The state has given inference a name, a statistic and a procurement category. Underneath the label, three different businesses are being run — and prices are moving in opposite directions.
DeepSeek now charges by time of day. Between 9 am and noon, and again between 2 pm and 6 pm Beijing time, its API costs double. Off-peak, the price is unchanged.
Peak-and-trough pricing is what utilities do. That is roughly the condition Chinese inference has arrived at: not a product being sold, but a supply being metered.
The state named the unit
In March, the head of China’s National Data Administration, Liu Liehong, told the China Development Forum that the country’s daily token consumption had passed 140 trillion. The figure was 100 billion in early 2024 and 100 trillion at the end of 2025 — a gain of more than a thousandfold in two years, and more than 40 per cent in the first quarter of 2026 alone.
In the same remarks he did something less commonly noticed abroad. He gave the token an official Chinese name — ciyuan, 词元 — and described it as a settlement unit for the intelligent era. The National Data Administration now reports token consumption the way a statistics bureau reports electricity output.
Liu also named token exports as evidence of Chinese industrial competitiveness. That framing matters for reading everything below: the movement of inference capacity abroad is not only a commercial strategy pursued by individual firms, it is a described national objective.
The supporting numbers arrived with it. More than 100,000 high-quality datasets, totalling over 890 petabytes, had been built by the end of 2025. JPMorgan projects Chinese inference token consumption growing roughly 370-fold between 2025 and 2030.
Three businesses, one label
“Token factory” — 词元工厂 or Token工厂 — is now attached to at least three commercial arrangements that have very little in common.
The first is a capacity framework. In April, China Telecom’s Ningxia branch opened procurement for what it called Token factory generation capacity services. The framework is estimated at 16.45 billion yuan before tax, or 17.44 billion including it, divided into eleven packages over a service period of sixty months. It is the first procurement of this scale in China to buy token generation capacity as such.
What the winners are contracted to supply is neither hardware nor cloud instances. It is standard token generation capacity, to be provided within thirty days of an order — a delivery window that was originally set at ten days and then relaxed. Package one’s candidate winners were published in mid-May: Huasheng Tiancheng with 18 per cent, Lexun Technology 17.5, Hubei Big Data Group 17, Qingxin Zhisuan 16.5, Chuangshiyun 16 and Dongfang Guoxin 15.
One detail in that result is worth holding onto. Six of the seven candidates bid between 99.2 and 99.8 per cent of the ceiling price. Only Chuangshiyun went materially lower, at 97 per cent. At the procurement layer, in other words, there is almost no price competition at all — which sits oddly beside what is happening to published API prices, discussed below.
These are estimated framework values, not executed contracts, and should be read as such.
The second is a hardware bid wearing the label. On 1 July, Digital China (000034.SZ) disclosed that its wholly owned Shanghai subsidiary had been named first candidate in the evaluation for a domestic compute Token factory project, with a bid of 717 million yuan. The equipment is the company’s own Kunpeng-and-Ascend-based supernode servers. It is a candidacy, not a confirmed award.
The scale is worth setting in context. In May the same company was shortlisted on China Mobile’s server procurement, where its bid on the first package alone was 7.84 billion yuan before tax. Measured against that, the token factory business is not yet where the volume is for a hardware integrator; it is where the language is.
The third is a managed operating model, and it is the only one of the three that is plausibly new. Approaching.AI, founded in Beijing in 2024 by a team out of Tsinghua’s high-performance computing institute, announced an A round on 13 July that took its cumulative raise past 1 billion yuan in six months. The round was led by a fund under Henan Investment Group, a provincial state holding company.
Approaching.AI does not own the chips. It sells a platform that converts compute into token output meeting agreed service levels — inference as an operated layer rather than a rented one. The company says per-machine token production efficiency has more than tripled since the Lunar New Year and that its total capacity has grown more than thirtyfold, with one trillion-parameter model sustaining daily output above a trillion tokens. Those figures are the company’s own. It says it is now building a facility with the lead investor targeting several trillion tokens a day.
Three arrangements, then: a utility-style capacity framework, a server sale, and an operated inference layer. They are being described with a single phrase.
Prices are moving both ways
The familiar story about Chinese inference is a price war. That is half right.
Chinese laboratories cut API prices six times in the first half of 2026, three of those declared permanent. As of prices surveyed in May and June, DeepSeek’s V4-Pro was listed at $0.87 per million output tokens, roughly a thirty-fourth of GPT-5.5 at about $30. Claude Opus 4.7 sat near $15. Across the five leading Chinese flagships, output pricing had fallen to somewhere between a tenth and a thirty-fourth of the international majors.
But over the same months, several Chinese vendors raised prices. Zhipu increased its rates three times — its coding plan by 30 per cent in February, GLM-5-Turbo by 20 per cent in March, GLM-5.1 by 10 per cent in April, with overseas pricing up far more steeply — citing demand growth. Alibaba Cloud withdrew basic tiers on its Bailian platform and raised AI compute pricing by up to 34 per cent. Tencent Cloud raised prices from 9 May, citing surging global compute demand and rising hardware supply chain costs.
One week at the end of April captured the whole pattern. DeepSeek released a V4 preview on Hugging Face under an MIT licence with pricing that undercut everything around it. Zhipu raised prices for the third time that year. Alibaba and Tencent raised theirs. And ByteDance moved in the opposite direction again: on 4 May its Doubao assistant posted subscription tiers on the App Store, formally launching them on 24 June at 68, 200 and 500 yuan a month.
The Doubao reversal is the sharpest single illustration. In May 2024 Doubao entered the API market at pricing reported as 99.3 per cent below its peers, and effectively started the price war. Two years later it is charging consumers. The reasoning is visible in the market data: CNNIC counted 602 million generative AI users in China by December 2025, a penetration rate of 42.8 per cent, and Doubao’s own monthly active users had reached 345 million — close to the ceiling of the domestic consumer market. Growth in users has slowed while compute costs have not. A Citi survey found 45 per cent of respondents willing to pay for advanced AI features, but at an acceptable average of 48.3 yuan a month, about 30 per cent below Doubao’s entry tier.
What is actually happening is not a price war but a separation. Generic inference is commoditising toward cost. Capability that is genuinely differentiated is charging more. Across seventeen vendors surveyed in mid-2026, published rates ranged from $0.022 to $180 per million tokens — a spread of roughly eight thousand times. Whether Chinese tokens are cheap depends entirely on which layer is being bought.
Where the cheapness comes from
Hardware accounts for around half the cost of a token. That places the domestic chip transition directly in the pricing story: domestically produced AI chips passed 40 per cent of shipments in China in 2025, with Huawei’s Ascend line taking about 47.8 per cent of that domestic share, and Ascend supernodes have been brought into support for DeepSeek’s V4 series.
Electricity is the other half of the equation. Runjian, the Guangxi operator of the Wuxiang Yungu token facility, puts power at 50 to 60 per cent of operating cost during inference — the company’s own figure, but a plausible one given industry estimates that Chinese compute centres consumed around 170 billion kilowatt-hours in 2025, about 1.6 per cent of national electricity use, with projections reaching 800 billion by 2030.
That is why token factories are being sited where the power is: Ningxia, Inner Mongolia, Guangxi. The cheapness is not principally a subsidy. It is an electricity tariff, a power mix and a chip supply chain, expressed as a price per million tokens. Regional differences in token prices, as one Chinese industry analysis put it, have stopped being a temporary market condition and become a durable structure shaped by energy costs, chip access and industrial policy.
Where the value ends up
The demand for these tokens is now visibly international. On OpenRouter, the API aggregator, Chinese model calls rose 127 per cent over three weeks in February 2026 and passed US models for the first time; four of the platform’s top five models that month were Chinese — MiniMax’s M2.5, Moonshot’s Kimi K2.5, Zhipu’s GLM-5 and DeepSeek’s V3.2 — together accounting for 85.7 per cent of top-five volume. A year earlier, Chinese models held under 2 per cent of that platform. Martin Casado of a16z has estimated that around 80 per cent of startups building on open-source AI stacks are running Chinese models.
The open commercial question is which layer of this supply chain keeps the money. The procurement frameworks pay integrators. The hardware bids are one-time revenue at integrator margins, which are thin. The operated layer — the Approaching.AI model — is the only one structurally positioned to earn a share of output rather than a fee for equipment, and it is also the least proven.
For buyers in the Global South, the arithmetic is straightforward and, for now, favourable: inference priced by the token removes the need to own high-end compute at all. But the pattern visible in China’s first export cases is that the factory does not travel. Runjian’s ASEAN cooperation agreements, signed at its own forum this month, extend a Nanning facility’s customer base rather than build capacity in Malaysia or Indonesia — a distinction covered separately in this publication’s Deployment Watch.
Which leaves the question the industrialisation of inference has not yet answered. Cheap metered tokens genuinely lower the barrier for countries that cannot build frontier compute. They also place another country’s inference layer on China’s grid, China’s chips and China’s operating standards. Both of those are true at once, and it is too early to say which one ends up mattering more.
Prices cited were surveyed between May and July 2026 and change frequently; API rates should be checked against vendor listings before use. Company performance figures are attributed where they are self-reported.
