
Produced by Bullet Finance
Author | Qiong Ma
Editor | Egg Chief
Art Editor: Qianqian
Reviewed by | Song Wen
As tokens are increasingly viewed as the 'utilities of the AI era'—akin to water, electricity, and coal—the capital race around the inference compute layer is rapidly intensifying. Recently, SiliconFlow submitted its prospectus to the Hong Kong Stock Exchange, aiming to become the first listed 'AI token factory.'
This young company, founded in August 2023, has completed seven funding rounds at an astonishing pace, setting a record for the largest single investment in China’s third-party MaaS (Model-as-a-Service) sector in 2026. Backed by investors including Alibaba and Meituan, it now boasts a post-money valuation of RMB 7.74 billion.
Behind this high-growth capital narrative lies a harsh reality: cumulative net losses of RMB 440 million over three years and a gross margin that turned negative in 2025, plunging to -24%. The negative gross margin trend has drawn particular market attention.
As an independent middleware provider, SiliconFlow offers broad compatibility with diverse chips such as NVIDIA and Ascend, supporting 170 mainstream models including DeepSeek. However, despite being China’s fourth-largest token supply platform, its 1.5% market share remains dwarfed by the top three players.
With industry leaders wielding significant pricing power and entrenched ecosystem barriers, third-party platforms like SiliconFlow face severely constrained room to operate. Its strategy of using subsidies to acquire users may even accelerate its own cash burn.
Meanwhile, price volatility further complicates SiliconFlow’s position. The company focuses on optimizing domestic computing resources and delivering private deployments for government and enterprise clients—areas deemed unattractive or uneconomical for big tech firms. Striking a sustainable growth rhythm between 'avoiding giants' and achieving 'economies of scale' may be one of the key tests for its IPO narrative.
Continuous capital injections have provided ammunition for this突围 (breakout), but also set a tight time window. As investor patience races against its burn rate, can its business model—as an independent AI infrastructure provider—ultimately withstand the test of profitability?
1. Explosive Growth Masks Persistent Profitability Challenges
The so-called 'token factory' refers to a MaaS infrastructure platform that standardizes underlying GPU or domestic chip computing power through compute orchestration and inference engines, then charges customers on a per-token basis.
According to its prospectus, SiliconFlow positions itself as a 'middleware layer' in AI inference infrastructure. By leveraging its proprietary inference engine and compute orchestration system, it converts raw computing power into standardized token supply—hence the analogy to a 'token factory.'
In terms of its business model, the company primarily follows two implementation paths:
First, private deployment services (i.e., on-premises solutions), which provide software solutions to enterprises that have already procured computing power, enabling them to instantly upgrade their internal computing resources into token factories;
Second, public cloud services offering ready-to-use token services to technology-focused enterprises.
In 2025, public cloud services accounted for 52.9% of revenue, while on-premises solutions contributed 47.1%, making public cloud the largest revenue source.

(Photo / Prospectus)
According to the prospectus, the company’s revenue surged from RMB 7.35 million in 2024 to RMB 55.33 million in 2025, representing a year-over-year increase of 653%. As of April 30, 2026, the platform had surpassed 10.28 million registered users, with an average daily token throughput of approximately 578.5 billion tokens.
This explosive growth was driven by a 1,602.6% year-over-year expansion in China's token supply market and the company’s rapid scaling across both its public cloud services and on-premises solutions segments.
Measured by token throughput, China’s token supply market grew by 1,602.6% from 2024 to 2025 and is projected to reach approximately 532 quintillion tokens by 2030, reflecting a compound annual growth rate (CAGR) of 638.3% from 2025 to 2030.
Notably, SiliconFlow is classified as a non-commercial company under Hong Kong Stock Exchange Chapter 18C. The company expects to meet the Chapter 18C revenue threshold for commercial companies by the end of 2026—namely, achieving HKD 250 million in revenue in its most recent fiscal year.At current exchange rates, this equates to approximately RMB 216 million, implying that SiliconFlow must achieve a minimum revenue growth rate of roughly 293% in 2026.
However, this industry-wide explosive growth has not translated into improved profitability.
SiliconFlow reported losses of RMB 12.22 million, RMB 81.92 million, and RMB 345 million for the periods of 2023 (from inception to year-end, spanning four months), 2024, and 2025, respectively,with cumulative losses over the three years amounting to nearly RMB 440 million.Even after excluding non-cash items such as share-based compensation, the company’s adjusted net loss for 2025 still reached RMB 187 million.

(Photo / Prospectus)
A sharp decline in gross margin more vividly reveals the core issue: the company’s overall gross margin fell from 83.3% in 2023 to 39.4% in 2024, and further turned negative at -24% in 2025.
In particular, the public cloud services segment posted a gross margin as low as -119%, severely dragging down the overall gross margin. Even the on-premises deployment solutions business, with a gross margin of 82.5%, could not offset this drag.

(Photo / Prospectus)
The compression in overall profitability primarily stems from substantial costs incurred through leased computing power. According to the prospectus, SiliconFlow’s sales and marketing expenses surged by 1,210% from RMB 6.39 million in 2024 to RMB 83.74 million in 2025, equivalent to 151.4% of its revenue.
Of this amount, RMB 54.213 million(accounting for 64.7%)was spent on distributing free token vouchers—a strategy essentially trading subsidies for user acquisition.
Moreover, high R&D expenditures have further intensified cash flow pressure. In 2025, the company’s R&D expenses reached RMB 209 million, 3.78 times its annual revenue, primarily allocated toward iterating its proprietary inference engine and computing orchestration system.
Although the company stated that continued investment in R&D is aimed at supporting long-term growth and scaling, the combination of high R&D expenses and a business structure featuring negative gross margins means losses are unlikely to reverse in the near term.
Regarding questions such as the negative gross margin, SiliconFlow told 'Bullet Finance' that its current overall positive gross profit primarily comes from dedicated instances and on-premises deployment services. The main source of losses stems from its public cloud Serverless API business serving small and medium-sized users, which is currently in a strategic loss phase. As losses from the Serverless API business continue to narrow, the company as a whole will achieve breakeven.
2. How credible is its claim of being the largest independent token supplier?
Currently, the token supply market has largely formed two business models: independent ecosystems and closed ecosystems.
In an independent ecosystem,token supply platforms typically do not tie themselves to any specific cloud provider, model, or use case. Instead, they offer multi-model, multi-infrastructure token supply services across deployment environments by connecting diverse computing resources, multiple models, various computing centers, and enterprise customer demands.
In a closed ecosystem,cloud service providers, large tech companies, or model developers leverage their own cloud resources, proprietary models, computing infrastructure, and application scenarios to integrate resources, enable model invocation, and deliver end-to-end services within their own ecosystems.
SiliconFlow positions itself as an 'independent ecosystem token supply platform,' unaffiliated with any single cloud provider, chipmaker, or model vendor, instead acting as a system software middleware layer connecting upstream computing power, midstream models, and downstream applications.
According to its prospectus, the company’s platform already supports a total of 170 models cumulatively, including mainstream advanced models such as DeepSeek, GLM, Kimi, and MiniMax; it ranks first globally in cross-chip and multi-model compatibility, supporting leading international chips (such as NVIDIA and AMD GPUs) as well as major domestic AI chips (including Huawei Ascend, Muxi, and Moore Threads).

(Photo / Prospectus)
According to Frost & Sullivan,By token throughput in 2025, SiliconFlow is China's largest independent ecosystem token supplier, with a market share of 1.5%.

(Photo / Prospectus)
However, when including closed-ecosystem vendors that rely on their own cloud resources and models, SiliconFlow ranks only as China's fourth-largest token supply platform, with a market share that lags far behind the top three—by more than an order of magnitude.
According to an IDC report, in the first half of 2025, Volcano Engine led the market with a 49.2% share, followed by Alibaba Cloud at 27% and Baidu Intelligent Cloud at 17%, collectively capturing approximately 93% of the entire market. (Editor’s note: This data covers only the public cloud MaaS market.)
In SiliconFlow’s view, this positioning holds unique value: closed ecosystems tend to lock customers into their proprietary computing power and models, whereas SiliconFlow maintains a neutral and open stance. Amid China’s diversified domestic computing landscape and a competitive yet cooperative industry structure, this neutrality fosters greater trust and positions SiliconFlow as a hub connecting chipmakers, model developers, and enterprise clients.
However, customer concentration remains a concern.In 2023, 2024, and 2025, the top five customers accounted for 100%, 85%, and 45% of revenue, respectively, while the single largest customer’s share declined from 83.3% to 61.1% and further to 13.6%.
Although concentration is declining, the top two customers still contributed 14% and 13% of revenue in 2025, respectively. In an industry environment where token pricing continues to fall, large customers’ bargaining power shows a sustained upward trend, limiting the company’s ability to pass costs upstream.
Further challenges stem from price wars.According to data from the China Academy of Information and Communications Technology (CAICT), the average price of domestic large-model APIs has dropped by over 90% since 2023, with leading large-model providers having slashed API prices more than ten times cumulatively.
Represented by DeepSeek, in June 2026, DeepSeek V4 Pro permanently reduced its pricing, with the new rates set at just one-quarter of the original price.
As a midstream player, SiliconFlow faces dual pressures: it cannot control upstream computing power costs and is subjected to continuous downward pricing pressure downstream, resulting in squeezed profit margins from both ends.
Regarding responses to market competition strategies, SiliconFlow’s senior management publicly stated last year thatthe company focuses on areas where large tech firms currently show little interest or find uneconomical to pursue,particularly those requiring deep optimization for domestic computing hardware, serving medium-to-large enterprises and government-affiliated clients who are cost-sensitive and require on-premises deployment. The company also aims to expand its product adoption from leading clients to a broader base of mid- and long-tail enterprises with scalable application needs.
Positioning itself in segments overlooked by big tech players is undoubtedly a rational choice to carve out survival space. Yet, this differentiated path—designed to avoid direct competition with industry giants—might it trap the company long-term in a narrow niche characterized by high investment, low gross margins, and slow growth?
In response to these concerns, SiliconFlow told Bullet Caijing that in the field of domestic chip compatibility, SiliconFlow offers a rich portfolio of computing solutions and is the only heterogeneous computing platform compatible with both NVIDIA and mainstream domestic chips, meeting diverse downstream client demands. Its on-premises deployment solutions have already penetrated high-value sectors such as internet services, finance, energy, and transportation, with landmark customer cases successfully implemented. This business primarily delivers standardized software solutions, maintains positive gross margins, and does not suffer from diseconomies of scale.
3. The Two Sides of Deep Vertical Integration Across the Entire Industrial Chain
Leveraging its neutral positioning, SiliconFlow has carved out a unique strategic collaboration path, securing investments from capital across the entire industrial chain—including internet platforms, chipmakers, computing infrastructure providers, energy companies, and telecom operators.
Judging by its fundraising pace, SiliconFlow has attracted strong investor interest, completing seven funding rounds in total since launching its angel round in December 2023.
On June 16, 2026, the company officially announced the completion of its Series B financing round exceeding RMB 2 billion, setting a record for the largest single funding round in China’s third-party MaaS (Model-as-a-Service) sector since 2026. Following this round, the company’s post-money valuation reached RMB 7.74 billion.
However, Bullet Caijing noted that according to the prospectus submitted by SiliconFlow to the Hong Kong Stock Exchange on June 30, the amounts raised in its Series B and Series B+ rounds were RMB 520 million and RMB 740 million respectively, totaling RMB 1.26 billion.

(Image / SiliconFlow Prospectus)
In fact, Chapter 18C of the Hong Kong listing rules explicitly requires uncommercialized companies to have a market capitalization of no less than HKD 8 billion. After announcing the 'completion of over RMB 2 billion in Series B financing,' the company rapidly elevated its perceived market valuation to RMB 7.74 billion (approximately HKD 8.9 billion), just meeting the listing requirement.
It should be noted that SiliconFlow’s prospectus disclosed a post-money valuation of RMB 5.02 billion as of April 30, 2026, from its Series B round—still below the listing threshold. However, after its Series B+ round in early June, its valuation quickly rose to RMB 7.74 billion, precisely hitting the required market cap, prompting it to swiftly file its listing application with the Hong Kong Stock Exchange on June 30—a move that has sparked external discussion.
The prospectus shows that SiliconFlow’s shareholder list reads like a 'who’s who' of industry leaders: Alibaba, Huawei Hubble, Zhipu AI, Meituan, SenseTime, Trip.com, JinkoSolar, China Unicom, Biren, Nio Capital, Kingdee, along with top-tier financial investors and state-owned institutions such as GGV Capital, Huakong Fund, and China Development Bank Financial Leasing.
This full-stack investment brings not only capital but also deep ecosystem synergies in terms of application scenarios, computing power, models, and market access.
For example, Biren Technology stated that it looks forward to deep collaboration with SiliconFlow on chip compatibility, inference acceleration optimization, and large-scale deployment of computing clusters to jointly build a high-performance token factory.
Moreover, in February 2025, SiliconFlow became the first company in the industry to launch the full-featured DeepSeek R1/V3 model powered by Huawei Ascend cloud computing, successfully pioneering the deployment of DeepSeek on domestic chips. This milestone quickly drew industry attention to SiliconFlow and positioned it as a key promoter of Huawei’s Ascend ecosystem.
However, this neutral business model also faces challenges. Some investors are themselves potential competitors. Shareholders such as Alibaba Cloud, SenseTime, and Zhipu AI are also active in the MaaS space, and their divergent strategic interests could influence long-term decision-making. For instance, Alibaba operates its own Bailian platform, SenseTime has its proprietary ‘Wanxiang’ infrastructure, and Zhipu runs its own API business—creating direct competitive overlaps with SiliconFlow.
Another noteworthy detail is the overlap between customers and suppliers.
Customer I (also Supplier J) was SiliconFlow’s second-largest customer in 2025 and its fifth-largest supplier in the same year, with both entities belonging to the same group. The company primarily provides AI model chip adaptation services to this customer, while the customer mainly supplies computing resources to the company through different entities within its group.
In 2025, SiliconFlow generated RMB 7.1 million in sales from Customer I, accounting for 12.8% of the company’s total revenue that year. During the same period, the company’s procurement from Supplier J amounted to RMB 12.6 million, representing 7.6% of its total procurement expenditure.

(Photo / Prospectus)
This dual role—where an investor is simultaneously a major customer and a major supplier—is not uncommon in business, but it often implies the company’s multifaceted dependence on the same counterparty.
Regarding questions about its operational independence, SiliconFlow told Bullet Finance that it retains full autonomy over its day-to-day business decisions and maintains a complementary, win-win ecological partnership with its strategic investors focused on expanding the overall market. SiliconFlow emphasized that its open and neutral positioning will continue in the long term.
For now, SiliconFlow’s future inevitably hinges on one practical challenge:Can economies of scale outpace its high R&D and computing costs?
Against the backdrop of tokens becoming the 'utilities'—the electricity, water, and coal—of the AI era, SiliconFlow, as an independent infrastructure provider, is well-positioned to occupy a critical niche in the ecosystem—provided it survives its current cash-burning phase and defends its market share amid intense competition from tech giants.
Ultimately, SiliconFlow’s professed neutrality must be validated by a sustainable unit economics model, rather than resting solely on narrative claims about its ecosystem position.
*The featured image in this article is sourced from SheTu.com under the VRF license.
Risk Disclaimer: The above content only represents the author's view. It does not represent any position or investment advice of Futu. Futu makes no representation or warranty.Read more
Comments
to post a comment
