Autumn Sound | Yuan Jinhui's New Company Rushes to HK IPO, Established Less Than Three Years 秋声 | 袁进辉新公司冲港股IPO，成立不到三年

TL;DR

• SiliconFlow filed for a Hong Kong IPO under Rule 18C, positioning itself as China's leading independent token supply platform with over 10 million registered users and a valuation of 7.7 billion RMB.
• The company operates as an AI inference infrastructure provider, aggregating heterogeneous computing power from NVIDIA, AMD, and domestic chips like Ascend to offer standardized token services.
• Despite explosive growth in user base and token throughput (12x increase in daily average tokens), the business model is currently unprofitable, with a negative gross margin of -24% in 2025 due to intense price wars in the AI API market.
• High operational costs are driven by compute leasing (86.9% of sales cost), massive R&D spending (378.1% of revenue), and aggressive sales/marketing efforts, resulting in a net loss of 345 million RMB in 2025.
• The company faces significant risks related to supplier concentration, dependency on open-source models, and the challenge of converting high-volume, low-margin developer traffic into sustainable profitability.

Why It Matters

This case highlights the critical disconnect between rapid user acquisition and unit economics in the current AI infrastructure sector, serving as a cautionary tale for investors and practitioners regarding the sustainability of "burn cash for growth" strategies in commodity AI services. It underscores the strategic importance of proprietary inference engines and heterogeneous compute management as potential moats against the commoditization of LLM APIs. Furthermore, it illustrates the evolving landscape of AI startups leveraging public markets (specifically HKEX Rule 18C) to fund operations despite lacking immediate profitability, signaling a shift in how early-stage AI hardware-software integration companies are valued.

Technical Details

• Infrastructure Architecture: SiliconFlow aggregates heterogeneous computing resources from NVIDIA, AMD, Huawei Ascend, MetaX, and Moore Threads. It utilizes a self-developed inference engine and compute orchestration system to abstract underlying hardware differences into standardized token supply services.
• Service Models: Offers two primary delivery channels: Public Cloud Services (Serverless Token Service and Dedicated Instances) targeting developers and SMEs, and Local Deployment Solutions targeting large enterprises and institutions.
• Scale and Performance: As of April 2026, the platform supports over 170 models. Daily token throughput grew 12-fold from December 2024 to April 2026. Serverless paid customers increased 292-fold year-over-year, reaching 716,000.
• Cost Structure Analysis: In 2025, sales costs were 68.63 million RMB against revenue of 55.33 million RMB. Compute resource leasing accounted for 59.63 million RMB (86.9% of COGS). R&D expenses reached 209 million RMB (378.1% of revenue), and Sales & Marketing expenses were 83.74 million RMB (151.4% of revenue).
• Supply Chain Dependency: The top five suppliers accounted for 70.8% of procurement costs in 2025, with the largest single supplier representing 20.4%. The company relies heavily on leasing rather than owning physical GPU clusters, exposing it to volatility in compute availability and pricing.

Industry Insight

• Profitability Pressure in AI Infrastructure: The negative gross margins indicate that the AI inference market is entering a severe price war phase. Companies acting merely as intermediaries for open-source models without significant proprietary efficiency gains or unique value-add services will struggle to survive unless they can drastically reduce compute costs or differentiate their offerings.
• Strategic Pivot Required: To achieve sustainability, AI infrastructure providers must move beyond simple token reselling. Investment should focus on optimizing inference efficiency through advanced kernel-level optimizations, speculative decoding, or specialized hardware acceleration to widen the gap between compute cost and service price.
• Risk Management in Supply Chains: Heavy reliance on third-party compute leasing creates vulnerability to supply shocks and price hikes. Successful players will likely need to diversify their compute sources further, negotiate long-term fixed-price agreements, or explore hybrid models that include owned capacity for baseline stability while using leased capacity for peak loads.