Anthropic prospectus exposed: revenue of $16.3 billion in half a year, book loss of $80.7 billion

By | Blue Character Plan, Author | Huang Xiaobin

On October 20, as the media exposed the draft IPO prospectus of Anthropic, a set of jaw-dropping numbers stunned the entire Silicon Valley:

In the first half of 2026, the revenue of this large-model giant surged to 16.3 billion US dollars, a nearly 14-fold increase from 1.1 billion US dollars in the same period last year; but looking down at the income statement, one's heart sinks—the net loss for the half year was as high as 80.7 billion US dollars...

That is equivalent to, on average, 440 million US dollars evaporating from the books the moment one opens one's eyes each day.

But if one's gaze stops only at the "80.7 billion loss," one would be misreading it.

In Silicon Valley's capital game, this 80.7 billion huge loss is actually a standard "paper illusion." And turning to another page of the prospectus, a more thought-provoking bill comes into view:

A batch of long-term computing power and infrastructure contracts spanning many years into the future, totaling more than 600 billion US dollars, that Anthropic signed with computing giants such as Google and Amazon.

If the new agreements under negotiation are finalized, the total commitment scale could reach as high as 750 billion US dollars.

Having earned only 16.3 billion US dollars in half a year, yet rushing to sign hundreds of billions of dollars in computing power leasing contracts in advance... Isn't Anthropic being a bit too bold?

"The more valuable, the more money-losing"

To see clearly what Anthropic is thinking, one still has to first understand what the situation is with that frightening 80.7 billion US dollar loss.

Why can a company that brought in 16.3 billion US dollars in half a year show an astronomical loss of 80.7 billion on its books? The answer is actually somewhat counterintuitive:

The company's valuation rose too sharply, which instead pushed up the book loss.

According to the draft prospectus disclosed by the media, of this 80.7 billion loss, a full 72 billion US dollars was not real money spent at all, but was instead related to the revaluation of financing instruments such as "convertible notes."

To put it in plain terms: when Anthropic raised funds in its early years, it issued to investors a type of "IOU" that could later be converted into stock. As the company became more and more valuable, these financing instruments in investors' hands also became more valuable.

Under the applicable accounting rules, the company needs to recalculate the value of these financing instruments; the related appreciation increases book liabilities and is recorded as a loss for the period. Thus, a surge in valuation is originally a good thing, but on the income statement it turns into a frightening loss.

What is hard to bear is that, despite such a large loss, this 72 billion US dollars was neither used to buy a single Nvidia graphics card nor deposited into any programmer's payroll card—it is purely a numbers game triggered by the company being too valuable.

So, after stripping out this 72 billion "fake loss," where did the money Anthropic spent actually go?

Computing power and infrastructure are one of the big items.

In the first half of 2026, Anthropic's spending on computing power and data center infrastructure alone approached 9 billion US dollars, nearly four times the 2.3 billion US dollars in the same period last year, and in half a year it even exceeded last year's full-year total computing power bill of 7.3 billion US dollars.

Converted into a more intuitive ratio: for every 1 US dollar received from customers, about 55 cents is spent on computing power and infrastructure, and the remaining about 45 cents still has to cover scientists' salaries, R&D, promotion, and other expenses.

However, if one does the accounts carefully, there is actually good news hidden here:

In the first half of last year, Anthropic earned 1.1 billion but had to spend 2.3 billion on computing power, with computing power costs at more than 200% of revenue; by the first half of this year, revenue soared nearly 14-fold to 16.3 billion, and although computing power spending rose to 9 billion, its share of revenue had already dropped sharply to about 55%.

Even by the third quarter of this year, single-quarter revenue further climbed to 16.8 billion US dollars, and on an adjusted basis the company had already begun making money.

By logic, seeing the speed of making money finally outrun the speed of burning money, Wall Street should be opening champagne to celebrate.

Why, then, are institutions instead sweating over it at this very moment?

Still living a tightrope-walking life

Behind Anthropic's tens of billions in revenue, the biggest patrons are those enterprise customers paying for Claude.

Relying on its model Claude's long-document processing, compliance and security, and coding capabilities—especially the hit tool Claude Code—the businesses of banks, law firms, insurers, and tech companies have become Anthropic's main battleground for making money aggressively.

According to a report published by Menlo Ventures in 2025, Anthropic's share of the enterprise API market rose from 12% in 2023 to about 40%, surpassing OpenAI's 27%.

However, opening the customer ledger, you will find that even if Anthropic has become the "sales champion," it does not mean it can rest easy.

According to Reuters, in 2025, more than 80% of Anthropic's revenue came from usage-based paying customers, like paying utility bills at home—pay for as much as you use. This model seems to show surging cash flow, but in fact hides two major risks.

First, a large chunk of its revenue is held in the hands of a few big patrons.

According to the new draft prospectus disclosed by The Wall Street Journal, of the 16.3 billion US dollars in revenue in the first half of 2026, a full one-fifth—that is, more than 3.2 billion US dollars—came from just two major customers.

Second, these big patrons may not have signed a "long-term meal ticket."

The traditional enterprise software business allows companies to sign three- or five-year procurement contracts with customers, locking in a portion of revenue in advance. But according to disclosures, many of Anthropic's largest customers have not signed long-term procurement commitments, so in theory these big Anthropic backers could reduce or even stop spending.

Once these mega-customers decide to cut budgets, pivot to building their own models, or hand some tasks to competitors, Anthropic's tens of billions in revenue could shrink sharply.

To sum up its situation in one sentence: whether customers will still be willing to spend this much next year is completely unknown.

If it were just a nimble software company, it could find ways to cut costs when customers dwindle, tighten its belt, and still get by. But the fatal problem is that, without having locked in long-term revenue, Anthropic turned around and signed an almost irrevocable sky-high bill on the cost side—namely the compute capacity reservation contract of up to 750 billion US dollars mentioned earlier.

No wonder Wall Street had to gasp at this.

Who exactly is working for whom?

More than 600 billion US dollars in forward compute and infrastructure commitments can be called the super bet Anthropic has placed.

Among them, Anthropic committed to purchasing about 111 billion US dollars from Google Cloud, about 110 billion US dollars from Amazon AWS, and about 31 billion US dollars from Microsoft. In addition to several cloud giants, it also procures Nvidia chip compute through suppliers such as Nscale and Lambda, with related agreements worth up to about 86 billion US dollars.

In addition, around leasing arrangements for Broadcom and Google TPU chips, it has taken on more than 160 billion US dollars in forward payment obligations. Beyond these main contracts, there are other compute and infrastructure arrangements, and overall commitments already exceed 600 billion US dollars.

But even harsher than the figure of several hundred billion US dollars are the rigid payment clauses in the contracts.

Documents disclosed by Reuters in September showed that of the then-total 518 billion US dollars in commitments, about 80% were irrevocable or required payment regardless of actual usage. Some agreements with Google and Amazon even require that if actual consumption falls short of the agreed amount, the difference must be made up.

This makes the whole thing seem very surreal.

It is like opening a trendy restaurant where all the customers are walk-ins who do not buy membership cards; even if the line stretches to the end of the street today, no one can guarantee they will come back to eat next year.

But to secure the venue and equipment first, you sign a ten-year lease in one go and promise suppliers to purchase a certain amount of ingredients each year. Even if three years later business is so slow you are swatting flies, the rent you owe and the procurement payments you must make cannot be short by a single cent.

So why does Anthropic, knowing the risks are huge, still sign such rigid contracts?

Because without securing compute in advance, it might not even have the qualification to stay at the table.

In this Silicon Valley game of power, Google, Amazon, and Microsoft play an extremely delicate "triple role": they are major shareholders investing money in Anthropic, landlords collecting compute tolls from it, and at the same time they hold their own beloved children such as Gemini and Copilot, fighting Claude head-on in the market.

The prospectus reveals a rather realistic detail: in fact, the scale of compute Anthropic wants has already exceeded the upper limit these cloud giants are willing to supply it. To avoid being strangled by giants that are "both shareholders and competitors," and to prevent one day offending enterprise customers because of a compute gap, it can only lock in big-company capacity with sky-high long-term contracts while gritting its teeth and turning to second-tier service providers with less track record, such as Nscale and Lambda, betting real money for the next decade on the table in advance.

At the top are the chip overlords and cloud giants sitting securely in the fishing pavilion, using ten-year rigid contracts to lock in hundreds of billions of US dollars in future cash flow with guaranteed returns regardless of drought or flood; while Anthropic, standing in the spotlight, appears on the surface to be a glamorous trillion-level tech unicorn, but in the whole chain it is merely a slightly plumper shrimp.

References:

Anthropic IPO draft prospectus (2026-10-20)

36Kr: "Anthropic Prospectus Latest Disclosure: Revenue 16.3 Billion, Book Loss as High as 80.7 Billion US Dollars"

Sina Finance: "Anthropic Submits IPO Prospectus: Revenue Soared 12-Fold Last Year, Net Loss 42 Billion US Dollars"

Shanghai Securities News: "Anthropic Sprinting Toward the Largest IPO in History"

New Intelligence Era: "Huge Loss of 80,000,000,000 US Dollars in Half a Year! Anthropic Prospectus Exposed"

Zhidongxi: "Anthropic, Huge Loss of 538.9 Billion in Half a Year"

AI Frontline: "Valuation Cut by 500 Billion US Dollars on the Eve of Listing! Anthropic's Latest Prospectus Reveals Huge Loss of 80.7 Billion in the First Half—How Much Longer Can the Myth of Huge Profits in Large Models Last?"