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, in the Xijiade dumpling restaurant downstairs at Tsinghua Science Park, Yuan Jinhui had just sat down when he overheard a nearby table discussing his company: "OneFlow had great technology, but in the end, they still couldn't make money and got acquired."

Recalling this scene to *LatePost* later, he said he sometimes wonders if he set a bad example: making the right technical judgments and working diligently, yet still failing to achieve the kind of success that earns widespread recognition.

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 aims to prove the proposition that was questioned in that dumpling restaurant: good technology can make money.

However, the prospectus first presented a harsh interim result: in 2025, for every 1 yuan of revenue the company earned, 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, processes it into tokens, and sells them downstream.

SiliconFlow's role is similar to that of an oil refinery.

An oil refinery doesn't own oil fields itself; it buys crude oil, refines it into gasoline, and sells it, earning a profit from the spread. Similarly, SiliconFlow doesn't own "oil fields." It rents NVIDIA GPUs from cloud providers, along with domestic chips like Ascend, MXC, and Moore Threads. Using its self-developed inference engine, it processes these varied computing resources into standardized tokens, selling them by volume to developers and enterprises. The rental cost is its purchase price, the token price is its retail price, and the spread is its profit margin.

The problem is, this spread is currently negative.

In 2024, when the company was still small, this business was profitable, with a gross margin of 39.4% from the purchase-sale spread.

By 2025, revenue skyrocketed 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, the purchase price is high: to handle the potentially massive influx of users at any moment, the company must reserve large amounts of computing power in advance. In 2025, sales costs surged from 4.452 million yuan the previous year to 68.632 million yuan, exceeding total annual revenue, and much of this reserved computing power remained underutilized.

On the other hand, the retail price is being slashed aggressively: major tech companies, vying for developers, have repeatedly cut prices. The cost per thousand tokens for some mainstream models has been reduced by over 90%. In May of this year, DeepSeek announced a permanent 75% price reduction for V4-Pro, with Tencent Cloud following suit and cutting prices by up to 97.5%.

More troubling is that SiliconFlow has no control over either the purchase price or the retail price. The rental cost of computing power is set by upstream cloud providers; the selling price of tokens is dictated by the price wars among large companies. When Alibaba and ByteDance cut prices, they subsidize the cost with profits from other businesses to capture market share. For an independent player like SiliconFlow, which relies solely on the spread, every price cut announcement directly erodes its profit margin.

A Torrent of Traffic, a Torrent of Bills

The company's most glorious moment perfectly illuminated the dynamics described above.

On February 1, 2025, DeepSeek went viral globally. Its official servers were overwhelmed, leaving millions of users unable to access the service.

SiliconFlow seized this window. Partnering with Huawei Cloud and utilizing Ascend chips, it was the first to deploy full-version R1 and V3 services, successfully capturing 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 this year. The platform processed 578.5 billion tokens daily, with a peak of over 1.07 trillion tokens in a single day. Based on 2025 throughput, it has become the largest independent token supply platform in China.

The cost of this success is recorded in other pages of the prospectus: the net loss for 2025 was 345 million yuan, 4.2 times that of the previous year. Even excluding non-cash items like share-based compensation, the adjusted net loss was still 187 million yuan. Operating activities consumed 172 million yuan in cash, averaging about 14.8 million yuan burned 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 fantastic news. But for a factory operating with a negative spread, a surge in orders means only one thing: the rate of losses also surges.

Why is it Worth 7.74 Billion?

At this point, you might ask: how can such a loss-making business raise funds across seven rounds in three years, with its valuation soaring from 280 million yuan at the angel round to 7.74 billion yuan? Why are Alibaba, Meituan, Huawei Hubble, SenseTime, Zhipu AI, and Sinovation Ventures all crowded onto its shareholder list?

In the eyes of the computing power industry, it still holds two significant cards.

The first card is neutrality. One of the biggest fears for developers is locking their entire business onto a single major cloud provider, making migration prohibitively expensive. A token platform that doesn't belong to any giant is naturally reassuring. The fact that funds from multiple giants appear on its shareholder list precisely indicates that all parties need this neutral middle ground that no one controls.

The second card is the true royal flush narrative: a substation for domestic computing power.

The constraints on NVIDIA chip supply are a stark reality for the entire Chinese AI industry. Domestic chips like Ascend, MXC, and Moore Threads are stepping up, but each has its own architecture and peculiarities. The barrier for developers to directly use them is extremely high.

What SiliconFlow does is translate these diverse domestic chips into standardized, universally usable tokens. The fact that the full version of DeepSeek can run smoothly on Ascend chips is underpinned by this translation capability.

Large companies may be unwilling to optimize for their competitors' chips, and chip manufacturers themselves may lack the software expertise, yet the entire domestic computing power ecosystem depends on it. It’s like a substation in a power grid: the upstream power plants and downstream users can change, but the substation's position is the most stable.

In June this year, the company repurchased all the intellectual property of its predecessor, OneFlow, specifically to strengthen this capability. This is also the confidence behind its decision to list on the HKEX under Chapter 18C (a channel designed for pre-revenue tech companies): The prospectus cites Frost & Sullivan's forecast that the China token supply market grew 16 times from 2024 to 2025 and will continue to expand at an annual rate of 638.3% over the next five years.

Is the token factory a good business? The prospectus's answer is split: selling standardized tokens based on the spread is a tough business right now; becoming the substation for domestic computing power could be a business for the ages. SiliconFlow's gamble is whether it can survive on the traffic from the former long enough to reap the harvest of the latter.

Persevering Despite Repeated Setbacks

To understand why this company is willing to make such a bet, we must look back at Yuan Jinhui.

Throughout his journey, almost every step has been a "close call." He earned his bachelor's degree from Xidian University, graduated top of his computer science class in 2003 to pursue a direct PhD at Tsinghua University under Academician Zhang Bo, and spent nearly a decade at Tsinghua, including his postdoctoral research. His original life plan was to become a professor, but choosing the niche interdisciplinary field of computational neuroscience meant he couldn't secure a faculty position. He was one step away from the lectern but never stood on it.

By the time he left the ivory tower, the golden age of the Chinese internet had passed him by. He worked at Youdao and 360, and later at Microsoft Research Asia, where he developed core systems adopted by companies like Kuaishou, winning the President's Award.

In 2017, he started a company based on a belief that few shared at the time: future models would be too large for existing frameworks, requiring a complete rewrite of the underlying system. This venture was OneFlow. The subsequent wave of large models proved his judgment correct, but the company didn't survive to see the harvest. In 2023, it was acquired by Wang Huiwen's Light Years Beyond at a valuation of $100 million. A few months later, Light Years Beyond was wholly integrated into Meituan.

Having the right judgment and surviving, yet still not winning – that's the truly painful core of the "comment" made in that dumpling restaurant.

At that time, Yuan Jinhui was receiving high-paying offers from major tech companies, and his team members also had respectable options. His choice was to start over for a third time, taking 35 members from his original team of 40. He explained publicly: within a massive company with vast operations, an AI framework might not be a top priority; but for this team, it was their sole focus. The new company was named SiliconFlow. "Silicon" refers to chips, and "Flow" refers to the software that makes computing power fluid – a name echoing OneFlow, as if adding a sequel to an unfinished story. In his WeChat Moments announcing the new venture, he wrote: "The past 15 years haven't been smooth, but we persevere despite repeated setbacks."

China's tech industry has never lacked for storytellers. What is scarce are people who correctly judge the direction, lose the outcome, and are still willing to return to the table. SiliconFlow's prospectus may not look pretty, but behind it stands an engineer who has spent two decades repeatedly proving the same conviction: good technology deserves to be matched with commercial success.

Regardless of the final IPO market valuation, I wish SiliconFlow good luck, and I wish Yuan Jinhui success in finishing the story this time. Through repeated setbacks, may spring finally arrive.