DeepSeek just put surge pricing on intelligence. That settles the argument about who owns AI.
The company that proved AI could be cheap now charges peak rates like an airline. Follow the compute: the meter was always the power, not the model.

Image: RoadMaster19 / Wikimedia Commons, CC BY 4.0
For a year and a half, DeepSeek was the counter-argument. When a Chinese lab shipped a frontier-grade model in early 2025 at a fraction of the going rate, gave the weights away, and briefly wiped a trillion dollars off the American chip trade, it was received as proof of a comforting thesis: that artificial intelligence would be cheap, abundant, and therefore beyond the control of whoever owned the most compute. This week DeepSeek quietly retracted that thesis. It did not issue a statement or reverse a model release. It changed its prices — and the shape of the change tells you more about where power actually sits in this industry than any keynote this year.
As of Saturday, DeepSeek charges by the clock. Its V4 Pro model, which listed at $0.87 per million output tokens, now costs $3.96 during peak hours and $1.98 off-peak — a peak rate roughly four and a half times what it was. V4 Flash moved the same way, from $0.28 to $1.32 at the peak. The company had promised in May to make its discounted pricing permanent; it has now un-promised that, explaining that dual pricing lets it 'allocate resources more reasonably' between busy and quiet hours. Read that sentence again, because it is the entire story compressed into corporate anesthetic. Allocating a resource by time of day is what you do when the resource is scarce.
The mechanism is the message
The number that matters is not the size of the increase; it is the mechanism. Surge pricing — peak and off-peak, the instrument an airline uses for seats and a utility uses for electricity on the hottest afternoon of the year — is an admission about the nature of the good. It says the constraint was never really the cleverness of the model, which DeepSeek proved could be made cheap. The constraint is the compute behind the model, and how many people want it at two o'clock in the afternoon. A model is software; it costs almost nothing to copy. The chips that serve it are physical, finite and contested, and their availability at the moment of peak demand is the actual product being priced. DeepSeek has simply started charging for the thing that was always scarce and that it spent a year letting people treat as free.
Allocating a resource by the hour is what you do when the resource is scarce. Intelligence didn't get more expensive. The compute behind it stopped being hidden.
The tell is not that DeepSeek raised prices — every company reserves that right. The tell is that it reached for time-of-day metering, the pricing psychology of a system running near its physical limit. Software companies do not surge-price, because serving one more customer costs them next to nothing. Utilities surge-price, because at the peak there is a hard ceiling and someone must be rationed by price. When the cheapest, most ideologically open lab in the field adopts the tariff structure of a power grid in a heatwave, it is telling you which kind of business inference has been all along. Not the software business it was sold as. The capacity business its skeptics said it was.
The strongest version of the other side
Let me put the best case against my own, because it nearly moved me, and I have reversed myself in this column about precisely this before. Dynamic pricing is not villainy; it is arguably good engineering and even good citizenship. Charging more at the peak flattens demand, protects latency for everyone, and pushes the flexible work — the overnight batch jobs, the patient agents — into the quiet hours when capacity would otherwise sit idle. And the absolute numbers remain remarkable: DeepSeek at its new peak rate is still a small fraction of what the Western frontier charges, where a comparable OpenAI model runs many times higher and Moonshot's Kimi sits somewhere in between. Cheap AI is not dead. On this reading, DeepSeek is just a maturing company pricing a genuine constraint honestly, and the alarm is misplaced.
I take that seriously, in part because I have been wrong in the optimistic direction before. I argued, more than once, that open weights and cheap challengers were a structural check on the concentration of power in AI — that as long as a scrappy lab could give a capable model away, nobody could truly own the field. I was too sanguine, and I was watching the wrong layer. I was watching the model, which is free, and not the inference, which is not. Surge pricing is what made me move, because it exposes where the leverage actually lives.
Cheap at 3 a.m. is still price-taking
Here is the answer to the good counter-argument. Cheap at off-peak is still price-taking. The user does not set the terms; the owner of the capacity does, and 'you may have it cheaply at three in the morning' is a term dictated to you, not a freedom you hold. Open weights were supposed to be the escape hatch — own the model, run it yourself, owe no one. But a weight you cannot afford to serve at scale is a car with no fuel. The moment your workload is real and time-sensitive, you are back to renting GPUs at whatever the peak commands, from whoever has them to rent. The leverage moved. It moved off the model, which anyone can now download, and onto the metered capacity to run it, which only a few can supply. The democratization story always depended, quietly, on someone else absorbing the capital cost of that capacity. Surge pricing is that someone deciding to stop.
Which relocates the power question, and then answers it. If inference is priced like a utility at the peak, the decisive advantage belongs to whoever owns the most idle capacity — the player who can absorb the two o'clock spike without surging, and who therefore sets the ceiling that everyone else's prices press up against. That is not the insurgent who gives models away. It is the incumbents with the largest fleets of chips and the longest power contracts, the compute barons this column keeps returning to, because the compute keeps turning out to be the company. DeepSeek did not break that structure by being cheap. It has now joined it, by learning to meter the peak the way anyone who owns capacity eventually must.
So take the price change as the honest disclosure it is. For eighteen months the DeepSeek moment let a great many people believe the scarcity had been engineered away — that intelligence would be a commodity too abundant to be owned. The surge tariff is the correction. Intelligence is abundant; the capacity to serve it at the instant everyone wants it is not, and those two things were never the same layer. Whoever controls the second layer sets the terms, and can always, in the end, charge for the peak. The model was the narration. The meter is the power. DeepSeek just told you which one it finally decided to run.


