Silicon Flow IPO: The 'Token Factory' Experiencing 653% Growth—Why Are Sales Increasing While Losses Mount?

08/30 2026 442

Author|Pan Yan

Editor|Wei Fanxi

// Valuation Hits Regulatory Threshold Right Before IPO Filing

Recently, Beijing Silicon Flow Technology Co., Ltd. (Silicon Flow), established merely three years ago, submitted its listing application to the Hong Kong Stock Exchange, aspiring to become the 'pioneer AI Token factory stock'.

The so-called Token factory is engaged in the wholesale and retail of computing power. Silicon Flow procures GPU computing power from cloud service providers, integrates it into its proprietary inference engine, and transforms it into Tokens for sale to developers and small-to-medium-sized enterprises (SMEs).

This business model has enabled Silicon Flow to secure seven rounds of financing over three years, with its valuation skyrocketing from 280 million yuan to 7.74 billion yuan. Its investor roster includes industry giants such as Alibaba, Huawei, and Zhipu AI.

However, the prospectus also reveals that Silicon Flow has incurred cumulative losses of 440 million yuan, including a net loss of 345 million yuan in 2025, with its gross margin plummeting to -24.04%.

What is impeding this Token factory, which is supported by renowned shareholders and riding the wave of AI?

Independent Yet Holding a Mere 1.5% Market Share

On paper, Silicon Flow has demonstrated remarkable growth.

Four months after its inception in August 2023, Silicon Flow generated cumulative revenue of 6,000 yuan. In 2024, its annual revenue surged to 7.345 million yuan, and further to 55.33 million yuan in 2025, marking a year-on-year increase of 653.20%.

This revenue explosion was fueled by the surge in demand for AI inference.

In early 2025, DeepSeek went viral, overwhelming its official servers. In February of the same year, Silicon Flow partnered with Huawei Cloud to launch DeepSeekR1 and V3 inference services based on Ascend computing power, capturing a significant portion of the overflowing traffic.

As of April 2026, Silicon Flow's platform had registered over 10 million users, with an average daily token throughput of approximately 578.5 billion and a single-day peak of around 1.07 trillion. It had served over 13,000 enterprise clients and supported over 170 models.

Despite these impressive figures, Silicon Flow ranked fourth in China's token supply market in 2025, holding just a 1.5% market share based on annual token throughput.

In contrast, the top three players—Volcano Engine, Alibaba Cloud, and Baidu Intelligent Cloud—accounted for a combined 87% market share.

The disparity in market share stems from differing business models. The top three operate closed-loop ecosystems encompassing computing power, models, and applications, with Token demands from Doubao, Tongyi Qianwen, and ERNIE Bot naturally flowing to their own platforms.

In contrast, Silicon Flow does not confine itself to any single cloud provider or large model. Instead, it connects with various computing powers, including NVIDIA and Ascend, and over 170 models. The prospectus describes this model as an 'independent ecosystem token supplier' and claims it to be China's first.

However, 'independence' comes at the cost of high customer churn.

In 2023, Silicon Flow had only two clients, accounting for all its revenue that year. In 2024, its top five clients contributed 85% of revenue, but this figure dropped to 45% in 2025, with no overlap in major clients over the three years.

During the same period, the number of local deployment solution clients for large institutions declined from 28 to 20, with each contributing 1.303 million yuan. Meanwhile, the number of serverless token service clients for individual developers and SMEs skyrocketed from 2,454 to 716,000, but each contributed only about 20 yuan on average.

Without barriers to independence, customers come and go, leaving Silicon Flow with escalating customer acquisition costs that ultimately weigh on its income statement.

The Token Business Turns into a 'Money-Losing Proposition'

From 2023 to 2025, Silicon Flow reported net losses of 12.223 million yuan, 81.915 million yuan, and 345 million yuan, respectively, accumulating over 440 million yuan in losses over three years. During the same period, its gross margin continuously declined, from 83.3% to 39.4% and then to -24.0%.

The divergence between revenue and profit began with shifts in business structure.

In 2025, revenue from public cloud services for selling Tokens jumped from 14.6% of the total in the previous year to 52.9%, replacing local deployment as the primary revenue source.

Every additional unit of revenue from this business incurs two rigid costs: computing power procurement and customer acquisition.

Silicon Flow purchases all its computing power from third parties. From 2024 to 2025, its sales costs surged from 1,000 yuan to 68.63 million yuan, a compound annual growth rate of 26,097.70%, far outpacing revenue growth. Computing power resource costs accounted for 86.9% of sales costs in 2025.

The prospectus explicitly states, 'Even with significant revenue increases, cost growth will continue to outpace revenue growth.' Financially, this translated to a gross margin of -119.0% for public cloud services in 2025, meaning Silicon Flow was still selling Tokens at a loss.

Customer acquisition costs also soared with the shift in business structure.

In 2025, to attract new users to its public cloud platform, Silicon Flow distributed free Token vouchers totaling 54.213 million yuan, accounting for 64.7% of sales and marketing expenses.

In comparison, these vouchers primarily drove growth in serverless token service clients, who generated only 14.3 million yuan in revenue in 2025, less than 30% of the voucher investment.

Research and development (R&D) expenses were another major outlay. In 2025, R&D spending surged by 224.43% year-on-year to 209 million yuan, accounting for 378.1% of current revenue, primarily for inference engine iteration and domestic chip adaptation.

In total, computing power, vouchers, and R&D expenses reached 331 million yuan in 2025, six times the revenue for that period.

Silicon Flow's cash reserves are depleting rapidly. From 2024 to 2025, its net operating cash flow was -46.51 million yuan and -172 million yuan, respectively.

As of the end of 2025, Silicon Flow had 172 million yuan in cash and cash equivalents and 100 million yuan in time deposits, with approximately 272 million yuan in available funds, just enough to cover short-term borrowings of 40 million yuan.

According to the prospectus, Silicon Flow's average monthly cash burn rate climbed from about 4 million yuan in 2024 to 14.8 million yuan in 2025. At this pace, its cash reserves will only last 18 months.

The real issue is that this money did not come from operations but from financing. If financing dries up, cash burn will not cease automatically.

Valuation Surges Prior to IPO

In three years, Silicon Flow has completed seven rounds of financing, raising a cumulative 1.951 billion yuan, with its post-investment valuation soaring from 280 million yuan to 7.74 billion yuan, and its per-share cost rising from 22.59 yuan to 303.66 yuan.

Its investor lineup spans the entire AI supply chain, including upstream computing power companies like Biren Technology, cloud purchasers like Alibaba and Huawei, and model application service providers like Zhipu AI, SenseTime, and Sinovation Ventures.

Behind the shareholder list lies a subtle interplay of competition and cooperation. Institutions like Zhipu AI and SenseTime are both shareholders of Silicon Flow and direct competitors in the MaaS (Model as a Service) sector.

Notably, in June 2026, Silicon Flow announced a high-profile completion of a Series B funding round exceeding 2 billion yuan, setting a record for the largest single funding round in China's third-party MaaS sector in 2026.

However, the actual amount disclosed in the prospectus fell significantly short. In April and June 2026, Silicon Flow completed Series B and Series B+ funding rounds, raising a combined 1.26 billion yuan, over 700 million yuan less than the claimed amount.

More intriguingly, the timing aligned precisely with regulatory requirements.

In April 2026, after completing its Series B funding, Silicon Flow's valuation stood at approximately 5.02 billion yuan, still short of the Hong Kong Stock Exchange's Chapter 18C requirement for 'non-commercialized companies to have a market capitalization of at least 8 billion Hong Kong dollars (about 7 billion yuan).' However, after the Series B+ funding in June 2026, its post-investment valuation of 7.74 billion yuan precisely met the listing threshold.

The gap between publicly claimed financing amounts and actual funds received, combined with the precise timing of valuation adjustments before the IPO filing, has raised doubts about the reliability of Silicon Flow's disclosures.

Additionally, the founder's cash-out ahead of the IPO warrants scrutiny. From January 2025 to June 2026, Silicon Flow completed four equity transfers, three of which involved founder Yuan Jinhui personally selling shares, cashing out a total of 38.5197 million yuan.

Ahead of the IPO, Yuan Jinhui directly holds 14.35% of the shares and controls another 30.14% through four equity incentive platforms, totaling 44.49% control.

Among external shareholders, Alibaba holds 7.42%, Huawei holds 4.07%, Sinovation Ventures holds 4.01%, and institutions like Glow Capital, Wang Huiwen, and Puhua Capital hold between 1% and 3% each.

With cash burning fast, Silicon Flow is also seeking lower-cost computing power sources.

Shortly after filing for the IPO in August 2026, Silicon Flow announced two computing power partnerships: a joint operation with Guizhou Suanjia Computing leveraging idle computing power resources in Gui'an New Area, and a 'Token factory' co-construction agreement with Shandong Railway Development Fund.

The direction is clear—light-asset computing power acquisition and securing state-backed support. However, neither partnership has disclosed specific computing power scales or investment amounts, leaving their progress to be verified.

While a path to reduce computing power costs exists, whether it can succeed remains uncertain. Meanwhile, Silicon Flow's cash reserves may not last much longer.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.