Silicon-Flow Blood-Soaked Sprint Toward an IPO: Is the Token Factory a Good Business?

Byline: 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 the next table discussing his company: “OneFlow’s technology is pretty good, but in the end they didn’t make any money either—they were acquired.”

When he later recalled this moment to LatePost, he said that sometimes he wondered whether he had set a bad example: his technical judgment was right, and he worked diligently, yet he still couldn’t achieve the kind of success that earns everyone’s 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 in hand. This time, he wanted to prove the proposition that had been questioned in the dumpling restaurant: good technology can make money.

However, the prospectus first delivered a brutal midterm result: in 2025, for every 1 yuan the company brought in, it had to spend 1.2 yuan on direct costs alone.

What kind of company is this?

SiliconFlow doesn’t build large models and it doesn’t make chips. What it does can be summed up in one sentence: lease computing power from upstream, process it into Tokens, and then sell it to downstream.

SiliconFlow’s role is similar to an oil refinery.

A refinery doesn’t have its own oil fields. It buys crude oil, processes it into gasoline, and sells it, earning money from the spread in between. Likewise, SiliconFlow doesn’t have its own oil fields. It leases NVIDIA GPUs from cloud providers, and also leases domestic chips such as Ascend, Mucsi, and Moore Threads. Using its own developed inference engine, it processes these disparate computing resources into standardized Tokens, selling them by quantity to developers and enterprises. The lease fee is its purchase cost, the Token is its retail price, and the spread between them is its profit margin.

The problem is that this spread is now negative.

In 2024, when the company was still small, this business was profitable. The spread left a gross margin of 39.4%.

By 2025, revenue surged 653.2% to 55.33 million yuan, but the gross margin fell to -24%. Among them, the hottest public cloud Token business had a gross margin of -119%—meaning that for every 100 yuan in Token sales, it effectively subsidized 119 yuan.

Why did this happen?

On one side, the purchase cost is expensive. To handle the massive surge of users that could arrive at any time, the company had to lease large amounts of computing power in advance. In 2025, sales costs jumped from 4.452 million yuan from the previous year to 68.632 million yuan—exceeding total annual revenue—while the computing power it stockpiled had not been fully utilized.

On the other side, the retail price is being slashed: big companies keep cutting prices to win developers. For some mainstream models, the price per thousand Tokens was cut by more than 90%. In May this year, DeepSeek announced a permanent 75% price cut for V4-Pro; Tencent Cloud followed with cuts too, with the most aggressive cut reaching 97.5%.

What’s even more troublesome is that neither the purchase price nor the retail price is something SiliconFlow gets to decide. The rental rate of computing power is set by upstream cloud vendors; the Token pricing is set by the price war among big companies. When Alibaba and ByteDance cut prices, they use money earned from other businesses to subsidize and grab market share. For an independent player like SiliconFlow that survives solely on the spread, every price-cut announcement is like watching its profit margin get shaved off a layer right in front of its eyes.

A deluge of traffic, a deluge of bills

The company’s most glorious moment actually illuminated the logic above the brightest.

On February 1, 2025, DeepSeek went viral worldwide. Its official servers were overwhelmed—millions of users wanted to use it but couldn’t.

SiliconFlow seized this window. In cooperation with Huawei Cloud, based on Ascend chips, it was the first to launch full-power versions of R1 and V3 services, quickly absorbing the overflow crowd. Paired with a user-acquisition gimmick of “get 14 yuan when registering, and get another 14 yuan when inviting,” website traffic soared by nearly 40 times. Registered users increased 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 per day, with a peak of 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 bill is written on other pages: in 2025, net loss was 345 million yuan, 4.2 times the previous year. Even after excluding book-based factors such as equity incentives, the adjusted loss was still 187 million yuan. Cash outflow from operating activities was 172 million yuan for the year—burning about 14.8 million yuan per month on average. Since it was founded in August 2023, the total loss over three years has been about 440 million yuan.

For a normal factory, an explosion in orders is great news. For a factory with a negative spread, an explosion in orders means only one thing: the speed of losing money is also exploding.

Why is it worth 7.74 billion?

At this point, you might ask: how can a money-losing business like this raise funding in seven rounds over three years, with its valuation rising from 280 million yuan at the angel round to 7.74 billion yuan? Why are Alibaba, Meituan, Huawei Hubble, SenseTime, Zhipu, and Innovation Works all squeezed into its shareholder list?

From the perspective of the computing power ecosystem, it still holds two truly real cards.

The first card is “neutrality.” One of the biggest fears developers have is tying the entire business to the cloud of one particular big vendor. In the future, if they want to move, the cost would be so high that they can’t move. A Token platform that doesn’t belong to any giant naturally inspires trust. At the same time, the fact that the giants’ money appears on its shareholder roster precisely suggests that every party needs such a middle ground that no one controls.

The second card is the real ace of the story: a substation for domestic computing power.

The limited supply of NVIDIA chips is a reality facing China’s entire AI industry. Domestic chips such as Ascend, Mucsi, and Moore Threads are stepping in—but each has its own architecture and temperament. As a result, it is extremely difficult for developers to use them directly.

What SiliconFlow does is unify these different domestic chips into standard Tokens that everyone can use. The fact that DeepSeek’s full-power version runs smoothly on Ascend chips is backed by this “translation” capability.

Big vendors may not be willing to optimize for competitors’ chips, and chip makers themselves may not be able to do it well—but the entire domestic computing power ecosystem cannot do without it. It is like a substation in a power grid: power plants upstream can change, power consumers downstream can change, but the substation’s position is the most stable.

This June, the company bought back all the intellectual property from its former employer, OneFlow—strengthening precisely this layer. This is also what gives it the confidence to sprint toward an IPO under Hong Kong Stock Exchange Chapter 18C (a channel for unprofitable tech companies). The prospectus cites Frost & Sullivan’s predictions: the size of China’s Token supply market increased 16 times from 2024 to 2025, and over the next five years it will keep expanding at a rate of 638.3% per year.

Is a Token factory a good business? The prospectus offers a split answer: selling standard Tokens based on the spread is currently a tough business; being a substation for domestic computing power could be a business on the scale of the era. SiliconFlow’s bet is whether it can survive on the traffic from the former until the latter starts to pay off.

Failing repeatedly, yet fighting on

To understand why this company dares to make such a bet, you have to go back to the person at its center: Yuan Jinhui.

Almost every step along his path has been marked by “just missing the mark.” He was an undergraduate at Xi’an University of Electronic Science and Technology, earning the top rank in the computer science department in 2003 to pursue a direct Ph.D. at Tsinghua University, where he studied under Academician Zhang Bo. He then spent nearly ten years on Tsinghua’s campus, pursuing his Ph.D. and postdoctoral work. His original plan in life was to stay and teach, but because he chose the niche and interdisciplinary track of computational neuroscience, he never waited for a faculty position. He was only one step away from the podium, but in the end he never stood on it.

When he left the ivory tower, China’s internet’s most prosperous golden age 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 such as Kuaishou, earning a special award.

In 2017, he started a business with a judgment that, at the time, almost no one believed: future models would be too large for old frameworks to hold, and the underlying systems would have to be rewritten. That was OneFlow. The later wave of large models proved he had the right direction, but the company couldn’t wait for the harvest season. In 2023, it was sold at a valuation of $100 million to Wang Huiwen’s Light Year Outside. A few months later, Light Year Outside was fully merged into Meituan.

He judged correctly and survived—but still didn’t win. That is exactly where the “comment” in the dumpling restaurant stings.

At that time, Yuan Jinhui held high-paying offers from big tech companies, and the people in his team each had respectable destinations too. His choice was to take 35 of the 40-person team and set out for the third time. He explained the reason externally: in a large company with massive operations, AI frameworks may not necessarily be a top priority; but for this team, it was the only priority. The new company was named SiliconFlow—“Silicon” refers to chips, and “Flow” refers to software that makes computing power flow—echoing OneFlow. It was as if he added another chapter to a story that hadn’t finished yet. In his WeChat Moments announcing the new venture, he wrote: “The past 15 years haven’t been smooth—failing repeatedly, yet fighting on.”

China’s tech industry has never lacked people who can tell stories. What is truly scarce, however, are people who judge the direction correctly, lose the outcome, and still are willing to return to the table. SiliconFlow’s prospectus may not look pretty, but behind it stands an engineer who has spent 20 years repeatedly validating the same belief: good technology should deserve commercial success.

No matter what the final IPO market price is, I wish SiliconFlow good luck, and I also wish Yuan Jinhui that this time he finishes the story: failing repeatedly yet fighting on, until spring finally comes.