You Can’t Sell a Fourth Meal

AI is a supply-side technology. The developed world’s problem is demand. That mismatch — not the technology — is the whole story.

I’m not down on AI — I think it’s one of the most important tools of our lifetime. But I want to be precise about where its value actually lives, because the trillions being spent are, I’d argue, largely pointed at the wrong place.

The core idea is this: AI is a supply-side technology — it makes one thing, cognition, abundant and cheap. Whether that grows an economy depends entirely on whether cognition was the constraint holding that economy back. In much of the developed world, it isn’t. The constraint is demand.

You cannot out-think a potato. Its limits are soil, water, sun, and season — none of which yield to a smarter thought. Spend millions optimizing how you grow the same potato the same way, and you won’t harvest billions in extra potatoes, because intelligence was never the scarce input.

That’s only true if the method stays fixed. Change the method — gene-edited cultivars, vertical farming — and cognition comes right back, as the constraint on discovery. The Green Revolution wasn’t more sun; it was Borlaug’s knowledge, and it fed a billion people. So the real test is this: AI creates value where the bottleneck is discovering something new, and almost none where the bottleneck is doing a task we already understand. Most enterprise AI is aimed at the second. That’s why it hits a ceiling fast.

The wall nobody in the capex debate talks about is demand, not supply. Everyone’s arguing whether the models will get good enough. The harder question: even if they work perfectly, what new demand do they unlock? Better marketing can’t sell you a fourth meal when you eat three. The developed world already produces more than it can sell. Pour a supply miracle into that and you don’t get growth — you get deflation and reshuffling. The pie stays the same size.

I won’t overclaim: “wants are satiated” is false. Demand always migrates — from food to cars to phones to experiences. But for AI specifically, the new demand it’s meant to unlock hasn’t been named yet, and betting trillions on an unnamed appetite is the risky part.

You don’t have to take it on faith, because the industry that adopted AI hardest — tech itself — is now pulling back. Developers “tokenmaxxed,” burning billable tokens on internal leaderboards that rewarded usage over results. Then the bills landed. Uber blew through its annual AI budget in four months and started rationing it. Lindy, a startup built on falling token costs, ripped out its models overnight to survive. And MIT found ~95% of enterprise AI pilots produced no measurable profit — not because the models failed, but because the value had nowhere to land.

The fair rebuttal is that this is just the messy middle of any adoption — waste, then correction, then maturity. Fair. Retrenchment isn’t proof of a permanent ceiling. But it does puncture the claim that today’s spending is justified by today’s returns. By the adopters’ own math, right now, it mostly isn’t.

Software is the instructive case, because it’s where the money went. Coding looks like a thinking problem — which is why it drew the heaviest spend — but most of it isn’t discovery, it’s acceleration: writing known patterns faster. That’s the doing side of the line, and it hits the potato ceiling, which is exactly why the bills ballooned without the returns. There’s a real discovery frontier in software — novel algorithms, verification, the non-obvious optimization — but it’s narrow. The rest is very expensive typing.

The same logic that finds the ceiling also finds the gold. AI pays off exactly where cognition was the scarce, rationed input. Two places stand out.

First, anything genuinely gated by thinking: drug discovery, materials science, protein and chip design, research, cybersecurity. One AI cybersecurity company built an autonomous system that became the first AI to top HackerOne’s US leaderboard, outranking the human researchers it was up against. Here the search space itself is the bottleneck, and that’s precisely what AI attacks. These aren’t cheaper potatoes; they’re potatoes we couldn’t grow before.

Second — and this excites me most — the underserved and developing world. The fourth-meal problem only exists for the already-full. Much of the world is on one meal. What stands between people and a doctor, a tutor, a lawyer, an accountant isn’t physics or saturation — it’s that expert cognition is scarce and unaffordable. There, AI isn’t a fourth meal nobody wants. It’s a first meal that never existed at any price. That’s where the landscape changes fastest, and where the real human gains will be.

The irony is that’s not who’s writing the checks. The trillions are a developed-world efficiency bet — aimed where the tool fits worst — while the places it fits best have the least capital to deploy it.

So the whole game is matching the tool to the constraint. Point cheap cognition at a saturated economy and you get tokenmaxxing and rationed budgets. Point it where thinking was actually scarce — a hard scientific problem, or someone who never had access to expertise at all — and you get the transformation the hype promised, just not where the hype is looking.

You can’t sell a fourth meal to someone who’s already full. But billions are still waiting on their first — and there are countless problems where thinking really was the thing we never had enough of. That’s where this technology changes the world, and we should point it there. Which is why the debate we’re having is too small. This isn’t really about US versus China, or closed frontier labs versus open-source Chinese models. It’s about aiming the most powerful supply of cognition ever built at the problems where cognition was the thing we never had enough of. Get that right, and who wins the race matters far less than what the race was for.