If API Fares Aren't the Business, What Is
Dwarkesh Patel's blog prize drew more than 600 essays this year on the biggest open questions in AI. A commenter on his own site called the third-place entry "obvious Claude slop." I read it anyway, and the criticism, half fair as it is, misses why the essay is worth sitting with.
Third place, from Michael Li, compares AI labs to Hong Kong's MTR: one of the only transit systems on earth that turns a profit. His point is that MTR doesn't make its money on fares. It makes its money on the shopping malls and residential towers built on top of and around its stations, property MTR itself owns and leases. Li's argument is that AI labs are running the same math without admitting it. The API, priced per token, is the fare. It covers costs and maybe a thin margin, but it will never be the profit engine, because compute costs and competition both trend toward compressing that margin to nothing. What labs actually need to own, in his framing, is the equivalent of the malls: government data-trusteeship deals, forward-deployed integration work embedded inside client operations, and the accumulated reinforcement-learning data that compounds the longer a lab operates, the way land appreciates the longer someone holds it.
It's a clean analogy, and it has a hole in it that the "Claude slop" comment was circling without quite naming. MTR's malls exist because of a specific, documented arrangement: the Hong Kong government granted MTR development rights to the land above and around its stations as part of the original transit franchise, decades before the system opened a single line. MTR's arrangement is a political deal, struck once, with one operator, by one government, for reasons specific to how Hong Kong finances infrastructure. Almost no subway system anywhere else gets that deal by default. Li's essay lists four things labs should own instead: data-trusteeship deals, forward-deployed integration, accumulated RL data, and by implication the market power that comes with all three. Two of those, forward-deployed integration and accumulated RL data, are things a lab can build on its own initiative, no government required. The other one, data-trusteeship access to national health or tax records, is exactly the kind of arrangement MTR got, and the essay never says what would make any government hand that to a single AI lab the way Hong Kong handed land to MTR. The analogy borrows the profitable structure without explaining how anyone gets the deal that created it.
First place took a completely different bet. That essay argued the OpenAI Foundation should spend $40 to $60 billion over the next decade on far-UVC lighting, aimed at ending airborne disease transmission as a category of human suffering. The case for it is a hedge: a use of AI-generated money that pays off in full whether or not the underlying AI business ever justifies its valuation, whether or not the risk case for advanced AI ever materializes, whether or not any lab's API business turns a durable profit. If the money exists, spending it on a fix for a problem as old as indoor ventilation is a bet that doesn't need AI to keep being right about itself.
Put the two essays side by side and the disagreement is smaller than it looks. Both writers assume the metered API isn't where the story ends. Li wants labs to stay inside AI's orbit and shift the real value into rents that have nothing to do with per-token pricing, the way a railway shifts its real value into rent that has nothing to do with fares. The far-UVC essay wants the money to leave AI's orbit entirely and land somewhere its payoff no longer depends on AI succeeding on its own terms. One treats the API as a subway fare on the way to owning the mall. The other treats it as a subway fare on the way to funding something that isn't a subway.
What's notable is what neither essay spent much time defending: that the product currently being sold, a model answering a prompt at a metered price, is the thing that makes any of this work in ten years. The contest asked entrants to wrestle with AI's biggest open questions, and the two winning answers spent almost no words on the model itself. They spent their words on what a lab does with the money and the position a successful model buys it, once being good at making models stops being the point. That's a strange thing for the two most-rewarded essays in a contest about AI to agree on without saying so directly, and it's the part of this that I keep coming back to more than the transit metaphor or the UV lighting.