Your AI Budget Assumes Prices Keep Falling. They Might Not.

Dwarkesh Patel makes his living interviewing the people who build and finance frontier AI. On July 29 he published an argument that runs against the one thing nearly every marketing plan takes for granted: that the price of using these models falls every year.

His case starts with a mismatch in growth rates. Compute at the labs, he writes, roughly triples each year, while revenue at the leading labs tracks something closer to a tenfold climb.

Compute here just means the rented computers that run the model when you press the button. Supply of those computers grows about 3x a year. Demand for what they produce grows far faster. When that happens to any other input, the price goes up.

Then he prices the ceiling.

“If a true human-level software engineer that could run on an H100 equivalent, at current market rates for software engineers, that H100 should rent for over $250k a year. That’s 15x today’s spot prices.”

You don’t have to accept the human-level premise for the mechanism to bite. The argument is that the value a chip can produce, not the cost of making the chip, sets what it rents for. Then comes the sentence marketers should read twice. As models get better at high-value work, Patel writes, “a lot of current popular applications of AI get priced out.” His example is short-form video, generated at volume, for cheap.

That is a description of a marketing department.

What this collides with

In July we covered OpenAI’s latest price cut and argued that cheaper per token isn’t cheaper in practice, because consumption climbs to meet the discount. We also covered open-weight models you can run yourself as a hedge against vendor pricing. Both pieces assumed the list price keeps drifting down and the fight is over volume. Patel is arguing that the list price itself has a floor under it, set by what else that computer could be doing instead.

The counterweight is real and worth stating. Chips get faster and cheaper to manufacture, competition adds supply, and smaller models keep getting good enough for routine work. Patel isn’t forecasting permanently expensive computers. He’s describing a squeeze in the middle, before manufacturing catches up.

But a squeeze that lands in 2027 lands on budgets being written now.

The bill nobody line-items

Ask operators what AI costs them and the token invoice isn’t the interesting number. Daniel Ruke runs blink, a Florida creative studio whose work has touched properties from Disney, Marvel, and Epic Games. He spends about $5,000 a month on AI tools and tokens.

“The real bill is hours,” Ruke said. “If my token bill is $5K, the labor wrapped around correcting the machine is worth multiples of that.”

“Nobody budgets for it, because it never shows up as a line item. It shows up as your afternoon.”

That’s the position most marketing teams are in before any price increase. The metered cost is small and visible. The correction cost is large and invisible, which is the same finding we hit reporting on the maintenance tax on marketing agents. Raise the metered cost 3x and the invisible half doesn’t shrink to compensate. It grows, because the cheap fallback models you’d switch to need more correcting.

Jon Winsett, founder and CEO of the Atlanta procurement firm NPI, watches this from the buying side.

“The nature of agentic AI is that it compounds on itself,” said Winsett, who called the current trajectory of spend unsustainable. “There’s going to have to be a reckoning.”

Gartner sees it landing in software prices rather than API bills.

“The cost of software is going up and both the cost of features and functionality is going up as well thanks to GenAI,” said John-David Lovelock, distinguished VP analyst at Gartner.

That’s the transmission mechanism for most marketers. You don’t rent an H100. You buy seats in tools whose vendors do, and their margin has to come from somewhere.

What’s already changed

The finance side has moved faster than marketing has. In the FinOps Foundation’s 2026 survey of 1,192 practitioners, 98% now manage AI spend, up from 63% a year earlier. 58% named AI cost management as the most wanted skill on their teams. Cloud finance teams built a discipline around a cost line that marketing still treats as a subscription.

Three things worth doing while the question is open.

1. Find out what your tools pay for compute. Any vendor pricing per seat while paying per token has a gap to close. Ask what happens to your price if their model costs double, and treat a non-answer as the answer.
2. Know which of your workflows is the short-form video slop. Every team has one high-volume, low-value generation job running because it’s currently almost free. Price it at 3x. If it dies at 3x, it was never a workflow, it was a discount.
3. Track corrected output, not output. Ruke’s point is the measurable one. If you can’t say how many hours your team spends fixing generated work, you can’t tell whether a price rise is survivable. You don’t know what you’re paying now.

Patel may turn out to be wrong about the size. The direction is the part to plan against. And nobody selling you an AI tool has told you what happens to your price when their costs go up. That silence is the forecast.