Anthropic Prospectus Discloses: $80.7 Billion Loss in First Half, Compute Orders Exceed $800 Billion

Add crowdsourced testing promotional copy Add crowdsourced testing promotional copy end Add App-exclusive traffic routing

Humanity's biggest IPO in history has arrived—or rather, it has arrived again.

Four months ago, Musk had just taken his SpaceX to break the IPO record, raising $85.7 billion at a valuation of $1.77 trillion.

This time Anthropic is even more exaggerated. Although the offering details haven't been announced yet, the rumored fundraising target has already hit $100 billion.

What does that mean?

Let me put it this way: last year, all of 2025, not to mention the A-share, Hong Kong, or US markets—the world's 1,293 IPOs raised a total of $171.8 billion. Put together, they still can't suck up as much as these two brothers...

Of course, even setting aside the exaggerated numbers, Anthropic's prospectus this time is still quite worth reading. It discloses a lot of billing details and operating conditions, packed with substance.

My impression after reading it is that Anthropic is already trying to redefine the AI market, because what it sells is the new concept of "knowledge labor."

And the size of this pie is even more imaginative than Musk's previous one, reaching a full $40 trillion...

First, the most eye-catching set of data: in Q1 2026, revenue reached $4.731 billion, directly crushing all of last year.

Q2 was even more vigorous, with $11.56 billion in revenue, up 1,368% year-over-year, more than 13 times. Q3 wasn't directly disclosed, but based on clues left in the business segment, a quick calculation puts it at around $16.8 billion, with growth still continuing...

At the same time, the loss figure for the first half of the year was also absurdly huge, directly hitting $80.748 billion.

No way—even if Anthropic took the money and burned it for fun, it couldn't lose this much, right?

Well, don't say that—there really is a hidden reason.

If you look carefully through the financial report, you'll find that of this $80.7 billion, $72.388 billion comes from changes in the fair value of financial instruments such as convertible bonds.

In plain terms, Anthropic previously raised a large sum of money, and many of the financing instruments can be converted into company stock in the future under certain conditions.

When the financing happened, the price was still low, but as the company's valuation kept rising, these things became more and more valuable.

For the investors holding them, this is of course a good thing—they can make money lying down—but for Anthropic, it's purely a liability, recorded on the books as a huge loss.

That's why there's this absurd picture: the company becomes more and more valuable, while the income statement looks worse and worse.

But to put it plainly, Anthropic didn't actually pay this out in real money from its bank account; this part just looks ugly.

If you strip these things out, under Anthropic's own disclosed non-GAAP basis, the first half of the year still had net profit after adjustments, and it had already reached $2.671 billion.

That's right—its cash flow is actually quite normal...

And the user data supporting revenue is also strong. There's a metric called net revenue retention (NRR); excluding subscribers, the figure is about 900%.

In plain language, last year's old customers spent about 9 times more this year than in the same period last year.

It's indeed sticky enough. With this level of payment appeal, why worry about not making money?

But this part doesn't have much to do with us. After all, you can occasionally still see a few grasshoppers after autumn, but a Claude account that survives a full year is truly rare.

What's funniest is that Anthropic also casually disclosed a set of data, saying that 5%–7% of revenue is deemed to come from customers using it in violation of the rules.

Remembering the science-popularization short videos made by Opus 5.5 all over the mountains and plains a few days ago, this is almost naming names...

With data stacked on data like this, according to their estimate, the economic value corresponding to global knowledge labor in 2025 exceeded $40 trillion.

Of course, this $40 trillion is also standard pie-in-the-sky stuff, just like Musk saying he wants to colonize Mars—please respect and understand it, but for now it's only fighting for the future.

After all, with a pie drawn this big, to bake it, how much batter do you need to mix, and how big a furnace do you need to build?

Well, Anthropic is also very honest—it immediately lays out its spending budget for you.

It went through all the computing power suppliers on the market—Google, Amazon, Microsoft, SpaceX, and so on. Orders from eight companies, with contracts scheduled as far out as 2045.

Among these orders, the tens-of-billions level is just an appetizer; only the hundreds-of-billions level can get a seat at the table. Stacked together, they total more than $800 billion.

Yes, Anthropic's pre-booked bill for fighting for the future is roughly equal to one Musk. In terms of computing power, this guy is truly scared of being poor.

But these contracts can't simply be seen as liabilities, because apart from delivery details, many of these companies are more than just creditors.

For example, in May this year, Amazon invested $5 billion in Anthropic, and Google was even more generous, investing $10 billion.

And if you've used their respective Agent products, like Amazon's Kiro or Google's Antigravity, you'll find that they also include Claude's models.

Beyond that, after the partnership they can also deploy Claude on their own servers and then sell the API. In the first half of 2026, 48% of Anthropic's revenue already came from sales through these third-party cloud platforms.

That's right, these big companies are even craftier than we thought: they make money on the investment, make money renting out computing power, and make another cut by helping sell the product...

This may also be Anthropic's awkward spot. To be honest, its ideals are still too idealistic.

Although it has always thought quite highly of itself—it even chose PBC (public benefit corporation) as its company registration type—tokens don't care about any of that. Without computing power, it simply can't spit them out.

Half your revenue depends on the giants selling for you, and you've already booked over $800 billion in orders waiting to recoup the investment—and now you still want not to be led around by commercial interests? That's a real headache.

After all, even someone as willful as Musk—complaining about Anthropic one moment, then landing a computing-power rental order the next—can smile and cooperate.

In such a chaotic capital market, when AI safety and making money come into conflict, can you, Company A, still hold on to the original aspiration you set for yourselves?

In response, they simply set up an extremely complicated equity structure.

As everyone knows, founders of listed companies, in order to safeguard their say, generally separate the right to dividends and profits from the right to vote on management.

Well, Anthropic did the same, and in even finer detail: they split the stock into 5 classes—Class A, N, E, T, and F.

Among them, Class A shares are normal stock, the kind ordinary investors buy, which can both earn money and carry voting rights.

N shares are mainly reserved for strategic partners, like Google and Amazon mentioned earlier: they can share in the profits but basically cannot participate in company decisions.

The really interesting ones are the latter three classes.

First, F shares, reserved specifically for the seven co-founders. There is only one such share in total, it does not participate in dividends, yet it holds 50.1% of the company's voting rights.

In other words, even if they raise round after round of funding in the future, as long as this mechanism remains, the seven founders can jointly control the decisive power of shareholder votes.

But relying on the founders alone is obviously not enough, so they created T shares and handed them to an independent organization called LTBT (Long-Term Benefit Trust).

This thing is even more extreme: it takes no dividends and does not participate in ordinary shareholder voting, yet it can decide the selection of the majority of the company's board seats.

Put plainly, they specifically found a group of people who do not make money from Anthropic's stock price to supervise it neutrally.

Finally, E shares are employee governance shares: they have no dividend rights, but can vote when amending the trust rules or when disputes arise over the selection of some directors.

In this way, each party has its own power, and no one can use a single class of stock to monopolize all decisions—adding several more layers of armor to their original aspiration.

In addition, they also set up a public benefit fund holding 5% of Anthropic's shares, and will keep adding to it as the stock price rises. They say it is dedicated to the public interest, but how exactly it will be distributed has not been specified.

Still, I think Musk's suggestion is pretty good: just give it all away to everyone and be done with it.

So don't mind that we usually keep griping about Anthropic—for such a big player to go public and be willing to go through a whole experiment in corporate governance is still worth discussing.

After all, all of this was previously sealed inside the company, kept secret layer by layer. Never mind the specific cost breakdown—we probably couldn't even see which companies they were making eyes at.

Once the prospectus is disclosed, at least we can see clearly what their business is actually like and exactly how they burn money.

As for their management methods and how much of their original aspiration they can ultimately hold on to, that is still for time to verify.

Speaking of which, OpenAI's IPO is not far off either. We're looking forward to that too, to see what new and unseen things this other veteran can pull out...article_adlist[

Editor in charge: Ruofeng

Report article content ]article_adlist Statement: This is an exclusive article from Sina. Reproduction without authorization is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin

Non-targeted 300*250 button end