DeepSeek spent eighteen months as the AI industry's standing argument that frontier intelligence ought to be nearly free. At 16:00 UTC on August 16, that argument got repriced. Output tokens on DeepSeek-V4-Pro, which had cost a flat $0.87 per million, now run $3.96 per million during peak hours — a 355% increase — and $1.98 off-peak, still more than double the old rate. Input tokens on a cache miss went from $0.435 to $1.32 at peak. On cached inputs, the mechanism that made DeepSeek genuinely cheap rather than merely cheap-looking, some line items rose by as much as 1,100%.
Three days earlier, DeepSeek had shipped V4-Pro out of a preview that had run since April 24, promoting the V4-Pro-0813 checkpoint to general availability alongside reasoning-effort tiers and a native OpenAI Responses API. And two weeks before that, on August 6, Bloomberg reported that the Hangzhou lab had restarted a funding round seeking close to $8 billion at a valuation of roughly $74 billion.
One correction to the record, because it changes the story. This is not DeepSeek's first big outside round. It is its second in roughly ten weeks. DeepSeek closed a maiden external round of about 50 billion yuan ($7.4 billion) at the end of May, at a $52 billion valuation, with Tencent putting in 10 billion yuan, CATL 5 billion, and NetEase, JD.com and China's National AI Industry Investment Fund filling out a syndicate of fewer than ten investors. Liang Wenfeng himself contributed 20 billion yuan, and the capital was routed through a limited partnership vehicle he manages rather than into voting shares — a structure designed to keep founder control intact. No foreign institutional money was accepted.
The second round is younger and messier. Preliminary talks surfaced July 14 at a pre-money valuation near $71 billion. On July 25, DeepSeek verbally told prospective backers the signings were off, days after a transcript of a May 20 closed-door investor meeting — in which Liang reportedly said China's lag behind the US is “essentially a compute gap,” that DeepSeek runs on roughly one-twentieth the compute of leading US labs, and that it remains dependent on Nvidia — was posted to GitHub on July 22 and went viral. Bloomberg reported the round back on August 6, with Moonshot backer Monolith Management in talks to join a syndicate now including IDG Capital, Loyal Valley and Shixiang Capital. The amount, timing and investor list can all still move. Nothing is signed.
The economics behind the clock
DeepSeek's own explanation for the increase is capacity, not margin. The company told users it planned “a significant increase” and asked them to plan accordingly, framing the peak/off-peak split as a way to “allocate resources more reasonably.” Seventeen of every twenty-four hours remain at the half-price off-peak rate; the peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC, land squarely on the Chinese working day, which means the schedule hits DeepSeek's home market hardest and leaves most Western buyers paying off-peak.
Analysts reading the schedule found the headline percentages misleading in both directions. “On paper, at peak, against the right comparator, DeepSeek's price advantage does disappear, and in places inverts,” said Sanchit Vir Gogia, chief analyst at Greyhound Research, but in practice “the schedule's own clock and cache hand most of it back to any buyer paying attention.” Gogia's sharper point is about the cache: DeepSeek's roughly 98% cache-hit discount, against an industry norm nearer 90%, is what kept its measured cost per task about 60% below OpenAI's GPT-5.6 Luna even after Luna's own 80% price cut. “The cache is where the advantage genuinely erodes,” he said. Flash's edge over Luna narrows from roughly sevenfold to threefold off-peak, and to 1.4x at peak.
Mark Tauschek, VP of research fellowships and distinguished analyst at Info-Tech Research Group, was blunter about the cause. “This isn't unexpected at all,” he said. “It's simple supply and demand: when demand goes up, pricing goes up, because supply becomes constrained.” He noted Anthropic raised prices for the same reason in April, and that DeepSeek V4-Pro still undercuts GPT-5.6 Terra and Sol even at peak.
Why It Matters
DeepSeek is the only major AI lab of its size with no cloud business underneath it. Alibaba, Tencent, Google and Microsoft can price tokens at or below cost because inference is a customer-acquisition line item for something larger. DeepSeek's annualized revenue, reported in mid-July at $400 million to $500 million, comes overwhelmingly from API sales. There is nothing to cross-subsidize with, and there is now an investor base — plus a planned Inner Mongolia data center expected to cost billions, plus a 140.8 million yuan ($20.8 million) stake in robot maker Unitree — that expects the unit economics to work.
The price war DeepSeek started in January 2025 was, in part, a function of having no shareholders. Taking $7.4 billion in May and asking for $8 billion more in August changes what the company is optimizing for. Gogia's warning is the one worth keeping: CIOs should read the new schedule with “relief and unease,” he said — unease because “a supplier that has learned to price the clock has learned something about its own leverage.”
What to Watch
Whether the round actually closes, and at what number. A $74 billion mark against a $52 billion valuation set in May implies a 42% step-up in ten weeks on roughly half a billion dollars of annualized revenue — a multiple that assumes the price increase sticks and volume does not flee to Qwen, Kimi or third-party hosts of DeepSeek's own open weights. Watch, too, for whether Chinese developers actually reschedule workloads into off-peak hours or simply route elsewhere, and whether the onshore IPO that every Chinese AI lab is now racing toward gives Beijing a say in how the round is priced.
“On paper, at peak, against the right comparator, DeepSeek's price advantage does disappear, and in places inverts.”— Sanchit Vir Gogia, Chief Analyst, Greyhound Research