Anthropic's IPO Shows the Real Winners of the AI Gold Rush Sell the Shovels

Share

Here is a set of numbers that deserves to sit on your desk for a minute. Anthropic filed to go public last week, and the S-1 tells a weirdly reassuring story on its face. Over sixteen billion dollars in trailing revenue, an operating profit, and customers who stay with astonishing loyalty, reportedly holding on to roughly ninety percent of their spend year over year at even the deepest end of the market. On paper this is a thriving business, not a speculative one. Any founder would kill to describe their company that way.

But the filing also carries a set of giant numbers that mostly got a shrug. Anthropic is on the hook for something like five hundred billion dollars in compute commitments over the next decade, much of it binding and non-cancelable, split among Amazon, Google, Microsoft, Broadcom and a few others, with well over a hundred billion more in future capacity that it expects to keep buying beyond that. It has to pay for much of this whether or not it actually uses it. Put the two sets of numbers next to each other and a strange picture emerges: enormous revenue, growing fast, and yet obligations to buy hardware and cloud capacity that run somewhere between twenty and thirty times that same revenue. The engine of the company appears to be running at full speed, and at the same time it is funneling almost everything it earns back out to the people who supply its inputs.

This is the part worth sitting with, because it upends the usual story about where artificial intelligence creates wealth. The past few years of commentary assumed the value would concentrate at the frontier labs that make the most capable models. They have the talent, the brand, the customers and the distribution, so the logic goes, and therefore they should capture the money. The most eye-catching number in Anthropic's own filing quietly makes a case for the opposite reading. It shows that building at the frontier is so extraordinarily expensive, and its appetite for compute so ravenous, that the margin in the whole enterprise ends up flowing to whoever sells the metal, the chips, the cloud and the power.

None of this is a knock on the lab. They are all trying to win a hard game, and for a while the bets looked like the only sensible move. If you believe intelligence will keep compounding, then renting the largest possible pile of compute is a rational way to defend the lead, even at a punishing cost. I think the S-1 actually earns respect for being honest about those costs instead of hiding them in footnotes. What is harder to defend is the assumption that because the frontier labs get the attention, they will also get the profit. The filing is an unusually clear piece of evidence that we might have the economics backwards.

The irony is that the industries doing the quiet best out of this boom are the least exciting ones. Chip designers that sell to every lab at once, clouds that host everyone's workloads, the utility companies that will eventually power these buildings, none of them have to guess which model will win. When OpenAI, Anthropic and Google all fire their compute budgets at the same handful of vendors, those vendors get paid no matter who finishes first. They sell to the gold rush itself rather than betting on any single miner. For every dollar of revenue the frontier makes, a meaningful share already pledges itself right back to this supply chain before the quarter even closes.

There is a fair objection to all this, and it deserves a real hearing. Revenue is compounding so fast, and retention is so strong, that the fixed commitments might simply get smaller relative to sales over time. If Anthropic can grow into its obligations, turning them into a rounding error by the early thirties, then the compute bill becomes a footnote and the current valuation looks cheap. Consumption revenue is the kind that scales with usage rather than shrinking, and the filing stresses how deeply its largest customers are embedded. That is a genuine possibility, and it is the bull case in its strongest form.

But I suspect the more durable lesson is the structural one, and it is the same lesson the first railroad boom taught. The men who actually got rich in the early railroad era were often not the railroad operators, who spent their lives fighting for routes and traffic and often ended up wiped out. They were the people who sold the operators everything they needed to run: the steel, the land, the financing, the labor. The operators competed with each other to the bone, driving down their own margins, while their suppliers were paid regardless of who won. Watching the AI build-out, I keep coming back to that pattern. When the essential input is scarce and every competitor must buy it, the seller has the leverage and the buyers have the race.

That insight does not only apply to a giant like Anthropic. It is arguably sharper for a startup, because a smaller company has less negotiating power and fewer places to turn. If the cost of intelligence itself keeps falling, and it is, then the margins have to come from somewhere other than just being big or first. They come from owning something a competitor cannot simply rent: distribution, a deep customer relationship, proprietary data, a workflow nobody else has bothered to build, the trust that makes a customer reluctant to leave. That is where I think the durable value of this technology is actually going to land over the next decade.

The other thing the S-1 clarifies is why the politics of frontier AI get so intense. When the five biggest players are collectively committing trillions to the same infrastructure, the winners are not just competing on the quality of their models. They are competing on access to compute, to capital and to power, and that is a very different game than the one where a small team with good ideas can disrupt a giant. It raises a real question about concentration that nobody in the funding frenzy seems to want to answer directly: if the value keeps flowing to the people who sell compute, then the race on top of it is being run under terms set almost entirely by someone else.

I do not want to overstate the case, because there is a healthy counterpoint hiding in the same filing. Anthropic is not merely a well funded research lab burning cash. It is generating real revenue with strong retention, and that is rarer in this industry than coverage suggests. A company that can grow revenue at this pace, even while buying an enormous amount of compute, has something real. The question is not whether the frontier labs can become good businesses. It is how much of the economic upside of their own success they end up keeping, once the bill for all those shovels comes due.

What interests me is how thoroughly the story flipped from a year or two ago. Back then the obvious way to bet on AI was to bet on the models, the labs and the brands. Today, the most reliable way to bet on it might be to own the thing every model needs, or to build a business whose value comes from the customer relationship rather than from renting the frontier. The gold rush narrative has always been about the lucky miner who strikes it rich. The history kept getting written by the people selling the pans.

The practical truth is that the AI build-out is not a debate about technology anymore. It is an enormous transfer of capital, decided by whoever controls the scarce input everyone must buy. I suspect that is where the next decade's fortunes get made, and the S-1 is one of the clearest invitations yet to look at exactly who that is. It is tempting to watch the race and cheer for a favorite. The smarter observation might be to notice who sells tickets to all of them, and to consider that the person collecting at the door may end up with the better business.