Anthropic Filed for an IPO. Its Prospectus Shows the True Cost of Frontier AI.

For years, the frontier AI business has been funded mostly behind closed doors.

Venture capital firms, sovereign wealth funds, Amazon, Google, NVIDIA, and assorted billionaires kept writing increasingly absurd checks while OpenAI, Anthropic, xAI, and everyone else kept buying increasingly absurd amounts of compute.

Now Anthropic is preparing to let the public markets look at the bill.

And wow, there is a bill.

According to an IPO prospectus seen by Reuters, Anthropic generated $4.59B in revenue in 2025, up roughly 12x from $386M the year before. That is bonkers growth for a company founded in 2021.

It also posted an $8.06B operating loss.

And buried underneath those two numbers is probably the most useful snapshot we've gotten yet of the economics of building frontier AI.

Anthropic's filing basically says: Claude is growing unbelievably fast. Claude is also unbelievably expensive to build.

Welcome to the frontier.

First, about that $42 billion loss...

You'll probably see a lot of headlines saying Anthropic lost $42B last year.

Technically true. Also slightly misleading without an explanation.

Anthropic reported a GAAP net loss of roughly $41.97B in 2025, versus $8.31B a year earlier. But about $34B of that came from an accounting charge tied largely to financing instruments that could eventually convert into Anthropic shares. As Anthropic's estimated value rose, the accounting value of those liabilities rose with it.

In other words, Anthropic did not wheel $34B in cash out behind the office and light it on fire. Contrary to popular belief!

Think of it like issuing someone a ticket whose payout depends on how valuable your company becomes. If your company suddenly becomes much more valuable, accounting rules may force you to record the increased value of that promise as a loss today, even though you did not actually spend that amount operating the business.

So for understanding the actual business, the more revealing number is the $8.06B operating loss.

Which is still... a lot!

The number I'd pay attention to is $7.33 billion

According to Reuters, and as reshared by Wall St Engine, Anthropic spent $7.33B on compute and infrastructure in 2025, nearly triple the previous year's spending.

That represented about 58% of Anthropic's $12.65B in total operating expenses.

Put differently: Anthropic spent roughly $1.60 on compute and infrastructure for every $1 of revenue it generated.

That's the number that makes the whole frontier AI race easier to understand.

See, most software companies love scale because once the product exists, serving another customer can be relatively cheap. Frontier AI currently works very differently.

Every Claude response needs computation. Better models require enormous training runs. More customers require more inference capacity. New agentic products can consume vastly more tokens because the AI may work for minutes or hours instead of answering one question.

And Anthropic keeps reserving more capacity before the demand even arrives.

Earlier this year, the company committed more than $100B over ten years to AWS technologies, securing up to 5 GW of computing capacity. It separately expanded its Google and Broadcom relationship for multiple gigawatts of next-generation compute.

Which brings us to the wildest number in the prospectus.

Anthropic has roughly $518 billion in future infrastructure obligations

According to Reuters, the prospectus outlines about $518B in cloud, compute, and infrastructure obligations as Anthropic builds out its systems.

Important distinction: that doesn't mean Anthropic is spending $518B tomorrow. These are forward commitments associated with the infrastructure Anthropic expects to need over time.

But the scale tells you what management believes has to happen next.

Anthropic isn't planning for Claude to become a moderately successful software product.

It's reserving infrastructure for a world where AI becomes one of the largest computing workloads on Earth. That also explains the strange relationship between the AI labs and Big Tech.

Amazon is both an Anthropic investor and its primary cloud and training partner. Google is both a strategic partner and a major source of TPU capacity. Anthropic increasingly depends on companies that can simultaneously finance its growth and sell it the infrastructure required to continue growing.

As you can see, from this frosty vantage point, he frontier model business starts looking less like traditional SaaS and more like software strapped to a gigantic infrastructure financing machine.

The revenue growth is very real, though

The other mistake one could make digging through these numbers would be looking at the spending and ignoring what's happening on the revenue side.

Going from $386M to $4.59B in one year is extraordinary.

And Anthropic says that acceleration continued into 2026. In April, the company said annualized revenue had passed $30B, up from approximately $9B at the end of 2025. It also said the number of customers spending more than $1M annually had doubled to over 1,000 in less than two months.

That's why a potential $2T valuation is even being discussed.

Reuters reports the offering could value Anthropic above that level, although the final valuation, offering size, and timing remain subject to change. Anthropic originally confidentially submitted its draft S-1 in June, when it explicitly said the number of shares and price had not yet been set.

We've been tracking Anthropic's road toward what could become a record-setting IPO for exactly this reason.

We're about to find out what public investors think a frontier AI lab is actually worth.

There's another risk hiding inside the growth

Anthropic also disclosed that its two largest direct customers each generated 12% of 2025 revenue.

Two customers. Almost a quarter of the business.

And Anthropic warned that many large customers aren't committed through long-term contracts, meaning spending can potentially shrink or disappear.

That doesn’t mean they will. But it creates a pretty wild mismatch: nearly a quarter of Anthropic’s 2025 revenue came from just two customers, while the company has roughly $518B in future cloud, compute, and infrastructure commitments.

In other words, Anthropic is making enormous long-term bets on demand that some of its biggest customers are not contractually required to maintain.

That's worth remembering when looking at growth rates this extreme.

Anthropic is expanding incredibly quickly, but some of those dollars are concentrated among a relatively small number of very large customers. Meanwhile, the infrastructure required to serve that growth has to be reserved and financed years in advance.

At the same time, the company finished 2025 with $20.28B in cash, cash equivalents, and short-term investments.

So Anthropic has both an enormous war chest and an enormous appetite.

The GPU vendors remain delighted by this arrangement.

Then there's the weirdest part of taking Anthropic public

Anthropic isn't only telling investors that AI could become enormously valuable.

It's also telling them the technology could become enormously difficult to control.

This will come as no surprise to Neuron readers, but Reuters says the prospectus discusses increasingly autonomous systems behaving unexpectedly in controlled testing, including examples involving code sabotage, fraud assistance, and information manipulation. Those disclosures reflect a tension that has followed Anthropic almost since its creation: the company wants to build extremely capable AI while simultaneously warning that extremely capable AI may create serious risks.

That tension becomes more interesting once quarterly earnings enter the picture.

We recently explored the same problem facing OpenAI as it considers life with public shareholders: what happens when an AI company believes safety considerations require slowing down, but investors are expecting growth?

Anthropic may soon get to find out first.

The IPO is really a referendum on the economics of frontier AI

Reuters reports Anthropic's public debut may now happen after the November U.S. midterm elections, with the schedule still subject to change.

If the election goes the way it seems likely to go, this could lead to a positive outcome for Anthropic and its very blatant attempts at encouraging regulatory capture, which is where a few industry leaders buy enough influence in Washington to shape the direction of how their industry gets regulated, effectively pulling the ladder up behind them so only the privileged, anointed few get to compete seriously, while everyone else is priced out by costly, difficult to comply with standards.

You can see why they might want to wait to IPO until after the path is made much more friendly for their business model, which relies on them being the middleman broker of intelligence that everyone else builds on top of. It's not a bad strategy... but it's not favorable to developers or businesses who will lose out on cheap intelligence if Anthropic gets what they're after. Of course, there's more than "cheap intelligence" at stake, but we write about that all the time, so we'll leave that for other articles on recursive self improvement and the like.

Anyway, the politics of that IPO timing are less interesting to me than what happens once the stock actually exists on public markets.

Until now, frontier AI companies could raise another private round whenever they needed another mountain of GPUs.

Public investors operate differently.

Every few months they'll get to ask:

• How fast is revenue growing?
• How much does serving that revenue cost?
• Are models becoming cheaper to train and run?
• Are agents increasing usage faster than efficiency gains decrease costs?
• And, eventually, does this business turn all that intelligence into free cash flow?

Anthropic's prospectus gives us our clearest look yet at the starting point.

$4.59B in revenue. $7.33B in compute spending. $8.06B in operating losses. Roughly $518B in future infrastructure obligations.

The bet is that AI becomes so useful, and the market becomes so enormous, that revenue eventually catches the infrastructure curve. It seems likely, but of course there's that pesky AI bubble pop that will probably happen somewhere in the middle.

If it does, today's spending could look like the construction costs for one of the world's most valuable technology platforms.

If it doesn't, we'll have discovered one very expensive limitation of the frontier-model business model.

And for once, we'll get to watch Wall Street put a price on that experiment in real time.

If the success of Amazon's public market debut taught us anything, public markets can be surprisingly resilient to unprofitability so long as growth is on the menu. The question is, can the rest of the global economy be that patient?

Next up: When will the actual S1 filing get leaked? When it does, it may be one of the most scrutinized S1 filings of all time...

And what about the 2026 numbers, which are missing from this analysis and likely much more favorable to Anthropic? When will they be leaked...