The Price War’s First Retreat: DeepSeek Just Raised Prices 355% — and Invented Surge Pricing for AI

For a year, China's weapon was radically cheap AI. Now DeepSeek is charging peak-hour rates like an Uber at rush hour. The monetisation reckoning watchchina.ai warned about all summer has quietly begun

The End of "Almost Free"

For sixty-two briefings, watchchina.ai has documented the single most disruptive feature of Chinese AI: its price. The "good enough at unbeatable prices" strategy, the models given away for free, the 140-trillion-token flood, the "death zone" where survival required either frontier capability or market-breaking cost. But this portal has warned, repeatedly and explicitly, that the diffusion strategy carried an unsolved problem beneath its dominance — monetisation. This week, the company that started China's AI price war made the first visible move to end it.

With the official general-availability launch of its V4-Pro model, DeepSeek announced a fundamental change to how it charges. DeepSeek-V4-Pro output tokens are rising to $3.96 per million at peak hours, up from the previous flat rate of $0.87 per million — an increase of roughly 355%. And more striking than the increase itself is its structure: DeepSeek is introducing peak and off-peak billing, with off-peak rates set at half the peak-hour price. This is surge pricing for artificial intelligence — the Uber-at-rush-hour model, applied to machine intelligence for the first time by a major lab. The company whose radically low prices forced OpenAI to cut its own rates just weeks ago is now, quietly, raising them.

Why the Cheapest Player Is Raising Prices

watchchina.ai has argued since the 140-trillion-token analysis that China's price war was a weapon with a rebound cost — that driving prices toward zero wins market share but destroys the margin needed to sustain the enormous compute bills behind frontier models. DeepSeek's surge-pricing move is the first hard confirmation that the rebound has arrived. Even a state-adjacent, efficiency-obsessed lab like DeepSeek cannot indefinitely serve near-frontier capability — a 1-million-token context window, agent-focused reasoning — at $0.87 per million tokens. The economics simply do not close.

The specific design of the pricing tells the strategic story. By introducing peak and off-peak rates, DeepSeek is doing two things at once. First, it is finally charging what its most valuable capability is worth during the hours when demand — and compute strain — is highest. Second, it is using the off-peak discount to shift less urgent workloads to quieter times, smoothing the load on its infrastructure. This is not a company abandoning the low-cost strategy; it is a company maturing it — segmenting customers by willingness to pay and workload urgency, exactly as watchchina.ai forecast when it predicted China would resolve its monetisation crisis through "enterprise tiers, cloud lock-in, and value-added services" rather than pure giveaways. Notably, DeepSeek is keeping a genuinely cheap option alive: its lighter V4-Flash model performed above expectations in independent tests, in some cases outpacing the April preview of the more expensive V4-Pro. The barbell is forming — cheap Flash for the masses, premium Pro for those who will pay.

What This Reveals About the Real State of the Race

watchchina.ai exists to resist easy narratives, and this story cuts against a comfortable one that has run through much of this summer's coverage — the image of limitless, unbeatably cheap Chinese AI steamrolling all competition. That image was always incomplete, and DeepSeek's price hike proves why. Even at the new peak rate of $3.96 per million output tokens, DeepSeek remains dramatically cheaper than premium American frontier models — the cost advantage watchchina.ai has documented is real and persists. But "dramatically cheaper" is a very different competitive position from "almost free," and the move signals that the era of Chinese AI as a pure loss-leading price bulldozer is ending, replaced by something more sustainable and more conventional: a business trying to make money.

This is, paradoxically, a sign of strength as much as strain, and the honest read holds both. On one hand, it validates watchchina.ai's persistent caveat that China's usage dominance had not yet translated into profit, and that the price war could not run forever — the reckoning is real. On the other, a DeepSeek confident enough to raise prices 355% is a DeepSeek that believes its capability is now valuable enough that customers will pay — the same confidence signal watchchina.ai identified when Beijing began weighing restrictions on its best models. A company only introduces surge pricing when it believes demand is inelastic enough to bear it. The deepest implication connects to the "two SKUs" framing from this series' sixtieth briefing: even China's price frontier is beginning to develop tiers of its own — a peak-priced premium layer, an off-peak discount layer, and a cheap-Flash mass layer. The global AI market is not settling into "American performance versus Chinese price." It is fragmenting into a far more complex structure where even the cheap SKU now has a surge meter running. The price war's first retreat is not a defeat. It is the moment China's AI industry stopped giving the future away and started charging for it — and that, too, is a milestone worth watching.

Sources: Quartz, Reuters, Global Times, Releasebot, DeepSeek API documentation