Dwarkesh Patel's New Essay Says GPU Rental Prices Could Go 15x. Do This Math Before You Trust Your Own Margin.
Podcaster and AI writer Dwarkesh Patel published an essay in late July arguing that GPU rental prices are set up to rise, not fall, as models get better at monetizing the compute they run on. His math: if a model running on an H100-equivalent can do roughly $250,000 a year of human-level software-engineering work, competitive bidding should push that GPU's rental price toward what it produces (somewhere around 15x today's spot rate), not toward the cost of manufacturing the chip. If you're pricing a product against current API rates and assuming they trend down the way they mostly have through 2026, this essay is worth 20 minutes, because it argues the opposite could just as easily happen.
The argument, stripped down
Today's GPU rental prices are largely set by supply and demand for raw compute capacity: how many H100-equivalents exist, how many people want to rent them, what the going hourly rate is. Patel's point is that this pricing model breaks down once a model running on that hardware becomes valuable enough on its own. If an H100 running a sufficiently capable model can do the work of a $250K/year software engineer, the GPU is worth what it produces, not what it costs to build. Competitive bidding among labs and companies that need that output should push the rental price up toward the value generated, not stay pinned to manufacturing cost, which by Patel's estimate implies something like a 15x jump from current spot prices.
Patel adds a specific mechanism for why this could accelerate rather than plateau: the Alchian-Allen effect, an economic pattern where adding a fixed cost to two goods of different quality makes the more expensive, higher-quality good relatively cheaper to use. Applied here: once GPU-hours themselves get expensive, running a weak, inefficient model on an expensive GPU-hour becomes wasteful in a way it isn't today, and demand shifts toward frontier models even though they cost more per token. That's the opposite of the "AI keeps getting commoditized and cheaper" intuition most solo builders have absorbed from two years of falling token prices, and it's the part of the essay worth sitting with longest.
Why the direction of this argument matters more than the exact multiple
I want to be careful here: 15x is Patel's estimate from a back-of-envelope calculation in a single essay, not a market forecast anyone should treat as precise. The value of the argument isn't the specific number: it's the mechanism. Nearly every solo AI-wrapper roadmap I've seen this year assumes token prices are a one-way ratchet downward, because that's what's happened for two years running: open-weight models from Chinese labs undercutting on price, competitive pressure between OpenAI, Anthropic, and Google compressing margins, free tiers expanding. Patel's essay is a structural reason that trend could reverse once frontier labs figure out how to actually monetize what their models produce, rather than giving away compute below its value to win market share.
That reversal hasn't happened yet, and there's no guarantee it will on any particular timeline. But "hasn't happened yet" is exactly the condition under which it's cheap to prepare and expensive to be caught flat-footed if it does.
The concrete exercise: run your numbers at 5x and 15x
Here's what I'd actually do with this essay, rather than just filing it as an interesting read. Take your current per-user or per-task AI cost (whatever you're paying today in tokens to deliver your product) and multiply it by 5x. Then by 15x. Look at what happens to your margin at each number.
If your pricing survives 5x with a haircut but still works, you have a normal margin, and this is a useful stress test you can file away. If your pricing doesn't survive 5x (if a 5x cost increase turns your product from profitable to underwater), you don't actually have a durable margin right now. You have a margin that exists because of a temporary price war between labs racing to acquire market share below cost, and that's a materially different thing to build a business on. A lot of thin-wrapper AI products fall into this second category without their builders having run the number explicitly, because the current prices feel stable enough not to question.
If you're in that second group, the fix isn't complicated, just uncomfortable: raise your prices now, while you have room to, rather than waiting until a cost shock forces the conversation with customers who've gotten used to the old price. Or build in enough architectural flexibility (model routing, caching, the ability to fall back to a cheaper model for lower-stakes requests) that a price shock on the frontier tier doesn't take out your whole product.
The honest counter-take
This is one economist-adjacent podcaster's essay, built on a specific and disputable assumption: that frontier labs can actually monetize compute at the rate Patel's math implies. AI monetization has consistently lagged behind AI capability throughout this entire cycle: models keep getting more capable faster than anyone figures out how to charge for that capability at its theoretical value. If that pattern holds, rental prices don't have to move at all, and Patel's scenario stays a thought experiment indefinitely. It's also worth noting that a 15x price increase would hit every AI-dependent business simultaneously, including the labs' own biggest enterprise customers, which creates real political and competitive pressure against it happening cleanly or quickly.
I don't think this essay is a prediction you should bet your roadmap on. I do think it's a well-reasoned argument for why the floor under current AI pricing is less solid than two years of falling prices has trained everyone to assume, and stress-testing your margin against a scenario that might not happen costs you an afternoon, while being unprepared for one that does happen costs you the business.
Author
Lukas
@lukcombinator