硅基流动流血冲刺 IPO，Token 工厂是否是一门好生意？

Original Author: Xiaobing

One day at the end of 2023, at the Xi Jia De Dumpling Restaurant downstairs in Tsinghua Science Park, Yuan Jinhui had just sat down when he overheard a neighboring table discussing his company: "OneFlow had good technology, but in the end, it still didn't make money and got acquired."

When he later recalled this scene to LatePost, he said he sometimes wondered if he had set a bad example: correct technical judgment, diligent work, yet still unable to achieve the success everyone recognized.

Two and a half years later, on June 30, 2026, he stood at the entrance of the Hong Kong Stock Exchange with a 347-page prospectus. This time, he wanted to prove the proposition doubted in that dumpling restaurant: good technology can make money.

However, the prospectus first revealed a harsh interim report: in 2025, for every 1 yuan the company earned, its direct costs alone amounted to 1.2 yuan.

What Kind of Company Is This?

SiliconFlow doesn't build large models or manufacture chips. What it does can be summed up in one sentence: It rents computing power from upstream providers, processes it into Tokens, and sells them downstream.

SiliconFlow's role is similar to an oil refinery.

An oil refinery doesn't own oil fields; it buys crude oil, processes it into gasoline, and sells it, profiting from the margin. Similarly, SiliconFlow doesn't own "oil fields." It rents NVIDIA GPUs from cloud providers, along with domestic chips like Ascend, Moffett AI, and Moore Threads. Using its self-developed inference engine, it processes this diverse computing power into standardized Tokens, which it sells to developers and enterprises by volume. The rent is its purchase cost, the Token price is its retail price, and the difference is its profit margin.

The problem is, this margin is currently negative.

In 2024, when the company was still small, the business was profitable, retaining a gross profit margin of 39.4%.

By 2025, revenue surged by 653.2% to 55.33 million yuan, but the gross margin plummeted to negative 24%. The most popular public cloud Token business had a gross margin of negative 119%, meaning for every 100 yuan worth of Tokens sold, the company lost 119 yuan.

Why did this happen?

On one hand, procurement costs were high. To handle potentially massive user surges, the company had to reserve large amounts of computing power in advance. In 2025, the cost of sales soared from 4.452 million yuan the previous year to 68.632 million yuan, exceeding total annual revenue, and the reserved computing power was not fully utilized.

On the other hand, retail prices were being brutally slashed. Major tech companies repeatedly cut prices to attract developers, reducing the cost per thousand Tokens for some mainstream models by over 90%. In May, DeepSeek announced a permanent 75% price cut for V4-Pro, followed by Tencent Cloud, with some reductions as steep as 97.5%.

More problematic is that SiliconFlow has no control over either procurement or retail prices. Computing power rental fees are set by upstream cloud providers, while Token prices are determined by price wars among the giants. When Alibaba and ByteDance cut prices, they subsidize the cost with profits from other business lines to capture market share. For independent players like SiliconFlow, which solely rely on this margin, every price cut announcement is equivalent to watching their profits being eroded.

Torrential Traffic, Torrential Bills

The company's most glorious moment perfectly illuminated this very logic.

On February 1, 2025, DeepSeek took the world by storm. Its official servers were overwhelmed, leaving millions of eager users unable to access the service.

SiliconFlow seized this window. Teaming up with Huawei Cloud, it became the first to launch a full-capability R1 and V3 service based on Ascend chips, absorbing the overflow of users. Combined with a user acquisition strategy offering "14 yuan for registration and another 14 yuan for referrals," website traffic surged nearly 40 times. Registered users grew from 127,000 at the end of 2024 to 10.28 million by the end of April. The platform processed 578.5 billion Tokens daily, peaking at over 1.07 trillion Tokens in a single day. Based on its 2025 throughput, it has already become China's largest independent Token supply platform.

The cost of this success was recorded on other pages: a net loss of 345 million yuan in 2025, 4.2 times that of the previous year. Even excluding accounting factors like equity incentives, the adjusted loss was still 187 million yuan. Operating activities consumed 172 million yuan in cash over the year, burning an average of about 14.8 million yuan per month. Since its founding in August 2023, the company has accumulated losses of approximately 440 million yuan over three years.

For a normal factory, a surge in orders is great news. For a factory operating at a negative margin, a surge in orders means only one thing: the speed at which money is lost also surges.

What Makes It Worth 7.74 Billion Yuan?

At this point, you might ask: For a business losing money like this, how did it secure seven rounds of funding in three years, with its valuation soaring from 280 million yuan in the angel round to 7.74 billion yuan? Why are Alibaba, Meituan, Huawei Hubble, SenseTime, Zhipu AI, and Sinovation Ventures all on its shareholder list?

From the perspective of the computing power landscape, it still holds two key cards.

The first card is neutrality. One of the biggest fears for developers is locking their entire business into a single major cloud provider, making future migration costly and difficult. A Token platform not owned by any tech giant naturally inspires trust. The fact that giants' funds appear simultaneously in its shareholder list precisely indicates that everyone needs this neutral ground controlled by no single entity.

The second card is the true narrative game-changer: The Substation for Domestic Computing Power.

The restricted supply of NVIDIA chips is a reality for the entire Chinese AI industry. Domestic chips like Ascend, Moffett AI, and Moore Threads are stepping up, but each has its own architecture and quirks, creating a high barrier for developers getting started.

What SiliconFlow does is translate these diverse domestic chips into uniform, easy-to-use standard Tokens. The fact that DeepSeek's full-capability version runs smoothly on Ascend chips is a testament to this translation capability.

Major tech companies might not be willing to optimize for their competitors' chips, and the chip manufacturers themselves may lack the expertise. Yet, the entire domestic computing power ecosystem depends on this capability. It's like a power substation: upstream power plants can change, downstream consumers can change, but the substation's position remains the most stable.

In June, the company bought back all intellectual property rights from its former employer, OneFlow, precisely to strengthen this layer. This is also the confidence behind its decision to list on the Hong Kong Stock Exchange under Chapter 18C (a special channel for pre-revenue tech companies): The prospectus, citing Frost & Sullivan's forecasts, states that China's Token supply market grew 16 times in scale from 2024 to 2025 and is expected to expand at an annual rate of 638.3% over the next five years.

Is a Token factory a good business? The prospectus offers a split answer: Selling standard Tokens based on the margin is a tough business right now. Becoming the substation for domestic computing power could be a business opportunity of the era. SiliconFlow's bet is whether it can survive on the traffic from the former long enough to reap the rewards of the latter.

Persistent Despite Repeated Defeats

To understand why this company dares to make such a bet, you have to look at Yuan Jinhui.

His journey has been marked by "almost succeeding" at nearly every step. He earned his bachelor's degree from Xidian University, went directly to a PhD program at Tsinghua University in 2003 as the top student in the Computer Science department, studied under Academician Zhang Bo, and spent nearly a decade at Tsinghua as a PhD student and postdoc. His original life script was to become a professor, but choosing the niche interdisciplinary field of computational neuroscience meant he didn't secure a faculty position. He was inches away from the podium but never stood on it.

When he left academia, the golden age of the Chinese internet had already passed him by. He moved through Youdao and 360, and later at Microsoft Research Asia, he developed a core system adopted by companies like Kuaishou, earning the President's Award.

In 2017, he started a company based on a judgment that almost no one believed at the time: future models would be too large for existing frameworks, and the underlying system had to be rewritten. That company was OneFlow. The subsequent wave of large models proved his bet was right, but the company didn't survive to see the harvest season. In 2023, it was sold to Wang Huiwen's Guangnianzhiwai (Light Years Away) for a $100 million valuation. A few months later, Guangnianzhiwai was fully acquired by Meituan.

He made the right call, he survived, but he still didn't win. That's what made the "comment" in the dumpling restaurant truly sting.

At that time, Yuan Jinhui had lucrative job offers from major tech companies, and his team members had respectable options too. His choice was to start over for the third time with 35 of the 40 people in his team. He explained his reasoning externally: within a large, diversified company, an AI framework might not be a high priority; for this team, it was everything. The new company was named SiliconFlow – 'silicon' referring to chips, and 'flow' referring to the software that makes computing power flow, echoing OneFlow as if continuing an unfinished story. In the WeChat Moments post announcing his new venture, he wrote: "The past 15 years haven't been smooth, but we persist despite repeated defeats."

China's tech industry has never lacked for storytellers. What's truly scarce are people like this – who correctly judged the direction but lost the ending, yet are still willing to return to the table. SiliconFlow's prospectus may not be pretty, but behind it stands an engineer who has spent two decades repeatedly testing the same belief: good technology deserves commercial success.

Regardless of the final IPO market valuation, good luck to SiliconFlow, and may Yuan Jinhui finally finish this story. Persisting despite repeated defeats, spring will eventually come.