The Chinese AI infrastructure startup's IPO prospectus shows revenue up 653% but public cloud gross margin at 119%, exposing a structural unit economics problem the next 12 months will test.
SiliconFlow wants to be Hong Kong's "AI token factory first stock," but the math inside its own prospectus tells a sharper story than the marketing tag suggests: every yuan of public-cloud compute it sold in 2025 cost more than a yuan to deliver.
The company, founded by ex-OneFlow engineer Yuan Jinhui and staffed largely by alumni of the same circle, filed its Hong Kong Stock Exchange listing application this week, pitching investors on a model that resells AI model access as a utility, measured in tokens rather than seats. The numbers in the HKEX prospectus PDF cover the period through April 30, 2026, and they paint a picture of growth that broke the unit economics under it.
Revenue did what a token-resale business is supposed to do: it exploded. SiliconFlow reported 55.33 million yuan in 2025 revenue, up 653.2% year over year from 7.346 million yuan in 2024, according to QbitAI's read of the filing. Registered users jumped from 127,000 at the end of 2024 to 10.28 million by late April 2026, while enterprise customers passed 13,000 and the platform began supporting more than 170 third-party models. Daily token throughput averaged 578.5 billion in April 2026. Backers include Alibaba, Meituan, SenseTime, NIO, and Zhipu AI across seven funding rounds that valued the company at 7.74 billion yuan.
The margin line tells the opposite story. SiliconFlow's gross margin swung from +39.4% in 2024 to -24.0% in 2025. Inside that aggregate, the public-cloud services line, the segment that overtook on-premise deployment as SiliconFlow's largest revenue source in 2025, carries a gross margin of -119.0%, per the prospectus summary circulated by Jiemian and Sina Finance. In plain terms: the company is paying more than a yuan in compute and platform costs for every yuan a customer hands over for tokens billed through the public cloud.
The pivot explains the math. Public-cloud revenue reached 29.261 million yuan (52.9% of the top line) in 2025, narrowly surpassing on-premise deployment at 26.069 million yuan (47.1%), according to the same prospectus coverage. That is the segment running at negative 119% gross margin. Cost of sales, driven by cloud compute and external technical service fees, climbed from 4.452 million yuan in 2024 to 68.632 million yuan in 2025, an arithmetic that turns a 7x revenue jump into a gross-profit reversal.
The market position SiliconFlow cites in its prospectus comes from consultant Frost & Sullivan. The filing describes the company as the fourth-largest token supply platform in China by 2025 throughput, with roughly 1.5% market share, and the largest among independent platforms not anchored to a single model developer, as reported by QbitAI. That rank is company-disclosed: the underlying Frost & Sullivan report is cited in the prospectus but not independently verified here.
The losses followed the pivot. Net loss reached 345 million yuan in 2025, roughly 4.2 times the 81.915 million yuan loss booked the year before, with adjusted net loss (excluding share-based compensation and similar items) at 187 million yuan. Operating cash outflow for 2025 was 172 million yuan, equating to a monthly burn rate near 14.8 million yuan, per the Sina Finance summary.
What the prospectus does not disclose, and what investors will press for in the upcoming listing hearings, is SiliconFlow's cash position and runway length. Without those figures, it is impossible to say whether the current burn is sustainable through a public listing or whether the public-cloud negative-margin segment is structural. Discussion threads on V2EX have already homed in on the same gap, suggesting retail and developer skepticism will follow the filing onto the order book.
Two questions frame the next six months. First, can SiliconFlow migrate its 13,000-plus enterprise customers upmarket into tiers where compute costs are amortized across higher prices: dedicated deployments, managed inference at premium rates, or private-cloud arrangements that retain the on-premise unit economics? Second, can the company negotiate cloud-provider pricing, or land direct hardware partnerships, fast enough to push public-cloud unit cost below the revenue line before cash runs out?
For now, the prospectus offers scale, not proof of margin recovery. Whether the "token factory" label survives Hong Kong's listing process depends on whether unit economics, not user count, is the story management chooses to lead with next.