Qutjing Technology Completes Over 1.3 Billion Yuan Series B Financing, Breaking Down Two Token Factory Accounting Models Against SiliconFlow's -24% Gross Margin

Beijing Qujing Technology recently completed a Series B funding round of over 1.3 billion yuan. Zhejiang Caixin Innovation, Xiaoshan Capital, ICBC Investment, Zhongguancun Science City, Xinglian Capital, and Qianjiang Century City jointly led the round, with SenseTime, Wangsu Science & Technology, CITIC Juxin, and others participating.

The company was founded less than three years ago, and has already raised a cumulative 2.4 billion yuan within 2026, averaging one round every two months.

The team's background is also worth a look. Founder Ai Zhiyuan holds a PhD from Tsinghua University's Department of Computer Science and previously served as an R&D director at Sangfor; co-founder Ren Xuyang is an early founding veteran of Baidu and founder of Zhenzhi Ventures; chief scientist Wu Yongwei is a Tsinghua professor, and Academician Zheng Weimin serves as chief scientific advisor. The core team comes from Tsinghua KVCache.AI and has ten years of accumulated experience in "high-performance computing + distributed storage."

The business it does is called "AI Token production services"—the industry more commonly calls it a "Token factory."

And in the wording of this funding round, there is one sentence more noteworthy than the 1.3 billion: the company states that its "overall business has crossed the cost line."

The contrast behind one sentence

First, let's look at what that sentence is worth.

In the same business, there is a company that has already filed with the Hong Kong Stock Exchange—SiliconFlow. Its prospectus data is as follows:

2024 revenue of 7.35 million yuan, 2025 revenue of 55.33 million yuan, a year-on-year increase of 653%; 2024 loss of 81.92 million yuan, 2025 loss of 345 million yuan; gross margin: 39.4% in 2024, turning to -24% in 2025, recording a gross loss of 13.3 million yuan.

In other words, this company, ranked first among China's independent ecosystem token supply platforms, has sold tokens down to negative gross margin.

The reasons need to be broken down. SiliconFlow's public cloud business gross margin is itself negative, at -271.6% in 2024, improving to -119% in 2025; the overall gross margin fell from 39.4% to -24%, mainly because the share of low-margin public cloud revenue rapidly rose from 14.6% to 52.9%, diluting the overall figure.

At the same time, the prospectus states: sales and marketing expenses surged 1,210% from 6.39 million to 83.74 million, of which "promotional computing power resource costs"—that is, the cost of free Token vouchers—accounted for 64.7%. This money is an expense, not the direct cause of the gross margin turning negative.

By the way, the industry's volume is exploding: the size of China's token supply market grew 1,602.6% from 2024 to 2025; average daily Token calls rose from about 100 billion in early 2024 to more than 140 trillion by March 2026. Volume grew more than a thousandfold, yet no money was made—this is the most counterintuitive part of this business.

Since 2023, leading large-model vendors have cut API prices more than ten times cumulatively, with the per-thousand-Token price of some mainstream models falling from several tenths of a yuan initially to a few cents, a cumulative drop of over 90%. In May 2026, DeepSeek announced a permanent 75% price cut for the V4-Pro API; Xiaomi followed, with cuts of up to 99%; starting June 3, Tencent Cloud lowered prices for the DeepSeek-V4 series, with ordinary input and output down 75% and the "cache hit" portion down by as much as 97.5%.

Downstream is cutting prices, while upstream costs are rigid. This is the structural dilemma of the Token factory.

So when Qujing says it has "crossed the cost line," it is actually saying: we have found another way to survive.

The Token business has two sets of accounts

The term "Token factory" actually contains two different sets of accounts: one sells by Token, the other builds by factory.

One is sold retail.

Packaging engineering capabilities such as chip adaptation, model deployment, inference tuning, and heterogeneous scheduling into standard APIs, hanging them on the public cloud, where anyone can come and buy, paying by usage. The core of this business is "retail"—the more users the better, and the lower the price the more competitive. SiliconFlow's public cloud business takes this path. Its public cloud revenue share rose from 14.6% in 2024 to 52.9% in 2025, with registered users exceeding 10 million and enterprise customers exceeding 13,000. But it also has another half: the local deployment business, which still accounted for 47.1% of revenue in 2025, and this part has higher gross margin.

But the script for the retail business is brutal: to attract new users, you have to hand out free Token vouchers; to retain customers, you have to follow the big vendors' price cuts. Squeezed at both ends, the gross margin disappears.

One is built by factory.

In addition to selling Tokens directly to model vendors, internet platforms, and large enterprises, Qujing also cooperates with local governments, industry parties, and computing power providers to jointly build "Token factories" in a region, with the two sides sharing revenue as agreed.

Beyond direct sales, Qujing is doing this at the same time. Its public statement is: working with industry partners to advance more than ten projects of daily-average trillion-level Token factories, jointly building production lines with SenseTime, and partnering with Parallel Technology, Hongxin Electronics, Wangsu Science & Technology, and others to build Token factories, laying out regional infrastructure in Henan, Hangzhou, and Hainan.

The difference between these two sets of accounts is not in technology, but in how money is collected.

The supermarket's customers are millions of developers, who are extremely price-sensitive and will switch over a one-cent difference.

The power plant's customers are local governments and industrial groups; what they buy is not "Tokens" but "computing infrastructure landing"—which involves land, electricity fees, subsidies, and taxes; the accounts are not calculated by Token.

What supports Qujing's self-described "crossing the cost line"

Qujing did not disclose financial data by business segment. What exactly supports "crossing the cost line" can only be inferred from a few pieces of public information.

The first is the cost structure. Building factories regionally may secure lower electricity rates and cheaper land. And in the cost of a Token, electricity and depreciation account for the bulk. Whether this holds true and how much can be saved—Qujing has not made public data available, so this can only serve as a speculation pending verification.

The second is the technology route. Qujing's technological proposition is "trading storage for compute"—using storage to replace computation. To put it plainly: rather than stacking more GPUs to compute, it is better to store intermediate results (KV Cache) for reuse, reducing redundant computation.

And this team happens to be the very group of people with a decade of accumulated expertise in "trading storage for compute."

Here "trading storage for compute" can be translated into plain terms: when a large model generates a response, it repeatedly uses intermediate results that have already been computed. If it recomputes each time, it repeatedly occupies GPUs; if these results can be stored and retrieved directly next time, a large amount of redundant computation can be saved. Qujing calls this "trading storage for compute," and the logic is to shift costs from "expensive memory" to "cheap storage."

This approach is not new, but the difficulty lies in engineering—it must run stably at the scale of trillions of parameters and trillions of Tokens, while also ensuring latency does not increase. If achieved, the unit Token cost can be a notch lower than "pure compute stacking."

The third is the revenue structure. Under the co-operation model, Qujing's revenue may include portions from engineering construction and delivery, rather than only billing per API call. But how the accounts are split specifically, how much this part accounts for, and what the gross margin is—the company has not disclosed.

So the phrase "crossing the cost line" may come from the combined result of its direct-sales and co-operation businesses. But the company has not disclosed financial data by business segment, so which segment this phrase actually corresponds to and how long it can last can only remain in doubt for now.

The co-operation path also has its costs

The co-operation path is not without cost. Several of its costs happen to correspond to the advantages of the retail model.

First, weak economies of scale.

In the retail business, each additional user still consumes servers, VRAM, and network according to call volume, so marginal cost is not zero; in the co-operation business, each additional customer requires renegotiating cooperation and building capacity, making expansion heavier.

Second, high customer concentration.

More than ten projects sounds like a lot, but if a few of them are the main sources of revenue, then any delay in any one project will directly affect cash flow.

Third, and most critically: its customers are themselves being battered by the price war.

If the retail price of Tokens keeps falling to a few cents, then the accounts for local factory building will change accordingly—an investment return that can be calculated today may not be calculable next year.

Two routes, each with its own accounts

Putting the two companies together, one can see the two kinds of accounts in this business today.

SiliconFlow takes the scale route centered on public cloud: trading scale for share, at the cost of sustained losses and negative gross margin. Its profit expectations are built on the narrowing losses of its public cloud Serverless business and future room for price increases.

Qujing takes the route of direct sales plus co-operation: selling Tokens on one hand, and helping localities and industry parties build factories and co-operate them on the other. Whether it can sustainably cover costs depends on whether the co-operation accounts add up and whether the cost-effectiveness of domestic computing power can be realized.

Both routes are viable.

The retail route is valued by user count and call volume; the co-operation route is valued more on orders and delivery. The former has greater room for imagination.

And the biggest variable in the entire track right now is not technology, but price cuts—when Tokens become as cheap as tap water, the business of selling water becomes increasingly difficult.

---

Back to Qujing Technology's 1.3 billion yuan.

It raised 2.4 billion yuan during the year, one round every two months, with the money mainly invested in "high-quality AI Token capacity expansion" and upgrading the ATaaS platform. Average daily Token output has exceeded one trillion, and the company disclosed that business revenue during the year reached nearly ten times that of the entire previous year.

These numbers are all impressive. But what truly determines whether it can keep going is whether the phrase "has crossed the cost line" can be maintained.

Because in this business, the cost line is not fixed. Downstream prices have cumulatively fallen by 90% over the past two years, while upstream factories, chips, and electricity costs are rigid. Crossing it today, it may be pushed back tomorrow.

SiliconFlow's gross margin fell from 39.4% to -24% in just twelve months.

The decline in gross margin shows that the matching of revenue and cost is deteriorating, and cannot be directly equated with a decline in unit cost. How long a company can stand on this line does not depend on how much money it raised this time, but on whether every cent it saves truly comes from technology.