Amidst the wave of large model companies seeking listings in Hong Kong, a specialized provider of underlying computing power and API services has officially initiated its IPO journey. SiliconFlow has submitted its listing application to the Hong Kong Stock Exchange (HKEX), aiming to become the first public "AI #Token Factory" on the main board.
Prior to this, the company completed 7 rounds of financing with a post-money valuation of 7.74 billion RMB. Its shareholder lineup is star-studded, featuring strategic industry giants like Alibaba, Meituan, SenseTime, Zhipu AI, NIO, and top-tier AI venture capital funds.
The founding team of #SiliconFlow consists of the original team of OneFlow (an AI Infra pioneer founded by Yuan Jinhui, a disciple of Academician Zhang Bo). After being acquired by Light Year Beyond in 2023 and subsequently integrated into Meituan, Yuan Jinhui chose to start up again, leading his team to establish SiliconFlow. According to the prospectus, the company's core narrative is highly focused: they do not develop base foundation models or consumer-facing apps. Instead, they position themselves as the "Token Factory" for the AI #inference era—packaging heterogeneous computing power, multi-model APIs, and enterprise invocation demands into a stable, metered, and billable Token supply service.
As of April 30, 2026, the registered users of SiliconFlow's platform reached 10.28 million. In April 2026, the average daily Token throughput reached 578.5 billion. The platform has supported over 170 open-source and proprietary models, serving more than 13,000 enterprise customers. According to Frost & Sullivan, based on 2025 annual Token throughput, SiliconFlow ranks as China's fourth-largest Token supply platform with a 1.5% market share, and ranks number one among all independent ecosystem Token supply platforms.
However, alongside rapid business growth, the financial data reveals intense pressure on profitability. In 2025, SiliconFlow's revenue reached 55.33 million RMB, a year-on-year surge of 653.2% compared to 7.35 million RMB in 2024. Simultaneously, its annual net loss widened to 345 million RMB, which is 4.2 times the 81.92 million RMB loss in 2024. The adjusted net loss stood at 187 million RMB, with an operating cash outflow of 172 million RMB and a monthly cash burn rate of about 14.8 million RMB.
Regarding margins, SiliconFlow's gross margin turned negative, dropping from 39.4% in 2024 to -24.0% in 2025, with public cloud service margins plunging to -119.0%. This was driven by a structural shift: public cloud service revenue reached 29.26 million RMB in 2025 (52.9% of total revenue), overtaking on-premise solutions (47.1%) for the first time. The explosive growth in registered users directly escalated back-end server and bandwidth costs. Sales costs jumped from 4.45 million RMB in 2024 to 68.63 million RMB in 2025. Although the public cloud business generates significant volume, it imposes heavy infrastructure cost burdens at this stage.
Even so, the fundamental business logic of the Token Factory has been validated. As the AI application market flourishes and Agentic workflows become more complex, global Token consumption will surge exponentially. Whoever can deliver stable, low-latency, and cheap Tokens at scale will occupy the core infrastructure layer of this new era. The IPO of SiliconFlow will test whether this massive, API-driven infrastructure play can successfully achieve long-term financial viability under public market scrutiny.
[AgentUpdate Depth Analysis] SiliconFlow's prospectus exposes a stark reality in the LLM era: even amidst soaring inference demand, the "Token API" business remains highly vulnerable to computing cost deficits and negative gross margins. However, for the AI Agent ecosystem, independent Token Factories are indispensable infrastructure. As Multi-Agent systems and protocols like MCP (Model Context Protocol) gain traction, Agent-to-Agent communication and autonomous planning will drive Token consumption exponentially (by 10x to 100x). Such workloads demand ultra-low latency and hyper-cheap inference networks. Unlike bundled hyperscaler clouds, an independent token provider offers flexible cross-chip scheduling and cost efficiency. SiliconFlow's ultimate path to profitability lies in software-defined hardware acceleration—using deep compilation optimizations to widen the margin between hardware costs and token pricing, ultimately serving as the "cheap electricity" for the upcoming Agent economy.
