Anthropic filed to go public yesterday, and buried in 300 pages of SEC-mandated disclosure is the most honest accounting we’ve gotten of what it actually takes to build AGI. The numbers are staggering, but what’s more interesting is what they reveal about the entire industry’s economic model.

The headline figure: Anthropic burned through $2.7 billion last year while bringing in $870 million in revenue. That’s a loss of roughly $1.8 billion, up from $1.4 billion the year before. Revenue tripled, but losses grew faster. This is what frontier AI actually costs.

Here’s the thing everyone keeps dancing around. The economics of training increasingly capable models don’t improve as you scale. They get worse. Each new generation of models requires exponentially more compute, more data, more infrastructure, and more specialized talent. Anthropic spent over $1 billion on compute alone last year. That number will grow.

The company admits this directly in the filing. “Competition for access to electrical supply and data center locations with adequate power availability is intense,” they write, noting their increasing reliance on natural gas to power training runs. They’re not alone. The entire industry is in an arms race where the cost of staying competitive keeps rising and there’s no clear path to profitability at the frontier.

What makes Anthropic’s filing particularly revealing is that they’re one of the best-positioned companies in this race. They have major enterprise customers. More than half of their $100,000-plus customers now use both the Claude API and Claude.ai, which is exactly the kind of cross-selling that creates sticky revenue. They’ve got Amazon and Google as strategic investors and cloud partners. Their models are genuinely good.

And they’re still losing nearly $2 billion a year.

The bull case, which the company makes explicitly, is that AI will eventually be so transformative that today’s losses won’t matter. They’re investing in artificial general intelligence, and AGI will unlock trillions in value. Maybe. But that’s a hell of a bet for public market investors to underwrite.

The more immediate problem is that the path from “good AI assistant” to “AGI” isn’t linear, and it’s not clear that simply spending more money guarantees you get there. OpenAI just announced three “major” discoveries in math, computer science, and physics using their latest models, but they’re also now relying on external “shepherds” to validate the work, which suggests even they aren’t entirely sure what their models are producing is correct.

This gets at the fundamental tension in Anthropic’s business model. They’re selling AI capabilities today while betting the company on capabilities they haven’t achieved yet. The revenue comes from enterprises using Claude for customer service, coding assistance, and document analysis. The losses come from training runs aimed at models that can do… well, everything.

The filing also reveals some genuinely weird governance choices. Dario and Daniela Amodei, the sibling co-founders, have structured the company with a “Long-Term Benefit Trust” that gives them effective control even after going public. The trust is designed to ensure Anthropic can pursue its AI safety mission without quarterly earnings pressure.

That sounds noble, but it also means investors are being asked to fund a multibillion-dollar research project with limited governance rights while the founders maintain control. That’s a tough sell in any market, but especially in a category where “AI safety” can mean anything from “we have good evals” to “we’re preventing human extinction.”

The real tell is in how Anthropic talks about competition. They list OpenAI, Google, Meta, and Mistral as competitors, but they also note that many of these same companies are customers and partners. The industry has created this strange dynamic where everyone is simultaneously competing and collaborating, sharing research through papers while racing to achieve breakthroughs first.

That works when everyone is venture-backed and losses don’t matter. It gets complicated when you have public market investors asking when you’ll be profitable and you have to admit you don’t know.

I don’t think Anthropic is being dishonest here. Their filing is actually refreshingly direct about the risks and uncertainties. But that directness reveals something uncomfortable about the entire frontier AI industry. The economics are predicated on a future that may or may not arrive, and in the meantime, the costs keep rising.

The filing notes that they’re “burning gas to feed the AI boom,” which is both literally true (those data centers run on natural gas) and metaphorically apt. Everyone is burning capital at an extraordinary rate to stay in the race, and the assumption is that the winner takes all, or at least takes enough to justify the spend.

But what if there isn’t a winner-takes-all outcome? What if we end up with several good AI companies that are all unprofitable because the cost of training state-of-the-art models stays permanently higher than what the market will pay? That’s not a crazy scenario. It’s arguably what the S-1 suggests is happening right now.

Anthropic is betting they can grow revenue faster than costs rise, achieve AGI before running out of money, and maintain their safety-focused culture while answering to public markets. That’s a lot of bets to get right simultaneously.

The IPO will probably happen anyway. There’s enough excitement about AI, enough enterprise adoption, and enough FOMO that investors will want in. But everyone buying should understand what they’re actually funding. Not a profitable SaaS business that happens to use AI. A multibillion-dollar research program aimed at building something that doesn’t exist yet, with no guarantee it ever will.

That’s either the investment opportunity of the century or an incredibly expensive way to find out that AGI is harder than anyone thought. Anthropic’s S-1 is honest enough to admit both outcomes are possible. Credit to them for that. But honesty doesn’t make the economics any less brutal.