When DeepSeek's servers buckled under demand in early 2025, the company that kept the open model reachable was not a cloud giant. It was SiliconFlow (硅基流动), a three-year-old Beijing startup that had just wired DeepSeek's models onto Huawei's Ascend (昇腾) chips.
On June 30, 2026, SiliconFlow filed for a Hong Kong IPO under Chapter 18C, aiming to be the city's first listed "AI Token (词元) factory." Its pitch is deliberately boring and possibly durable: it does not train models and does not sell GPUs. It aggregates heterogeneous compute — Nvidia, AMD, Huawei Ascend, MetaX (沐曦), Moore Threads (摩尔线程) — and turns it into standardized, billable tokens.
What the business actually does
SiliconFlow runs an inference platform called SiliconCloud. Developers call an API; SiliconFlow routes the request across whichever chips are cheapest and most available, optimizes the run with its own SiliconLLM engine, and returns an answer. The neutrality is the product: a model builder or chipmaker is also a potential customer, so SiliconFlow works with all of them instead of competing.
The numbers in the prospectus are stark. Registered users passed 10 million by April 2026, with more than 13,000 enterprise customers and 150-plus models online. Daily token throughput reached 578.5 billion (5785亿) by April 2026, with single-day peaks above one trillion. By Frost & Sullivan's count, SiliconFlow was China's largest independent token supplier by 2025 throughput, top five overall at roughly 1.5% share.
The money and the gap
SiliconFlow was founded in August 2023 by Yuan Jinhui (袁进辉), a Tsinghua PhD and former OneFlow founder. It has completed seven funding rounds; a June 2026 B+ round lifted its valuation to RMB 7.74 billion (≈ US$1.08 billion / HK$8.5 billion). Across three years and seven rounds, that marks a rise from its 2022 valuation of about RMB 228 million (RMB 2.28亿) to RMB 7.74 billion — roughly a 34× increase. In mid-June the company had announced a "more than RMB 2.0 billion (≈ US$280 million)" Series B haul — the largest single round in China's third-party MaaS (Model-as-a-Service) track that year — though the prospectus later disclosed a smaller combined B/B+ equity figure, a gap the company attributed to a not-yet-closed tranche.
The backer list reads like a map of the industry: Alibaba (7.42%, largest external shareholder), Huawei's Hubble unit (4.07%), Meituan (1.05%), plus SenseTime, 360, China Unicom and energy and chip-capital investors. The logic is that a neutral platform can take money and partnerships from rivals who would never fund each other.
The uncomfortable half of the story
Growth is real but unprofitable. 2025 revenue was RMB 55.33 million (≈ US$7.8 million), up 653% year over year, while the net loss widened to RMB 345 million (≈ US$48.6 million). The public-cloud business, now over half of revenue, carries a negative gross margin — every yuan of tokens sold costs about RMB 1.24 in compute. The company warns in its prospectus that it may never achieve profitability.
That is the core bet investors must price: can a neutral aggregator compress costs faster than the hyperscalers (Volcano Engine, Alibaba Cloud, Baidu AI Cloud) that bundle their own chips and models? SiliconFlow's answer is multi-vendor flexibility and a developer-first posture — LangChain, Dify and Cherry Studio ship it as a built-in provider.
For now, the company's worth rests on a bet that neutral, multi-vendor inference becomes the default plumbing of China's AI stack — a position hard to copy but also hard to defend if the hyperscalers close their own cost gap.
What readers can do now
- If you build on Chinese open models (DeepSeek, Qwen, Kimi, GLM), SiliconFlow is a credible multi-chip fallback to a single cloud lock-in; test latency and price against the hyperscalers before committing.
- Treat the RMB 7.74B valuation and "34× in three years" narrative as prospectus framing, not a verdict — the negative gross margin on core cloud revenue is the number to watch.
- For HK-listed-tech watchers, track whether SiliconFlow hits the Chapter 18C "commercialized" revenue threshold (recent fiscal-year revenue of HK$250 million) before year-end 2026.
Honest limitations
Figures come from SiliconFlow's official site, the IPO prospectus as reported by Securities Times (证券时报), United Daily News and Southern Finance (南方财经); I have not seen the filing directly. The B-round "RMB 2.0B+" announcement and the lower prospectus-disclosed equity total are reported as a known discrepancy, attributed by the company to an unclosed tranche — I have not independently reconciled the books. Market-share and throughput rankings cite Frost & Sullivan via the prospectus and are vendor-commissioned. All financials are point-in-time and pre-IPO, so they may be restated.
Key takeaways
- SiliconFlow filed a Hong Kong IPO on June 30 2026 at a RMB 7.74B (≈ US$1.09B) valuation, positioning as China's largest neutral "token factory" with 10M+ users.
- Its model aggregates Nvidia, AMD, Huawei Ascend and domestic chips into standardized API tokens — neutrality lets it partner with rival chipmakers and model builders.
- The open risk: 2025 revenue was only ≈ US$7.8M against a ≈ US$48.6M loss, with negative gross margin on core cloud tokens.