New Stock Outlook | Without tying models or selling computing power, how does the silicon-based flow rely on to capture 1.5% of the traffic throat?

When the big model competition enters the second half of the "landing application" phase, a seemingly low-level but crucial track quietly emerges - token supply. If AI big models are likened to "intelligent factories", tokens are the "raw materials" flowing in these factories, and the token supply platform is responsible for transforming computing power into these raw materials. In the current exponential expansion of AI inference demand, this pipeline itself is becoming a big business. Recently, Beijing SiliconFlow Technology Co., Ltd. (referred to as "SiliconFlow") submitted a listing prospectus to the Hong Kong Stock Exchange, intending to knock on the main board door as an 18C (specialized technology company). It is understood that SiliconFlow positions itself as an "open, independent token supply platform", serving as a technical hub between computing resources and AI models, and converting heterogeneous chips into standardized token services through self-developed inference engines and resource scheduling systems. Founded in August 2023, the company has experienced a rapid increase in valuation from zero to tens of billions in just three years, driven by the explosive growth of token supply as an infrastructure layer demand in the context of AI big models transitioning from training to inference applications. The attractiveness of this story lies in: when big models are no longer scarce, how to efficiently and cost-effectively "consume" model outputs is becoming the core bottleneck for AI application. However, while the company's revenue has increased sevenfold, its net loss has expanded to over 300 million yuan. Is SiliconFlow paving the way for the future or burning money to fill holes? AI inference demand surges, SiliconFlow targets token supply infrastructure The core business of SiliconFlow can be summarized in one sentence: enabling customers to "consume" model outputs more efficiently and cost-effectively. It differs fundamentally from closed ecosystem players who both model and provide computing power. The company does not bind to any specific cloud service provider, model vendor, or application scenario, but aims to become a neutral technological hub. From the product matrix perspective, SiliconFlow has built two main lines of business: public cloud services and on-premise deployment solutions. Among them, public cloud services are further divided into serverless token services and dedicated instances, with the former targeting small developers and startups using a pay-as-you-go shared computing model; and the latter targeting enterprise customers with higher performance, stability, and data isolation requirements by providing dedicated resources on public cloud infrastructure. The on-premise deployment solution deploys the company's software system directly in the customer's private environment to meet the data sovereignty requirements of heavily-regulated industries such as finance and energy. This "light asset" and "heavy service" model essentially penetrates different customer groups: first, quickly accumulate users and brand awareness through low-threshold public cloud services, and then achieve deep value mining through dedicated instances and on-premise deployment solutions. According to the prospectus, as of April 2026, SiliconFlow's registered user base has exceeded 10 million, with a daily token throughput of 57.85 billion times, indicating initial market recognition of its platform. In terms of industry positioning, a key label emphasized by SiliconFlow is "China's largest independent ecological token supplier". According to Frost & Sullivan data, based on the token throughput in 2025, SiliconFlow ranks fourth among all token supply platforms in China, with a market share of about 1.5%, but ranks first in the independent ecological camp. From industry trends, according to Frost & Sullivan data, China's token throughput surged from about 142.5 trillion times in 2024 to about 242.63 trillion times in 2025, with an annual growth rate exceeding 160%, and is expected to reach about 5.32 quintillion in 2030, with a compound annual growth rate of about 638% from 2025 to 2030. Behind this growth is the expansion of AI applications from simple question-answer conversations to more complex scenarios such as multi-step task execution, complex goal decomposition, and multimodal interaction. With the exponential growth in inference calls and context length, token supply is no longer just a technical detail on servers but has become the core bottleneck restricting the efficiency and cost of AI application deployment. However, the competition structure in this market is not stable. Currently, cloud providers, big model companies, and chip manufacturers are all extending their capabilities to the inference layer, making the "middle layer" not a naturally safe zone. While SiliconFlow's emphasis on an open neutral platform strategy helps expand the ecosystem, it also means a lack of natural traffic entry points, requiring continuous reliance on technological efficiency and cost advantages to maintain competitiveness. Revenue surges, losses widen, scale growth has not crossed the profitability inflection point If business structure reflects growth potential, financial data reveals that the company's business model is still in the early stage of validation. It is understood that the company's total revenue skyrocketed from 7.346 million yuan in 2024 to 55.33 million yuan in 2025, an increase of 653.2%. Among them, revenue from public cloud services increased from 1.069 million yuan to 29.261 million yuan in 2025, an increase of over 26 times; revenue from on-premise deployment solutions also increased from 6.277 million yuan to 26.069 million yuan, showing impressive growth rates. However, on the cost side, in 2025, SiliconFlow's sales cost was 68.6 million yuan, significantly higher than the revenue of 55.33 million yuan, with an overall gross profit margin plummeting from 39.4% in 2024 to -24.0% in 2025, meaning that the more the company sells, the more it loses. The collapse of the gross profit margin is driven by two core factors: the first is the deterioration of the revenue structure: although the gross loss rate of public cloud services narrowed from 271.6% in 2024 to 119.0% in 2025, it still represents a real loss-making business, and its revenue share increased from 14.6% in 2024 to 52.9% in 2025, structurally dragging down the overall gross profit level; The second is the heavy upfront investment in computing power costs: to support the explosive expansion of user scale, the company needs to lease a large amount of computing resources in advance, and these resources have low utilization rates in the initial stage, resulting in costs far exceeding the ability to cover current revenue. In 2025, the company's computing resource costs (i.e. leasing fees) reached about 59.6 million yuan, accounting for 86.9% of sales costs, a year-on-year increase of over 14 times, faster than the revenue growth rate. At the bottom of the income statement, in 2025, the company's net loss was as high as 345.5 million yuan, 4.2 times the net loss of 81.9 million yuan in 2024. More intriguing is that the adjusted net loss was 187.1 million yuan, while total revenue for the year was 55.33 million yuan. This means that even after excluding non-cash items such as stock payments, the company still needs to lose more than 3 yuan for every 1 yuan of revenue generated. Among them, research and development expenses are the largest cash outflow, reaching 209 million yuan in 2025, accounting for 378.1% of revenue. Although this ratio has significantly decreased compared to 877.7% in 2024, operating leverage is slowly starting, but the absolute value is still rising rapidly. Therefore, the company is currently in a typical stage of scale inefficiency. From an efficiency perspective, the future improvement of the financial structure depends on three variables: the speed of reducing unit token costs, the magnitude of increasing computing power utilization rates, and the percentage increase of high-margin enterprise customers. If these three form a positive cycle, the company may gradually enter the stage of economies of scale; conversely, if industry competition leads to a continuous decline in token prices, it may fall into a pressure model of "larger scale, thinner profits". Overall, SiliconFlow represents a new type of infrastructure enterprise in the AI inference era. It does not produce models or directly target end applications but aims to establish a standardized intermediate layer between models and applications and capture the growth dividend of AI calls through a token billing system. From industry trends, this track has a clear growth logic, with AI inference demand expansion almost certain, but from a business structure perspective, this model is still in the early validation stage. The key to determining the company's value in the future lies not in the growth rate itself, but in whether the growth can be transformed into a stable unit economic model improvement. For the market, these companies are more like a long-term bet on the AI infrastructure path. In the early stage of the AI inference era, the window is still open, but the business model has not yet converged.