Preface:

Silicon-Based Flow recently announced the completion of its Series B+ and Series C financing rounds, raising cumulative equity financing of nearly 2.9 billion yuan by 2026. This milestone comes just 82 days after the company submitted its prospectus to the Hong Kong Stock Exchange (HKEX). Founded 34 months ago, the company has completed seven funding rounds, achieving a post-money valuation of 7.74 billion yuan—narrowly exceeding the HKEX’s Chapter 18C market capitalization threshold of 8 billion Hong Kong dollars for pre-revenue firms.

Author | Fang Wensan
Image Source | Internet

653% Growth: The ‘Water Seller’ Rides the AI Wave

Yuan Jinhui’s resume reads like a technical manifesto. A Tsinghua University Ph.D. mentored by Academician Zhang Bo, he led deep learning framework research at Microsoft Research Asia before founding a startup acquired by Guangnian Beyond (later integrated into Meituan). Fifty days after the acquisition, Yuan rebuilt his team to launch Silicon-Based Flow, focusing on optimizing computing power efficiency for software applications. The company generated just 6,000 yuan in revenue during its first four months.

The turning point arrived before the 2025 Spring Festival. Liang Wenfeng urged Yuan to deploy 20 H800 servers for DeepSeek V3, incurring monthly computing costs of 6 million yuan—with no guarantee of full utilization. Yuan hesitated. After DeepSeek’s explosive success, he publicly admitted his mistake. Over the Lunar New Year holiday, his team worked around the clock to launch full-scale DeepSeek R1 and V3 services. Website traffic surged 40-fold, temporarily outpacing Tencent Cloud and Alibaba Cloud.

The numbers tell the rest: Revenue jumped from 7.346 million yuan in 2024 to 55.33 million yuan in 2025. Daily token throughput skyrocketed from 47.8 billion to 578.5 billion, peaking over 1 trillion. The platform now serves 10 million registered users, 13,000 enterprise clients, and hosts 170+ models.

However, this growth wasn’t solely company-driven. China’s token supply market expanded 1,602.6% in 2025, with enterprise-level MaaS (Model-as-a-Service) calls increasing 16-fold. As the tide rose, even modest players appeared formidable.

MaaS Achieves Scale, But Profitability Remains Elusive

The prospectus reveals a stark reality: Gross margins collapsed from 83.3% in 2023 to 39.4% in 2024, then plunged to -24.0% in 2025. Sales costs (68.632 million yuan) surpassed annual revenue (55.33 million yuan), with computing power leasing alone costing 59.627 million yuan (86.9% of sales expenses).

For every 1 yuan in token revenue, the company spent 1.24 yuan on computing and operational costs. The prospectus warns: “Cost growth will continue to outpace revenue gains.”

Customer acquisition drove much of the loss. In 2025, sales and marketing expenses hit 83.7 million yuan, including 54.213 million yuan in free token subsidies (64.7% of the total). These incentives attracted serverless token clients who generated just 14.3 million yuan in annual revenue—less than 30% of the subsidy value. Combined with 209 million yuan in R&D investment, the company reported a 345 million yuan net loss and 172 million yuan in operating cash outflows, burning 14.8 million yuan monthly.

The flip side: On-premises deployments maintained gross margins between 82.5% and 92.4%, with software licensing, implementation, and maintenance fees providing steady cash flow.

Silicon-Based Flow’s model resembles a funnel: Developers trial services on public clouds at low cost, then upgrade to dedicated instances or on-premises solutions as usage scales and data security/compliance needs rise. Public clouds drive traffic; on-premises deployments generate profits. Amid the 653% revenue surge, the less flashy on-premises business proved most lucrative.

The profit formula is clear: Profit = Token Price × (Computing Cost ÷ Inference Efficiency × Utilization Rate). For scale to translate into profit, the denominator (costs) must shrink faster than the numerator (revenue). In the recent price war era, the numerator plummeted at an alarming rate.

Neutrality: A Valuable Asset, But a Risky Gambit

AI startups typically align with a single cloud provider, trading loyalty for resources. Silicon-Based Flow defied convention, refusing to bind to any cloud, chip, or model vendor from day one. Yet it secured funding from internet, chip, computing power, energy, and telecom sectors.

Alibaba, Huawei Habo, Meituan, SenseTime, Zhipu AI, Ctrip, Jinko, and China Unicom affiliates—competitors and partners alike—appear on its shareholder list. After seven funding rounds, its valuation soared from 280 million yuan (angel round) to 7.74 billion yuan—a 33-fold increase in 2.5 years.

The prospectus reveals that in 2025, Alibaba and Huawei Habo supplied 19.7 million yuan and 12.62 million yuan in computing power, respectively (nearly 20% of procurement). Yet Alibaba Bailian and Huawei Cloud MaaS are direct public cloud competitors. Some subsidy funds flowed directly into rivals’ pockets.

According to IDC, ByteDance’s Volcano Engine led the 2025 MaaS market with 40%+ share, Alibaba Cloud held ~30%, and Silicon-Based Flow ranked fourth with 1.5%.

Giants can afford MaaS losses, cross-subsidizing with cloud storage, databases, and legacy contracts. Vertically integrated model vendors could devour the middle layer from both ends.

Neutrality granted Silicon-Based Flow supply chain access but no pricing power. Its true moat lies in inference engineering—squeezing more tokens from the same computing power. Yet engineering capabilities depreciate rapidly; each new framework release by competitors requires rebuilding the moat. In business, neutrality sometimes means freedom—and sometimes means having no allies.

Behind Each Token: A ‘Factory’ Heavier Than Models

Interestingly, Silicon-Based Flow avoided positioning itself as a “large model company.” The prospectus defines it as an “open and independent token supply platform”—akin to a power plant and grid operator in the AI era. It connects upstream GPUs, NPUs, and diverse computing clusters; deploys middleware like inference engines, resource schedulers, model adapters, and load balancers; and delivers model capabilities to developers via APIs, monetizing through tokens, instances, or on-premises solutions.

In production environments, customers care about reliability and cost. At this layer, tokens acquire an industrial, utility-like quality. HKEX filings explicitly define tokens as basic data units processed during inference, serving as the primary metric for computing consumption and billing.

Founder Yuan Jinhui’s background underscores this strategy. As a principal researcher at Microsoft China (2013–2016) leading LightLDA, and founder of OneFlow in 2017, he has long focused on deep learning frameworks, distributed computing, and hardware utilization. The company’s DNA prioritizes “maximizing machine utilization, speed, and cost-efficiency” over model parameter arms races.

This business operates within an “independent ecosystem.” Large cloud vendors control segments of chips, clouds, models, and application gateways, benefiting from resource abundance and short chains. Independent platforms bet on “horizontal connectivity”: not tying to a single cloud, chip, or model, nor requiring clients to use only one model.

Frost & Sullivan divides the market into “closed ecosystems” and “independent ecosystems.” By 2025 token call volume, China’s market totaled ~2,426.3 trillion tokens, with Silicon-Based Flow contributing ~36.7 trillion tokens (1.5% share), ranking fourth overall but first among independent platforms.

Silicon-Based Flow is betting on the complexity arising from proliferating models and diversifying chips. AI infrastructure follows a counterintuitive rule: The richer the choices, the more developers need someone to clean up the mess. Model diversity is romantic, but in data centers, it translates into compatibility tickets.

Conclusion:

Future financial reports must prove this business deserves the label “factory.” The term implies an iron law: Inputs must cost less than outputs. The HKEX gong merely moves this law from the boardroom to the trading floor.

Partial References:
• arXiv: “Serving Large Language Models on Huawei CloudMatrix384”
• 36Kr: “Another AI ‘Shovel Seller’ IPO Backed by Alibaba, Huawei Habo”
• Phoenix Finance: “Silicon-Based Flow Secures Two More Rounds: 55.33M Revenue Last Year, Nearly 2.9B Raised This Year”
• IPO Radar: “-24% Gross Margin! Silicon-Based Flow Rushes for HKEX ‘AI Token First Stock’”