Leaderbored

Is it a race or a run?

From a conversation with a smart friend,¹ a fun theory:

1. To do AI, you need a bunch of very fancy computers. However, because lots of people want to do lots of AI, there is a shortage of these computers. In a few years, we might have enough of them—we are putting data centers in Texas and in Tennessee; we are putting data centers in the ocean; we are putting data centers in orbit, on the moon, we might put data centers on Mars; can we borrow your computer to use it as a data center?—but, until then, we are short. Not everyone who wants a computer can have one.

2. This is even true—especially true?—at a handful of huge companies, like OpenAI, Google, Anthropic, Meta, and SpaceX-X-xAI.² Though they have a lot of fancy computers, they also want to do the most AI, and still have to make hard choices about how to use their computers.

3. Specifically, they have to choose if they want to use their computers to build AI or sell AI. Training new frontier models is how you build the smartest AI, and customers want the smartest AI. But offering people a way to use that model—that is, using the model you trained to provide inference, via websites like ChatGPT or programmatic APIs—is how you make money. The model is your secret recipe; inference is the food on the plate. Customers don’t want to eat bad food; customers can’t eat a recipe. You have to invent a good recipe, and use it to cook the food.

4. But, the computers. You do not have enough computers to do both.

5. So, you make a plan: “First, we’ll use most of our computers to build the smartest AI. It will cost a zillion dollars, but that’s ok, because once we have the best AI, we will aim to achieve turnaround.” Sow today, reap tomorrow, a classic business move.

6. Sometimes, it doesn’t work. You spend a zillion dollars, and do not build the best AI. Ah, awkward, an A for effort, I guess.

7. Other times, it does work! You spend a zillion dollars, and build the best AI. Everyone loves your AI. They love the products that use your AI. They want to stop using that company’s AI and use this company’s AI. They want to run your AI everywhere, all the time. This is great; this is exactly what you build the best AI for.

8. But, the computers! There is so much demand for your AI, and you only have so many computers. So you have a choice: Do you capture the moment, and use some of your training computers as inference computers? Or do you let it slip, and not sell the thing that you just spent a zillion dollars making?

9. I mean. You probably sell the AI.

10. Now, though, you have fewer computers to train an even better AI. And other companies—which no longer have the best AI—might get to thinking: “The company with the best AI is using a lot more computers selling AI. What if, while they’re doing that, we use most of our computers to build an even smarter AI? It will cost a zillion dollars, but that’s ok, because once we have the best AI…”

11. Round in round, round after round.

Yes, sure, this cuts a lot of corners. Good AI isn’t just built by computers; it’s also built by good researchers, unique training data, and, increasingly, the product around the AI. AI companies can do many things at once; their computers are not so easily fungible; many of their computers are actually, famously, also their competitors’ computers, leased back and forth in a giant trillion-dollar web of incomprehensible partnerships and frenemy-ships. There is no knob that says “build” on one side and “sell” on the other.

Still, that is the pattern, isn’t it? In 2023 and 2024, OpenAI was the dominant default, Anthropic was the IYKYK upstart, and Google was incinerating its reputation, and many billions of dollars, one disastrous demo at a time. Then, Google built the best new models—“Gemini 2.5 is the new SoTA;” “I tested Google’s new Nano Banana image AI, and it’s insane”—and became the smart money. Then, Anthropic put out Opus 4.5; it was a “world-changing shift;” everyone became obsessed with Claude Code; “move over, ChatGPT.” Anthropic blew past OpenAI’s hype, its revenue, and its valuation; OpenAI started missing revenue targets. But revenue is a lagging indicator. Over the last six months, while Anthropic was selling Opus and Claude Code—and struggling to find enough computers to do so—OpenAI was building GPT 5.5, out last week. And GPT 5.5 has it all; GPT 5.5 blew people away; GPT 5.5 is a new frontier in everything; cracked devs on X are abandoning Claude Code and coding on Codex.

And OpenAI is an inference company now. Build phenomenal cosmic power; get everything that goes with it. Round and round, round after round.

The implication, I suppose, is that this old post was partially right: AI labs are bad businesses. Models really are fast-evaporating moats, clever employees really are two weeks from leaving, and every billion-dollar model is one competitive release away from being leapfrogged into near-immediate obsolescence. But, this comparison was wrong:

There is, however, one enormous difference [between AI labs and cloud providers like AWS]: You can’t build a cloud vendor overnight. Azure doesn’t have to worry about a few executives leaving and building a worldwide network of data centers in 18 months. AWS is an internet business, but it dug its competitive moat in the physical world. …

What, then, is an LLM vendor’s moat? Brand? Inertia? A better set of applications built on top of their core models? An ever-growing bonfire of cash that keeps its models a nose ahead of a hundred competitors?

The problem is the premise. AI companies are actually no different than cloud providers: They build models, but they sell computers. And their lasting moat—their only lasting moat—is their computers. Models converge; talent migrates. Computers endure.

The other implication—the more useful one, maybe—is that AI whiplash is structural. The vertigo is systemic. The hype cycle is scripted theater. In science fiction stories, the AI company that builds the best model has a compounding advantage over its competitors. Someone builds the smartest model; that model helps its developers improve itself faster than the second-smartest model; a small gap becomes a big gap; a big gap becomes an insurmountable gap. There is urgency in small differences, because you never know when pulling ahead could explode into pulling away.

Perhaps that is backwards. Perhaps the AI race is less of a space race and more of a bike race: The headwinds are borne by whoever is in the front. Perhaps it’s Mario Kart: There are hardcoded advantages to being behind. Perhaps there is an invisible asymptote to every AI company’s lead: The further ahead your AI gets, the more people want to buy it, and the less new AI you can build. Perhaps it’s not a race at all, but a ___ run.³

Spectating from the dizzying center of the arena, maybe there is something steadying in that. When the peloton moves together, does it matter who’s winning right now? When no lead is safe, should we bother worrying about the leaderboard?

1. I have a friend with a two-letter last name that’s also a top-level domain. Their personal website is first-name-dot-last-name, which is, as far as personal websites go, obviously perfect. Are any others as good? Arguably, first-name-dot-com, because, man, what a flex. Then, there are a million miles between those two domains and whatever is next best.

Or, not! Because calv.info—the initialing, the efficiency, the .info for a site about with info about you—might be the third perfect domain.

2. XaiXspaceX? spaceXXXai? space.xxx.ai?

3. I do not know what to put in this blank. It’s this kind of run, high school coaches used this name, and it doesn’t seem like we should call it that?