For developers using the DeepSeek API, few things are as troublesome as a sudden price hike. On August 17, 2026, DeepSeek overhauled its entire pricing structure, and depending on the category, the new rates reached up to 12.1 times the old ones. But beyond the sheer scale of the increase, what catches the eye is that the newly introduced peak-hour window aligns precisely with standard Beijing office hours: 9 a.m. to noon and 2 p.m. to 6 p.m. This alignment—something you wouldn't notice just by glancing at a price table—reveals where DeepSeek's actual heavy users are located. The same pattern is visible, in a different form, in Alibaba's Qwen announcement of "3 billion downloads in 6 months."
A New Pricing Structure Where Only Beijing Time Becomes Peak Hours
On August 17, 2026, DeepSeek overhauled its API pricing structure and simultaneously introduced a peak-hour billing system. The peak windows are two time slots in Beijing time—9 a.m. to noon and 2 p.m. to 6 p.m.—which translate to 10 a.m. to 1 p.m. and 3 p.m. to 7 p.m. in Japan time. According to the official price list, the price difference between peak and off-peak hours is a uniform 2x across every category. For the flagship model V4-Pro, output tokens rise exactly 2x from $1.98 off-peak to $3.96 peak per million tokens, and input tokens (cache hit) rise exactly 2x from $0.022 to $0.044.
What shouldn't be overlooked in this revision is that the pricing structure itself changed, not just the intraday fluctuation. For V4-Pro, input tokens (cache hit) rose up to 12.1x from the old pre-August-17 price to the new peak price, input tokens (cache miss) rose 3.0x, and output tokens rose 4.5x. While the peak/off-peak intraday difference follows a simple uniform 2x rule, the magnitude of the increase from the old rates varies significantly by category. The eye-catching "up to 12x" figure that many people focus on refers to the price gap before and after the revision, and is a separate metric from the intraday time-slot difference (a uniform 2x).
The price hike itself can be explained as a mechanism for efficiently allocating computing resources during peak times. For developers who use the DeepSeek API on a daily basis, this design makes it harder to predict billing amounts, but it also serves as a rare external clue for inferring which time slots see concentrated server load at the company.
It's common for AI model usage fees to have separate unit prices for "input tokens"—the volume of instructions sent to the model—and "output tokens"—the volume of the response the model returns. Input tokens are incurred based only on the amount of the question text or attached data, while output tokens scale with the length of the generated response. Since DeepSeek's peak pricing applies a uniform 2x weighting to both input and output, users who feed large volumes of documents or code into the model all at once will see a correspondingly larger absolute increase in their bill during peak hours.
The practice of varying cloud computing prices by time of day is nothing new. Major cloud providers' spot instances also use a system where prices rise during high-demand periods and drop during low-demand periods. Since there's an upper limit on the number of GPUs that can be secured in advance, raising prices during periods of concentrated demand makes rational sense as a way to distribute usage and prevent server overload. However, what tends to be overlooked is that the information about which time slots see concentrated demand is itself a mirror reflecting that service's user base.
What the Peak Window—Which Even Reproduces the Lunch Break—Tells Us About the User Base
Lining up the peak windows along a timeline reveals that the 7 hours are split into two blocks. Only the 3-hour block from 9 a.m. to noon and the 4-hour block from 2 p.m. to 6 p.m.—7 hours total—become more expensive, while the 2-hour window from noon to 2 p.m. remains at off-peak pricing. This happens to be exactly the time slot when typical Chinese office workers take their lunch break. If the underlying idea were simply "it's busy during the day, so raise prices," it would be simpler as a system design to just set the entire block from 9 a.m. to 6 p.m. as peak, including the noon-to-2 p.m. window. The fact that they didn't do this suggests that actual traffic visibly drops during the lunch break—in other words, that the primary users are human office workers making active, deliberate queries.
Another clue is that this peak window doesn't overlap at all with standard U.S. business hours. Converting the peak window to UTC gives 1 a.m.–4 a.m. and 6 a.m.–10 a.m., but typical U.S. East Coast business hours (roughly UTC 13:00–22:00 during Eastern Daylight Time) don't brush up against this peak window at all. If DeepSeek's heavy users skewed toward North American companies and developers, the price hike's target hours would need to reach North America's demand peak as well—but in reality, those hours remain at off-peak pricing all day.
Useful reference data comes from Similarweb figures compiled in March 2026: desktop visitors to DeepSeek's official website were 44.17% from China, with Russia at 10.35% and the U.S. at just 5.68%. This is a different metric from the breakdown of paying API users, and the survey wasn't conducted at the same time as the August price revision, but as a regional breakdown of visitors to the free chat interface, it points in the same direction as the hypothesis suggested by the peak-window design. A seemingly technical pricing system design ends up unintentionally reflecting the distribution of its user base.
Looking at the time allocation as a proportion, the peak window accounts for only 7 out of 24 hours in a day—29%—while the remaining 71% stays at off-peak pricing. If DeepSeek wanted to equally capture demand from North America and Europe as well, there would be no reason to fix this narrow 29% window exclusively to Beijing business hours. If the company were assuming a user base distributed across the globe, the peak window would need to be spread more broadly across the 24 hours, or span multiple time zones.
It's also worth noting that the late-night-to-early-morning hours remain at off-peak pricing. Batch processing, where companies process large volumes of data in bulk, is an automated workload that doesn't require human involvement, and since it doesn't demand fast response times, it can easily be shifted to late-night hours. DeepSeek's pricing design actually favors this kind of automated batch usage, while making daytime interactive use by humans relatively more expensive. In this sense, the peak-window design also functions as an economic mechanism that steers usage patterns toward human interactive use.
The 2025 "DeepSeek Shock" and the Rise of the "Chinese AI Dominance" Narrative
The reason every move DeepSeek makes draws this much scrutiny traces back to memories of January 2025. On January 20 of that year, when DeepSeek released its reasoning model "R1," which it claimed had been developed at low cost, the conventional wisdom around the computing resources needed for AI development began to shake. A week later, on January 27, Nvidia's market capitalization lost roughly $600 billion in a single day—about ¥88.2 trillion at ¥147 to the dollar. This plunge, reported by outlets including the Washington Post, became the trigger that spread concern throughout U.S. industry that "China might threaten U.S. AI dominance."
While the stock price itself recovered within a few weeks, the trend of Chinese open models expanding their global share never stopped afterward. DeepSeek's new pricing system and Qwen's performance data are easily read as the latest installment in this year-and-a-half-long saga of concern.
Behind this plunge lay the fact that the very premises of AI development had been shaken. At the time, the industry-standard view was that developing cutting-edge models required tens of thousands of the latest GPUs and massive investment. For the V3 model underlying R1, DeepSeek disclosed GPU hours and costs for the final stage of pretraining at a dramatically lower level than expected, and this figure cast doubt on future GPU demand projections themselves, directly impacting Nvidia's projected future earnings. What the market revised downward was arguably not just the stock price itself, but the underlying premise of AI investment across the entire industry.
Qwen's Derivative Models Reach 2.6 Times Meta's Count
Looking specifically at Qwen's track record—often cited as the latest chapter in this ongoing concern—the sheer scale is striking. According to a report published by Hugging Face in August 2026, derivative models based on the Qwen series numbered 151,448 on the platform—2.6 times the total number of Meta-based models, and 4.7 times the number if limited to the Llama series. A derivative model refers to a model that has been modified—through additional training, quantization, or similar techniques—from an existing open-weight model and published as a separate model. A higher count is evidence that developers worldwide are building on that model as a foundation for their own use cases.
The same report also tallied download counts: over the first seven months of 2026, Qwen-related models reached roughly 2.05 billion cumulative downloads, while Google's models reached only about 418 million and Meta's models about 227 million. While the count-based gap is 2.6x, in terms of downloads Qwen reaches roughly 9 times Meta's figure. Even this figure—calculated by Hugging Face itself—conveys clearly enough that Qwen overwhelmingly dominates the distribution volume of open models compared to others.
Whether the number of derivative models grows also depends on licensing terms. The Qwen series is released under a relatively permissive license that allows modification and redistribution, including commercial use, making it easy for companies and individual developers to publish models they've tuned for their own use cases as-is. When licensing is more restrictive, derivative models are more likely to be kept for internal use only and not published on Hugging Face. The gap in counts likely reflects not just differences in the models' underlying performance, but also institutional factors related to how easy it is to publish.
When developers select a base model for a new project, the abundance of derivative models can serve as useful reference information. The more prior examples built on the same foundation exist, the more know-how around quantization and fine-tuning, as well as troubleshooting information, accumulates within the community. The fact that Qwen's derivative model count significantly exceeds Meta's is practical material that directly influences the decision-making of developers choosing their next model. A higher count directly translates into a larger volume of accumulated know-how.
"3 Billion Downloads" vs. "2.05 Billion"—Which Number Is Correct?
In August 2026, Alibaba announced on its own that "the Qwen series has surpassed 3 billion downloads in 6 months." Japanese media outlets, including GIGAZINE, reported this figure as-is. But according to figures published by Hugging Face itself around the same time, the cumulative total for the first 7 months of 2026 was only about 2.05 billion downloads. Simply taking the ratio, Alibaba's self-reported figure is about 1.5 times Hugging Face's measured value.
What's odd isn't the ratio itself, but the relationship between the time periods. The "6 months" that Alibaba specified is a shorter period than the "first 7 months" that Hugging Face tallied. A self-reported figure for a shorter period ends up larger than a third-party tally for a longer period. This suggests that Alibaba's "3 billion" may include distribution channels beyond Hugging Face—for example, its own ModelScope platform, API usage via Alibaba Cloud, or downloads from various regional mirror servers. Hugging Face has published official documentation explaining how it counts access to target files (such as how it handles GET/HEAD requests and the possibility of duplicate counting for GGUF format), but there's no public explanation of which channels Alibaba's "3 billion" figure aggregates or how.
The download-count metric itself is inherently prone to large variation depending on the calculation method. Tallies on Hugging Face can differ significantly depending on how far they count duplicate access to the same model file, retrievals via mirror servers, and periodic re-fetches by automated build pipelines. Many Chinese companies publish the same models on their domestic ModelScope platform in addition to Hugging Face, so it's possible the figure represents a sum across multiple platforms. However, this remains just one hypothesis for explaining Alibaba's "3 billion" figure—it's not publicly disclosed which channels were actually aggregated to produce that number. What remains true is that this is a comparison made on a different basis from Hugging Face's standalone tally.
What's easy to overlook is that Hugging Face's standalone figure alone (roughly 2.05 billion) already shows a 9x gap over Meta. Given that Qwen already overwhelmingly dominates the field in actual terms, the decision to nonetheless put out a larger, self-reported figure without disclosing the basis for its calculation surely calls for some explanation.
Beyond the Battle for Market Share Lies Dominance of the Ecosystem
DeepSeek's peak pricing and Qwen's download tallies might seem like two unrelated topics at a glance. But laid side by side, they reveal a reality different from the fear-driven framing that spread in January 2025—that "China is seizing the AI market from the U.S." DeepSeek's main customers are Asian users operating on Beijing business hours, and the company hasn't deeply penetrated U.S. corporate workflows. Qwen, too, has built up numbers that surpass Meta on the metric of the Hugging Face developer ecosystem—a growth pattern different from a direct battle for market share against U.S. companies.
This structure isn't unrelated to Japanese developers, either. Alibaba Cloud offers a Tokyo region for its enterprise AI development platform, "Model Studio," creating an environment where domestic companies can access Qwen-based models via a data center within Japan. Qwen's open-weight models themselves have long been available for download and self-hosting via Hugging Face, GitHub, and ModelScope, but the Tokyo region adds a separate option: a managed API with lower latency and support infrastructure for domestic companies. Pricing remains denominated in US dollars, so it doesn't eliminate exchange-rate exposure entirely, but the expanded options in terms of access latency and support are not insignificant.
DeepSeek is building up numbers in actual usage across Asia, while Qwen is building up numbers in developer-ecosystem distribution volume—each accumulating strength along a different metric. As Model Studio's launch in Japan shows, it won't be U.S. companies that first feel the expansion of this playing field, but developers across Asia, including Japan.