Anthropic's IPO Filing Shows a Company Betting Its Valuation on Becoming a Digital God

Anthropic filed the paperwork this week to go public on the Nasdaq, and the numbers inside that filing are the closest look anyone has gotten at what it actually costs — and pays — to run a frontier AI lab. The company could start trading as early as next month, with early estimates putting its valuation somewhere between $1.5 trillion and $2 trillion. That alone would make it one of the most valuable companies on Earth before it's traded a single public share.

But the valuation headline buries the more interesting story. Anthropic's S-1 runs hundreds of pages, and buried in it is a detailed, audited picture of how an AI lab's finances actually work — something outside analysts have mostly had to guess at until now. Joey Brookhart, an analyst at the chip and AI research firm SemiAnalysis who co-leads the firm's "Tokenomics" model (its effort to map the unit economics of AI businesses), told The Verge's Decoder podcast that the filing landed about as well as a company could hope. "It was an awesome S-1," Brookhart said, pointing to the company's growth to more than $70 billion in annualized recurring revenue in a short window, alongside margins that moved in exactly the direction investors want.

The margin turnaround is real, but it's not the whole story

Two years ago, Anthropic's gross margins on serving its models — what the industry calls inference — were negative 94 percent, according to Brookhart. Today that number has flipped to roughly positive 85 percent. Revenue per megawatt of inference compute has climbed from the mid-teens in millions last year to around $65 million now, and SemiAnalysis's modeling suggests upcoming hardware and model generations could push that figure to $100 million to $120 million per megawatt. Meanwhile, the company is paying a little over $10 million per megawatt in compute costs. That's a wide and growing spread.

Here's the catch: only about 30 percent of Anthropic's compute actually goes toward serving paying customers. The rest funds research, experimentation, and training the next model. And according to Brookhart, Anthropic wants to push that ratio further toward training, not less — because the company isn't optimizing to run a profitable software business today. It's optimizing to build what it believes could be artificial general intelligence, and it has the margins to fund that bet. "If you believe you're building a digital god or going towards AGI, and as long as you have the gross margins of 85 percent-plus on inference to do so, you can," Brookhart said.

That framing matters because it explains a number that alarmed plenty of people reading headlines this week: Anthropic reported an $80 billion loss in the first half of 2026. Brookhart's explanation is more mundane than it sounds. It's a non-cash accounting charge tied to the company's rising valuation and the warrants and convertible notes it has issued, including to Amazon. As Anthropic's value climbs, the accounting value of those instruments climbs too, and that shows up as a paper loss even though no cash left the building. The real cash burn, Brookhart said, has been about $10 billion — a fraction of the headline number — against more than $120 billion in cash and marketable securities already on the balance sheet, with potentially another $100 billion coming after the IPO.

Where the money actually comes from — and where it has to come from next

The harder question the S-1 doesn't fully answer is who, exactly, pays for the trillion-dollar future Anthropic CEO Dario Amodei has talked about. Right now, the two leading labs combined generate roughly $120 billion in annualized revenue, which Brookhart describes as still a narrow market. The top 1 percent of customers are spending $50,000 to $100,000 per employee, and a huge share of Anthropic's business specifically comes from coding — Brookhart called it "almost a single vertical company."

Getting from here to the kind of revenue that justifies a multitrillion-dollar valuation requires breaking into broader white-collar work, moving from AI that assists employees to AI that replaces tasks outright. Brookhart pointed to Microsoft's employee buyout offers as an early signal of that shift happening elsewhere in the industry. But penetration into the Fortune 500 remains tiny in dollar terms: Anthropic counts half of the Fortune 500 as customers, yet that spending represents just 0.1 percent of those companies' total operating expenses, according to figures in the filing. On a per-user basis, that works out to something like $8 a month if you exclude compute costs, or under $50 a month including them. Separately, the top 5 percent of Claude Code users spend around $18,000 a year on average, while typical non-coding users spend well under $75 a month — a gap that underscores how concentrated the current business really is.

The competition is catching up, and it's not just the other big labs

Anthropic had an unusually dominant first half of the year, capturing roughly 90 percent of the industry's net-new annualized revenue. That changed in the third quarter, when Brookhart said OpenAI likely took the majority of new industry revenue instead. Open-source models, meanwhile, have grown from roughly $1 billion in token-inference revenue in the first quarter to more than $10 billion by the third quarter.

Enterprises are also pushing back against sending sensitive data to any single provider. Brookhart noted that some companies, including law firms, are now running AI on their own infrastructure rather than through a cloud API, and many enterprises are fine-tuning open-weight models themselves rather than relying exclusively on frontier labs. Well-funded competitors like Google's DeepMind and Meta Superintelligence Labs have comparable training capacity, and newer entrants, including compute buildouts tied to Elon Musk's ventures, are adding more capacity to the race.

Brookhart's counterargument is that demand for frontier-level intelligence doesn't disappear just because open-source catches up on simpler tasks — it shows up in adversarial or high-stakes fields instead. If an opposing party in litigation or a cyberattacker is using the most capable model available, he argued, you don't want to defend yourself with a weaker one. Anthropic is leaning into that logic by building out industry-specific products for finance, legal, and cybersecurity work, including giving away free usage credits to cyber customers to seed adoption.

The cloud relationships are both the fuel and the risk

Nearly half of Anthropic's business — 48 percent, according to Brookhart — flows through cloud marketplaces like AWS Bedrock and Google's Vertex platform, where customers never directly interact with Anthropic's own sales team. Those cloud providers take a revenue share in return, and Brookhart said that arrangement has been lucrative enough that it's visibly boosting Amazon's and Microsoft's own cloud margins.

But the S-1 also lists a risk that's easy to overlook amid talk of AI consciousness and existential scenarios: Anthropic's biggest compute suppliers — AWS, Azure, and especially Google Cloud — are also its competitors. Google sells its own AI chips to outside customers, including Anthropic's rivals, and has said it doesn't particularly care whether that capacity ends up strengthening a competitor, as long as it gets paid. Anthropic has started buying chips, including Google's TPUs, more directly through special purpose vehicles that then lease the hardware back to the company, rather than renting everything straight from a cloud provider. Brookhart estimated that pipeline at 3.5 to 4 gigawatts and growing. The company's relationship with xAI, Elon Musk's AI venture, carries similar tension, given that xAI is simultaneously a compute partner and a future rival training its own frontier models.

Then there's the matter of the compute Anthropic has already promised to pay for. One estimate put future compute commitments at more than $800 billion over the next decade; Anthropic's own disclosures point to roughly $500 billion to $600 billion over seven to ten years. Brookhart's view is that this isn't as alarming as the raw figure suggests, since current spending levels translate into a relatively modest amount of actual compute capacity relative to where the company is headed. The bigger variable, he said, is how much more revenue Anthropic can squeeze out of each unit of compute as efficiency keeps improving, versus how much of that capacity it chooses to redirect toward training instead of serving customers.

None of this resolves the central question hanging over the filing: whether the diffusion of AI into ordinary white-collar work happens fast enough, and broadly enough, to justify the numbers Wall Street is being asked to underwrite. Brookhart's own read is bullish — he thinks that if Anthropic's AGI thesis plays out, the company could eventually be worth north of $10 trillion, a figure large enough that near-term shareholder dilution from future fundraising wouldn't much matter. That's the bet investors are being asked to make when Anthropic's shares start trading. The S-1 proves the current business works. Whether the much bigger one Anthropic is chasing actually arrives is still, by its own admission, unproven.