On the final day of the second quarter, SiliconFlow filed a listing application with the Hong Kong Stock Exchange. The IPO was submitted under Chapter 18C as a pre-commercial company. Huatai International and Haitong International are the joint sponsors.

For artificial intelligence developers, SiliconFlow may be a familiar name. When DeepSeek first gained attention, SiliconFlow was reportedly the platform that first deployed the company’s V3 and R1 large models on Chinese-made chips to run token services.

The figures in SiliconFlow’s prospectus point to rapid growth: registered users exceeded ten million in just over two years, average daily token throughput increased 12-fold from December 2024 to April 2026, paying customers rose about 292-fold in one year, and its valuation reached RMB 7.7 billion (USD 1.1 billion) in two and a half years.

But the prospectus also shows the cost of that growth. SiliconFlow is selling tokens at a loss. Its gross margin in 2025 was -24%.

Founded in August 2023, SiliconFlow is less than three years old. Its prospectus describes it as China’s leading open and independent token supply platform. In essence, it is an AI inference infrastructure company. Inference refers to the stage when a trained AI model generates output for users.

SiliconFlow aggregates heterogeneous computing power from Nvidia, AMD, Ascend, MetaX, Moore Threads, and others, then uses its proprietary inference engine and computing power orchestration system to turn underlying computing resources into standardized token supply. It delivers that supply through two business lines: public cloud services, including serverless token services and dedicated instances, and on-premises deployment solutions.

According to third-party data cited in its prospectus, SiliconFlow was China’s largest independent ecosystem token supplier by annual token throughput in 2025 and ranked among the top five token suppliers overall.

Selling tokens below cost

SiliconFlow’s revenue growth has been steep. In the first four months after its founding in 2023, it generated just RMB 6,000 (USD 881) in revenue. That rose to RMB 7.346 million (USD 1.1 million) in 2024 and RMB 55.33 million (USD 8.1 million) in 2025. Revenue grew more than sevenfold in a year, but the cost structure beneath that growth shows the underlying pressure on the business.

Its gross margin went from 83.3% in 2023 to 39.4% in 2024, then to -24% in 2025. In 2025, SiliconFlow’s cost of sales reached RMB 68.632 million (USD 10.1 million), equivalent to 124% of its RMB 55.33 million in revenue that year. In other words, for every RMB 1 (USD 0.15) of tokens it sold, it spent RMB 1.24 (USD 0.18) on computing power and other costs. Gross profit was negative, with a gross loss of RMB 13.302 million (USD 2 million) for the year.

This reflects the current price war in AI inference. Prices for large model application programming interfaces, or APIs, have been cut sharply, with major companies pushing token prices lower. As a middle-layer platform, SiliconFlow’s main cost pressure comes from leasing computing power. In 2025, computing resource costs totaled RMB 59.627 million (USD 8.8 million), accounting for 86.9% of cost of sales. The prospectus also says the sharp rise in cost of sales in 2025 was due to the rapid expansion of computing resource consumption to support growing demand for its services.

There were two other major expense lines.

R&D expenses reached RMB 209 million (USD 30.7 million) in 2025, equivalent to 378.1% of revenue that year. Over three years, cumulative R&D spending totaled RMB 284 million (USD 41.7 million), rising from RMB 10.84 million (USD 1.6 million) in 2023 to RMB 64.48 million (USD 9.5 million) in 2024 and RMB 209 million in 2025.

Sales and marketing expenses reached RMB 83.743 million (USD 12.3 million) in 2025, equivalent to 151.4% of revenue.

Revenue was RMB 55.33 million, yet R&D and sales alone consumed nearly RMB 300 million (USD 44.1 million). As a result, net losses from 2023 to 2025 were RMB 12.223 million (USD 1.8 million), RMB 81.915 million (USD 12 million), and RMB 345 million (USD 50.7 million), respectively. In 2025, its net loss margin was -624.4%, and loss per share was RMB 33.33 (USD 4.9). Even on an adjusted basis, after adding back RMB 135 million (USD 19.8 million) in share-based payments and RMB 23.076 million (USD 3.4 million) in interest on redemption liabilities, adjusted net loss in 2025 still reached RMB 187.1 million (USD 27.5 million).

Cash flow from operating activities was also negative for three consecutive years. In 2025, net cash used in operating activities was RMB 172 million (USD 25.3 million). The company has completed seven financing rounds. At the end of 2025, it had RMB 171 million (USD 25.1 million) in cash and cash equivalents, plus another RMB 100 million (USD 14.7 million) in time deposits.

In its 2025 revenue mix, public cloud services generated RMB 29.261 million (USD 4.3 million), accounting for 52.9% of revenue. Of that, serverless token services contributed RMB 14.3 million (USD 2.1 million), while dedicated instances contributed RMB 15 million (USD 2.2 million). On-premises deployment solutions generated RMB 26.069 million (USD 3.8 million), accounting for 47.1%. In other words, SiliconFlow is pursuing two tracks at once: scaled traffic from developers and small and midsize companies, and higher-ticket projects from large enterprises and institutions.

Among its serverless paying customers, the number rose from 2,455 to 716,000, an increase of about 292 times. It has served 13,000 corporate customers, and its platform has supported more than 170 models in total. The number of active serverless users was 5.453 million in 2025 and 1.454 million in the first four months of 2026. From 2024 to 2025, dedicated instance customers increased from seven to 49, with 20 in the first four months of 2026. On-premises deployment customers fell from 28 to 20, with five in the first four months of 2026, but revenue per customer rose significantly.

SiliconFlow does not directly purchase chips. Instead, it mainly leases computing resources through partners. In the use of proceeds section of its prospectus, several major categories state that funds will be used to lease computing resources. Its procurement is highly concentrated among its five largest suppliers, which are computing power and cloud service providers. In 2025, its five largest suppliers accounted for 70.8% of total purchases, while the single largest supplier accounted for 20.4%. In addition, two major business partners were both customers and suppliers.

In its risk disclosures, SiliconFlow said it depends to a large extent on the availability of high-quality open-source AI models. If those models stop being open source, or if licensing terms change, its core supply could be affected.

Judging only by operating data, SiliconFlow is growing quickly. At the end of 2024, the end of 2025, and the end of April 2026, it had 127,000, 9.197 million, and 10.282 million registered users, respectively. From December 2024 to April 2026, average daily token throughput increased about 12 times. It rose 241.1% from 2024 to 2025, then grew another 254.7% from December 2025 to April 2026.

But all that traffic translated into only RMB 55.33 million in revenue in 2025, and it did so at a negative gross margin. In 2025, SiliconFlow’s cost of sales reached RMB 68.632 million, equivalent to 124% of revenue, resulting in an overall gross loss margin of 24%. Between the rapid growth in users, throughput, and paying customers, and the thin, negative-margin monetization that followed, lies a wide gap.

The prospectus also acknowledges that serverless token services are aimed at developers and small and midsize enterprises, and are mostly prepaid, low-priced consumption products. This kind of traffic is large in scale and low in unit price, while supply must be provided close to cost, or even below cost.

Valuation rises to RMB 7.7 billion

From December 2023 to June 2026, SiliconFlow completed seven financing rounds in two and a half years. It has used only about 32% of its net financing proceeds.

Its post-money valuation rose from RMB 280 million (USD 41.1 million) in the angel round to RMB 7.74 billion (USD 1.1 billion) in the Series B+ round. That was a 27.6-fold increase in two and a half years, while cost per share rose 13.4 times.

In 2026 alone, SiliconFlow completed three rounds, with its valuation rising from RMB 3.1 billion (USD 455.2 million) in the Series A+ round to RMB 5 billion (USD 734.2 million) in the Series B round, and then to RMB 7.74 billion in the Series B+ round. Notably, the Series B+ round was completed just one month before it filed its listing application. Based on the Series B+ post-money valuation of RMB 7.74 billion, equivalent to about HKD 8.4 billion (USD 1.1 billion), SiliconFlow just met the listing threshold for pre-commercial companies under Chapter 18C.

In its shareholding structure, the controlling shareholder is Jinhui Yuan, who directly holds 14.35%. Through employee incentive platforms, Yuan controls about 44.48% of voting rights in total. Alibaba Group holds 7.42% and has appointed Chen Yingjie as a director. Huawei holds 4.07%, Sinovation Ventures holds 4.01%, Suzhou’s Glory Ventures, a semiconductor venture capital firm, holds 3.01%, Nanjing Lvyong holds 3.01%, Wang Huiwen holds 2.06%, Puhua Capital holds 1.81%, Guotai Junan holds 1.56%, SenseTime holds 0.64%, Biren Technology holds 0.39%, while Kingdee and Meituan are also shareholders.

Yuan is 45 years old. He holds a doctorate in computer science from Tsinghua University and a bachelor’s degree from Xidian University. From OneFlow to SiliconFlow, he has held multiple roles, including board chair, CEO, general manager, and finance chief.

From 2013 to 2016, Yuan was a principal researcher at Microsoft China, where he led the development of LightLDA, described in the prospectus as the world’s fastest large-scale topic model training system at the time. From 2017 to 2023, he founded and led OneFlow, a deep learning framework company, which was later acquired by Light Year and ultimately folded into Meituan.

In August 2023, Yuan started SiliconFlow with OneFlow’s core team. CTO Juncheng Liu and COO Zhao Zhen were both OneFlow veterans, while deputy general manager Zeng Hua previously served as a vice president at JD.com.

Nonexecutive director Chen Yingjie, 49, is managing director of Alibaba’s strategic investment department. He also serves as a nonexecutive director of MiniMax.

KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Peng Xiaoqiu for 36Kr.

Note: HKD, RMB figures are converted to USD at rates of HKD 7.84 = USD 1 and RMB 6.81 = USD 1 based on estimates as of July 9, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.