SiliconFlow Bleeds Cash in IPO Sprint, Is the Token Factory a Good Business?
TechFlow Selected TechFlow Selected
SiliconFlow Bleeds Cash in IPO Sprint, Is the Token Factory a Good Business?
Good technology should be worthy of commercial success.
Navigating Web3 tides with focused insightsAuthor: Xiao Bing
One day at the end of 2023, in the Xi Jia De dumpling restaurant downstairs from Tsinghua Science Park, Yuan Jinhui had just sat down when he heard the next table discussing his company: "OneFlow's technology was quite good, but in the end, it didn't make money and got acquired."
He later recalled this scene to LatePost, saying that he sometimes wondered if he had set a bad example: the technical judgment was correct, and he worked diligently, yet still failed to achieve the success recognized by everyone.
Two and a half years later, on June 30, 2026, he stood at the doors of the HKEX with a 347-page prospectus. This time, he wanted to prove the proposition questioned in the dumpling shop: good technology can make money.
However, the prospectus first gave a cruel mid-term result: in 2025, for every 1 yuan this company took in, direct costs alone spent 1.2 yuan.
What kind of company is this?
SiliconFlow does not build large models, nor does it make chips. What it does can be explained in one sentence: rent compute from upstream, process it into Tokens, and sell to downstream.
SiliconFlow's role is similar to an oil refinery.
An oil refinery does not have its own oil fields; it buys crude oil, processes it into gasoline, and sells it, earning the spread in between. SiliconFlow similarly has no oil fields; it rents NVIDIA GPUs from cloud vendors, and also rents domestic chips like Ascend, MetaX, and Moore Threads. It uses its self-developed inference engine to process this diverse compute into standard Tokens, selling them by quantity to developers and enterprises. Rent is its cost price, Tokens are its retail price, and the spread in between is its profit margin.
The problem is, this spread is currently negative.
In 2024, when the company scale was still small, this business was profitable, with a buy-sell spread leaving a gross margin of 39.4%.
By 2025, revenue surged 653.2% to 55.33 million yuan, but the gross margin dropped to negative 24%. Among them, the best-selling public cloud Token business had a gross margin of negative 119%, equivalent to subsidizing 119 yuan out of pocket for every 100 yuan of Tokens sold.
Why is this so?
On one side, inventory is expensive: to handle the massive users that might flood in at any time, the company must rent large batches of compute in advance to stockpile. In 2025, cost of sales soared from 4.452 million yuan the previous year to 68.632 million yuan, exceeding total annual revenue, while the stockpiled compute was not yet fully utilized.
On the other side, retail prices are being slashed aggressively: large firms cut prices repeatedly to grab developers, with prices per thousand Tokens for some mainstream models cut by over 90%; in May this year, DeepSeek announced a permanent 75% price cut for V4-Pro, Tencent Cloud followed, with the hardest cut reaching 97.5%.
What's more troublesome is that neither the cost price nor the retail price is up to SiliconFlow to decide. Compute rent is set by upstream cloud vendors; Token prices are set by the price wars of large firms. When Alibaba and ByteDance cut prices, they use profits from other businesses to subsidize and grab territory; for independent players like SiliconFlow that rely solely on the spread to eat, every price cut announcement equals watching helplessly as their profits are shaved off layer by layer.
Overwhelming Traffic, Overwhelming Bills
The company's most glorious moment happened to illuminate this logic the brightest.
On February 1, 2025, DeepSeek went viral globally, official servers were overwhelmed, and millions of users wanted to use it but couldn't.
SiliconFlow seized this window, partnered with Huawei Cloud, and based on Ascend chips became the first to launch full-performance versions of R1 and V3 services, successfully catching the overflow crowd. Coupled with user acquisition tactics like "Register get 14 yuan, invite get another 14 yuan," website visits 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 the highest day breaking 1.07 trillion. By 2025 throughput, it has already become China's largest independent Token supply platform.
The bill is written on several other pages: 2025 net loss was 345 million yuan, 4.2 times the previous year; even excluding accounting factors like share-based compensation, adjusted loss was still 187 million yuan, cash outflow from operating activities for the year was 172 million yuan, burning about 14.8 million on average per month. Counting from its founding in August 2023, cumulative losses over three years were about 440 million yuan.
For a normal factory, a surge in orders is a tremendous good thing. For a factory with a negative spread, a surge in orders means only one thing: the speed of losing money is also surging.
Why Worth 7.74 Billion?
Seeing this, you might ask: for such a losing business, why seven rounds of financing in three years, with valuation rising from 280 million in the angel round to 7.74 billion? Why are Alibaba, Meituan, Huawei Hubble, SenseTime, Zhipu AI, and Sinovation Ventures all squeezed into the shareholder list?
In the view of the Compute Bureau, it still holds two real cards.
The first card is called Neutrality. One of the things developers fear most is tying their entire business firmly to one large firm's cloud, making future migration costs too high to move. A Token platform that does not belong to any giant naturally reassures people. The fact that giants' money simultaneously appears on its shareholder list precisely shows that all parties need this intermediate zone where no one holds control.
The second card is the real ace narrative: Substation of Domestic Compute.
NVIDIA chip supply constraints are a reality facing the entire Chinese AI industry. Domestic chips like Ascend, MetaX, and Moore Threads have stepped up, but they each have their own architectures and temperaments; the threshold for developers to use them directly is extremely high.
What SiliconFlow does is translate these distinct domestic chips into standard Tokens that everyone knows how to use. The fact that the full-performance version of DeepSeek can run smoothly on Ascend chips is behind this translation ability.
Large firms may not be willing to do this for competitors' chips, and chip factories themselves cannot do it precisely, but the entire domestic compute ecosystem cannot do without it. It is like a substation in the power grid: upstream power plants can change, downstream electricity users can change, but the substation's position is the most stable.
In June this year, the company bought back all intellectual property from its former employer OneFlow, strengthening exactly this layer. This is also the confidence behind its daring to rush for listing via HKEX Chapter 18C (a channel opened specifically for unprofitable tech companies): the prospectus cites Frost & Sullivan's forecast that the China Token supply market scale 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 a Token factory a good business? The answer given by the prospectus is split: selling standard Tokens by spread is currently a tough business; being a substation of domestic compute might be an era-defining business. SiliconFlow's bet lies in whether it can survive on the former's traffic while waiting for the latter's harvest.
Persevere Despite Repeated Failures
To understand why this company dares to make such a bet, one must return to Yuan Jinhui himself.
His path has almost every step stepping on "almost." Bachelor's from Xidian University, in 2003 he went directly to PhD at Tsinghua University as top of the Computer Science department, studied under Academician Zhang Bo, PhD plus Postdoc, stayed in Tsinghua for nearly 10 years. His original life script was to stay and teach, but because he chose the niche interdisciplinary track of computational neuroscience, he didn't wait for a teaching position. Only one kick away from the podium, he finally did not stand on it.
When leaving the ivory tower, China's internet's most lucrative golden age had already brushed past him. He shuttled between Youdao and 360, later at Microsoft Research Asia created core systems adopted by companies like Kuaishou, winning the Dean's Special Award.
In 2017, he came out to start a business with a judgment that almost no one believed at the time: future models would be too large for old frameworks to hold, and the bottom-layer system must be rewritten. This is OneFlow. The later large model wave proved he bet on the right direction, but the company didn't wait for the harvest season; in 2023 it was sold to Wang Huiwen's Light Year Beyond at a valuation of 100 million USD, and a few months later, Light Year Beyond was merged into Meituan as a whole.
Judgment correct, survived, still didn't win; this is where the "comment" in the dumpling shop is truly heartbreaking.
At that time, Yuan Jinhui held high-salary offers from large firms, and team members each had decent places to go. His choice was to take 35 out of the 40-person team and depart for the third time. He explained the reason externally: in business-huge large companies, the priority of AI frameworks is not necessarily high; but for this team, this is the only thing. The new company was named SiliconFlow, "Silicon" refers to chips, "Flow" refers to software that lets compute flow, echoing OneFlow from afar, like adding a stroke to the previous unfinished story. In the Moments announcing the re-startup, he wrote: "Past 15 years not smooth, persevere despite repeated failures."
China's tech industry never lacks people who can tell stories; what is scarce instead is this kind of person who judged the direction right, lost the ending, and is still willing to return to the card table. SiliconFlow's prospectus is not beautiful, but behind it stands an engineer who used 20 years to repeatedly verify the same belief: good technology should deserve commercial success.
Regardless of how the final IPO market pricing turns out, wish SiliconFlow good luck, and also wish Yuan Jinhui finishes the story this time, persevere despite repeated failures, and finally see spring.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
