Written by: Xiao Bing

On a day at the end of 2023, at the Dumpling House of Xi Jiade located under the Tsinghua Technology Park, Yuan Jinhui had just sat down when he overheard the neighboring table discussing his company: “OneFlow technology is really good, yet in the end, it didn’t make any money and got acquired.”

Later, when reminiscing about this scene with LatePost, he said he sometimes wondered if he had set a bad example: the technical judgment was fine, and he worked diligently, yet he still failed to achieve the success recognized by everyone.

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

However, the prospectus first presented a cruel mid-term result: in 2025, for every dollar earned by this company, it had to spend 1 dollar and 20 cents just on direct costs.

What kind of company is this?

Silicon-based Flow does not create large models or manufacture chips; what it does can be explained in one sentence: Rent computing power from upstream, process it into Tokens, and sell it to downstream.

The role of Silicon-based Flow is similar to that of a refinery.

A refinery does not have its own oil fields; it buys crude oil and processes it into gasoline for sale, earning a spread in the middle. Silicon-based Flow also does not have oil fields; it rents Nvidia GPUs from cloud vendors and also rents domestic chips like Ascend, Moxiang, and Moore Threads, using its self-developed inference engine to process these various types of computing power into standardized Tokens, selling them by quantity to developers and enterprises. Rent is its purchase price, Tokens are its selling price, and the spread in between 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%.

By 2025, revenue soared by 653.2% to 55.33 million yuan, but the gross margin fell to -24%. Among them, the most popular public cloud Token business had a gross margin of -119%, which means for every 100 yuan worth of Tokens sold, it was actually losing 119 yuan.

Why did this happen?

On one end, the purchasing cost was high: in order to accommodate the massive influx of users that could come at any time, the company had to rent a large amount of computing power in advance; in 2025, sales costs skyrocketed from 4.452 million yuan the previous year to 68.632 million yuan, exceeding the total annual income, while the stored computing power wasn't utilized efficiently.

On the other end, the retail price was being fiercely undercut: large companies repeatedly slashed prices to compete for developers, with some mainstream model prices being cut by more than 90%; in May of this year, DeepSeek announced a permanent 75% price cut for V4-Pro, and Tencent Cloud followed suit with cuts as severe as 97.5%.

What’s more troubling is that neither the purchasing price nor the retail price can be controlled by Silicon-based Flow. The rent for computing power is set by upstream cloud vendors; the selling price for Tokens is determined by the price wars among large companies. When Alibaba and ByteDance cut prices, they use profits from other businesses to subsidize and grab market share; for independent players like Silicon-based Flow, who rely solely on the spread for survival, every price cut announcement is like watching their profits being sliced away.

Massive traffic, massive bills

At its most glorious moment, this company highlighted the above logic the most.

On February 1, 2025, DeepSeek became a global sensation, with its official servers overwhelmed, and millions of users unable to access its services.

Silicon-based Flow seized this window, partnering with Huawei Cloud, and was the first to launch full-version R1 and V3 services based on Ascend chips, successfully capturing the overflow of users. Coupled with “Register and receive 14 yuan, invite others and receive another 14 yuan” promotional tactics, website traffic surged nearly 40 times. Registered users surged from 127,000 at the end of 2024 to 10.28 million by the end of April this year; the platform processed up to 578.5 billion Tokens daily, with a peak surpassing 1.07 trillion Tokens in a single day. Based on the throughput in 2025, it has become the largest independent Token supply platform in China.

However, the bills tell another story: 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, with operating activities draining 172 million yuan in cash over the year, burning an average of about 14.8 million yuan each month. Since its establishment in August 2023, it has accumulated a loss of about 440 million yuan over three years.

For a typical factory, a surge in orders is a great blessing. For a factory with a negative spread, a surge in orders simply means one thing: The speed of losing money is also soaring.

What justifies a valuation of 7.74 billion?

Having reached this point, you might wonder: how can a loss-making business secure seven rounds of financing over three years, with its valuation rising from 280 million yuan in the angel round to 7.74 billion yuan? How come Alibaba, Meituan, Huawei Hubble, SenseTime, Zhizhi, and Innovation Works are all listed among its shareholders?

In the eyes of the computing power bureau, it still holds two true cards.

The first card is neutrality. One of the biggest fears for developers is being locked into a single cloud provider, making it expensive to move later. A Token platform that does not belong to any giant provides natural reassurance. The presence of the giants’ investments in its shareholder list only shows that all parties need such a neutral zone with no single entity controlling it.

The second card is the real ace narrative: A transformer station for domestic computing power.

The limited supply of Nvidia chips poses a reality for the entire Chinese AI industry. Domestic chips like Ascend, Moxiang, and Moore Threads have emerged, but each has its architecture and quirks, resulting in a high barrier to entry for developers.

What Silicon-based Flow does is to unify these distinctly different domestic chips into standardized Tokens that everyone can use. The full-version DeepSeek can run smoothly on Ascend chips precisely because of this translation capability.

This task may not be something large companies are willing to do for competitors’ chips, and chip manufacturers can’t perfect it either, yet it is essential for the entire domestic computing power ecosystem. It acts like a transformer station within the power grid: the upstream power plants can change, the downstream consumers can change, yet the location of the transformer station remains the most stable.

This June, the company bought back all the intellectual property from its former parent OneFlow, which strengthens this layer. This is also the confidence behind its rush to list on the Hong Kong stock market with Chapter 18C (a channel specifically for unprofitable technology companies): The prospectus cites a forecast from Frost & Sullivan indicating that the Chinese Token supply market will grow 16 times from 2024 to 2025 and will 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 standardized Tokens at a spread currently is a tough business; acting as a transformer station for domestic computing power might be a once-in-a-generation business. Silicon-based Flow's bet lies in whether it can survive on the flow from the former until the latter bears fruit.

Resilience in adversity

To understand why this company dares to make such a bet, we must return to the person of Yuan Jinhui.

Almost every step of his journey has tread on the line of “almost there.” An undergraduate from Xi'an University of Electronic Science and Technology, he entered Tsinghua University as the top student in the Computer Science Department in 2003, studying under Academician Zhang Bo, completing his PhD and postdoctoral work, spending nearly ten years in Tsinghua. His original life script was to stay and teach, yet due to selecting the niche field of computational neuroscience, he never got to wait for a teaching position. He was just a step away from the podium but ultimately never stood on it.

Upon leaving the ivory tower, China’s internet boom had already slipped past him. He had ventures at Youdao and 360, later working at Microsoft Research Asia, where he developed a core system adopted by companies like Kuaishou, winning a special award from the director.

In 2017, he ventured out to start a business with a judgement that almost no one trusted at the time: that future models would be too large to fit within old frameworks, and the underlying system would need a complete rewrite. This was OneFlow. The later wave of large models proved he had made the right bet, but the company failed to reap the rewards and was sold for 100 million USD to Wang Huiwen’s company A Year Beyond, which was then fully merged into Meituan a few months later.

He judged right, survived, yet still didn’t win; that’s what makes the comment in the dumpling house truly poignant.

At that time, Yuan Jinhui held high-paying offers from large companies, and the people in his team also had decent prospects. His choice was to venture out a third time with 35 people from his team of 40. He has explained his reasons publicly: within large companies with vast operations, AI frameworks may not have a high priority; but for this team, this was everything. The new company was named Silicon-based Flow, with “Silicon-based” referring to chips, and “Flow” indicating the software that enables computing power to flow, resonating with OneFlow, as if adding a line to a story that was left unfinished, declaring in his circle of friends about the re-entrepreneurship: “The past 15 years have not been smooth; I have faced repeated defeats and fights.”

The Chinese tech industry has never lacked those who can tell stories; what is scarce are those who have made accurate judgments, lost in the end, and are still willing to return to the table. The prospectus of Silicon-based Flow is far from pretty, but behind it stands an engineer who has validated the same belief over twenty years: good technology should deserve commercial success.

Regardless of how the final IPO market prices it, we wish Silicon-based Flow good luck, and we hope Yuan Jinhui can finally complete this story, fighting through repeated defeats until spring comes.