revenue.

More than $8 billion in operating losses.

At least $518 billion of future infrastructure commitments.

That is not simply a startup preparing to go public.

It is one of the world's biggest financial experiments in whether artificial intelligence can support the staggering cost of creating it.

And when Anthropic eventually rings the opening bell, investors will not just be buying into Claude.

They will be putting a market value on the economic future of frontier AI itself.