Nearly 2.9 billion yuan of financing was raised in one year, the "national team" institutions have made intensive investments ahead of its IPO, and Silicon Flow is sprinting to become the "first listed stock of the Token factory track".

An AI company founded only in 2023 with revenue of just 55.33 million yuan in 2025 is becoming a hot target for "national team" capital to compete for investment.

On September 20, Silicon Flow announced the completion of the second phase of its B+ round and Series C financing, with investors including China Internet Investment Fund, China New Era Holding Group Fund, China Mobile Chain-Length Fund and other state-backed capital entering the market intensively. To date, the company's total accumulated financing amount in 2026 has approached 2.9 billion yuan. More than three months ago, Silicon Flow just submitted its prospectus to the Hong Kong Stock Exchange, aiming to become the "first share of AI Token factory".

Behind the capital enthusiasm, the other side of Silicon Flow is the continuously expanding loss: in 2025, the company's revenue increased by 653.2% year on year, but its gross profit margin dropped from 39.4% to -24.0%, and the annual net loss reached 345 million yuan.

On one hand, the national team is making intensive bets, on the other hand, the company sees growing losses as its revenue expands. What exactly does Silicon Flow sell? Why can this AI infrastructure company founded only three years ago win the bets from state-owned capital?

On the eve of listing, the "national team" enters the market intensively

Founded in August 2023 and headquartered in Haidian, Beijing, Silicon Flow is a third-party "Token supplier".

According to data from Frost & Sullivan, calculated by the annual Token throughput in 2025, Silicon Flow has become the largest independent ecological Token supplier in China, ranking top 5 among all Token suppliers across the country.

With the recent successive completion of Series B+ and Series C financing, Silicon Flow's total financing amount in the year has exceeded 2.9 billion yuan. In June this year, the company publicly announced the completion of over 2 billion yuan of Series B financing, setting the record of the largest single financing in China's third-party MaaS track since 2026.

The intensive influx of capital has also helped Silicon Flow cross the threshold for listing on the Hong Kong Stock Exchange.

According to the previously disclosed prospectus, after the completion of the Series B+ financing in June, the post-investment valuation of Silicon Flow reached 7.74 billion yuan, which exactly crossed the market capitalization threshold for uncommercialized companies set out in Chapter 18C of the Hong Kong Stock Exchange — the expected market capitalization at listing is no less than 8 billion Hong Kong dollars, equivalent to about 7 billion yuan.

Looking back at all previous financings, Silicon Flow's shareholder lineup is already very strong, covering almost the upstream and downstream of the AI industrial chain: there are upstream computing power enterprises such as Biren Technology, cloud computing and computing power purchasers such as Alibaba and Huawei, as well as model and AI industry participants such as Zhipu AI, SenseTime and Sinovation Ventures.

The prospectus shows that among the external shareholders before IPO, Alibaba holds 7.42% of the shares, Huawei holds 4.07%, and Sinovation Ventures holds 4.01%; the shareholding ratios of institutional and individual shareholders such as Yuntu Capital, Wang Huiwen and Puhua Fund are mainly concentrated between 1% and 3%.

However, on the eve of listing, a noticeable change has taken place in Silicon Flow's capital territory.

In the just-completed Series B+ and Series C financing, the "national team" funds began to enter the market intensively. The investors include China Internet Investment Fund, China New Era Holding Group Fund, China Mobile Chain-Length Fund, China Orient Asset International, as well as a number of central and local state-owned capital platforms such as Jingneng Fund, Shanghai Inesa Intelligent Computing Fund, Beijing Venture Capital, Chengdu Science and Technology Innovation Investment, Jiangxi Financial Holding; at the same time, Tianchuang Capital also participated in the investment together with Shandong Iron and Steel Investment and Shandong Railway Fund.

Different from the in-depth participation of industrial capital such as Alibaba, Huawei, Zhipu AI and SenseTime in previous rounds, the presence of central SOE funds, local state-owned capital and government industrial funds has increased significantly in the two rounds of financing near the IPO.

Why did the "national team" choose to enter Silicon Flow intensively right on the eve of its listing?

The answer may lie in more than the upcoming listing of an AI unicorn. As a Token infrastructure platform connecting domestic chips, cloud computing, large models and downstream AI applications, the position of Silicon Flow and its business model different from ordinary large model companies may be the key to understanding this round of intensive entry of state-owned capital.

Not building large models, but acting as an independent third-party "Token factory"

From the perspective of business model, what Silicon Flow does is not complicated. It connects the underlying computing power, AI models and upper-layer applications through system software, and converts the originally scattered and heterogeneous computing power resources into standardized and callable Token services.

Silicon Flow positions itself as an "open and independent Token supply platform". It does not bind a single cloud vendor, chip vendor or model vendor. Instead, it aggregates different types of computing power resources through self-developed inference engine and computing power orchestration system, and then provides customers with inference services of mainstream AI models through a unified API interface.

To put this business more plainly, Silicon Flow obtains GPU resources from cloud vendors and computing power suppliers, integrates them into its own inference engine and scheduling system, then converts these computing power into standardized Tokens and sells them to developers and enterprise customers.

The value of this model is first reflected in solving the problems of "computing power fragmentation" and "model deployment complexity".

The performance, software ecology and adaptation methods of different chips vary greatly, and different models have different requirements for computing power. For a large number of small and medium-sized enterprises and developers, purchasing GPUs, deploying models and adapting to different chips by themselves will not only bring high cost, but also high technical threshold. What Silicon Flow does is to encapsulate these complex works in the underlying layer, so that customers only need to call the model through a unified interface.

In early 2025, DeepSeek became a hit, which served as an important verification of Silicon Flow's business model. At that time, the official server of DeepSeek was once congested. In February of the same year, Silicon Flow cooperated with Huawei Cloud to take the lead in launching DeepSeek-R1 and V3 inference services based on Ascend computing power in the industry, undertaking a large amount of overflow demand.

The scale of the platform has expanded rapidly as a result.

As of April 2026, the number of registered users on the Silicon Flow platform has exceeded 10 million, with an average daily Token throughput of about 578.5 billion times and a single-day peak of about 1.07 trillion times; the number of enterprise customers has exceeded 13,000, and a total of more than 170 models have been supported.

This growth rate is even more obvious on the revenue side.

Within the four-month operation period after its establishment in August 2023, Silicon Flow's total revenue was only 6,000 yuan; its annual revenue in 2024 was 7.345 million yuan, which jumped to 55.33 million yuan in 2025, a year-on-year increase of 653.2%. Although the absolute revenue scale is still small, its growth curve is very steep.

On the product side, Silicon Flow is also continuously supplementing its infrastructure capabilities from model calling to computing power scheduling.

In May 2024, the company officially launched the MaaS platform, putting forward the "Token factory" mode with standardized, large-scale Token production and supply as the core. In April this year, Silicon Flow further launched a new generation of computing power scheduling engine "Elastic GPU", which can carry out unified scheduling and elastic scaling for different types of heterogeneous computing power, and support customers to deploy models on their own.

As of June 21, 2026, Silicon Flow has served more than 13,000 enterprise customers in total, supported more than 170 mainstream AI models, and completed the adaptation of more than 10 chips from vendors such as NVIDIA, Huawei Ascend, MetaX and Moore Threads.

This is also an important difference between Silicon Flow and ordinary large model companies. It does not directly bet on a certain model, but tries to become the intermediate infrastructure layer connecting chips, computing power, models and applications. The more models, chips and heterogeneous computing power there are, the more prominent the scheduling and standardization value of this layer will be.

The larger the scale, the more losses, what is the difficulty of the Token business?

The rapid growth of revenue has not improved the profit side, but instead made Silicon Flow fall into the dilemma of "the more it grows, the more it loses".

In 2025, the company's gross profit margin fell from 39.4% in 2024 to -24.0%, behind which is the fierce price competition in the Token supply industry.

With the open source of models such as DeepSeek, the API call price continues to decline. For Token suppliers, only when the decline rate of unit computing power cost continuously outpaces the decline of Token price, can scale expansion bring real profits.

The problem of Silicon Flow is that it has to face the price war, and lacks the internal traffic that top cloud vendors naturally own.

Calculated by the annual Token throughput in 2025, Silicon Flow ranks fourth in China, but its market share is only 1.5%; the top three players including Volcano Engine, Alibaba Cloud and Baidu Intelligent Cloud account for 87% of the total market share. The latter have computing power, models and application ecology at the same time, and the Token demand generated by Doubao, Tongyi Qianwen, Wenxin and other services can be directly precipitated on their own platforms.

In contrast, Silicon Flow's "independence" brings cross-chip and cross-model flexibility, but also means that customers need to be acquired one by one from the market.

In 2024, the top five customers of the company contributed 85% of the revenue, which dropped to 45% in 2025, and there was no overlap in the list of major customers for three years. In the same period, the number of high-unit-price local deployment customers dropped from 28 to 20, while the number of serverless Token customers for developers and small and medium-sized enterprises soared from 2,454 to 7.16 million, with an average revenue per customer of only about 20 yuan.

The number of users has exploded, but the user value has not increased synchronously.

In 2025, the revenue proportion of public cloud Token services rose from 14.6% to 52.9%, becoming the largest business for the first time, but its gross profit margin was as low as -119.0%. Silicon Flow does not own large-scale GPU assets itself, and its computing power is mainly purchased from third parties. Its sales cost rose to 68.63 million yuan that year, of which the cost of computing power resources accounted for 86.9%.

Customer acquisition also burns a lot of money. In 2025, the company issued 54.213 million yuan of free Token vouchers, accounting for 64.7% of sales and marketing expenses. At the same time, the R&D investment reached 209 million yuan, a year-on-year increase of 224.4%, equivalent to 378% of the annual revenue, which is mainly used for inference engine optimization and domestic chip adaptation.

The more Tokens Silicon Flow sells, the more computing power it needs to purchase; in order to compete for customers, it also needs to continue to cut prices and provide subsidies; and R&D investment is difficult to reduce.

From 2023 to 2025, the net losses of Silicon Flow were 12.22 million yuan, 81.92 million yuan and 345 million yuan respectively, with a total accumulated loss of more than 440 million yuan in three years. The net outflow of operating cash flow further expanded to 172 million yuan in 2025.

What Silicon Flow needs to prove to the market is not how fast the Token scale can grow, but whether the unit Token cost can drop fast enough after the scale expands, so as to finally reverse the situation of "the more revenue grows, the more losses expand".

This article is from WeChat official account "Pai Finance Official" (ID: paicj314), written by Wang Zheping, and published with authorization from 36Kr.