Produced by | Entrepreneur Frontline

Art Editor | Xing Jing

Reviewer | Song Wen

As AI large models begin to charge fees, the "factories" that sell tokens have gradually become one of the most lucrative AI tracks in the market.

Overseas, Fireworks AI—which Jensen Huang dubbed "the TSMC of AI factories"—secured $1.505 billion in financing in July, with its valuation climbing to $17.5 billion. Domestically, on October 12, Shanghai-based AI infrastructure company Infinigence AI confirmed the completion of a 2.3 billion yuan Series C round, bringing its cumulative financing to 4.3 billion yuan.

Amid this capital enthusiasm, companies building token "factories" have begun to step into the public eye, with firms such as SiliconFlow and PaiXiang Future successively filing listing applications with the Hong Kong Stock Exchange.

Among them, PaiXiang Future's founder is Yao Xin, a post-80s academic high-flyer who previously founded PPTV. After exiting PPTV, he quietly positioned himself in the computing power track well before the AI computing power explosion.

PaiXiang Future is an independent distributed cloud computing service provider. Its business began in the edge cloud computing services field, and its revenue had already reached 770 million yuan by 2025. According to China Insights Consultancy, the company is China's largest independent AI cloud computing service provider.

However, the company has yet to achieve profitability, and its operating cash flow remains in a net outflow state. Moreover, under the pressure of convertible conditions, the company's need for listing is very urgent.

But what is really behind this token business that has successively achieved "big results" in the capital market?

1. Building a Token Factory, Aiming to Be the "Didi of Computing Power"

Simply put, a token factory is a "workshop" specialized in scaled production of inference services, continuously converting electricity, chips, models, and data into measurable, schedulable, and priceable model invocation capabilities.

Traditional cloud computing giants mainly follow the "centralized big-tech" logic, building several super data centers across the country and transmitting computing power to users through long-distance optical fibers. This model is prone to physical bottlenecks in the 5G and intelligent era, which demands ultra-low latency and massive concurrency.

Unlike public cloud platforms such as Alibaba Cloud and Tencent Cloud, PaiXiang Future does not build its own data centers. Instead, it has built China's largest distributed computing power network through a "swarm intelligence" approach.

The company found that there is a large amount of dispersed, idle, and heterogeneous third-party idle computing power resources in society, and there is a major mismatch between massive computing power demand and dispersed, not yet fully utilized computing power supply.

Therefore, the company built a computing power scheduling platform, treating these resources as "private cars," aggregating and managing the idle computing power of heterogeneous resources, and deploying it to customers in an optimal way, thereby accelerating content transmission and improving low-latency computing experience.

In an interview, founder Yao Xin compared the company to the "Didi of computing power," saying that the company does not purchase or hold large-scale GPU servers and data centers, does not independently build its own intelligent computing parks, and procures and integrates all computing power resources externally, then matches supply and demand through its self-developed distributed scheduling system.

(Image / Prospectus)

It can be said that PaiXiang Future's core asset is not hardware, but an extremely smart distributed cloud computing architecture and unified scheduling platform. It is precisely through this scheduling capability that the company has revitalized more than 4,600 idle nodes across more than 1,340 counties and cities nationwide, successfully building an edge computing power network and pushing computing power directly down to the very edge of the network.

(Image / Prospectus)

This "asset-light, scheduling-heavy" model ran straight into the super outlet of the large model explosion in 2023. The prospectus shows that after distributed scheduling by the company, GPU utilization remained above 75% throughout 2025, far higher than the industry average of 40%-50%, and model inference costs could theoretically fall by more than 80%. When the AI era arrived, this advantage was amplified.

(Image / Prospectus)

Building PaiXiang Future was not Yao Xin's first entrepreneurial venture.

In 2004, Yao Xin, then a first-year graduate student at Huazhong University of Science and Technology, keenly captured the characteristics of P2P technology that "the more people, the faster it gets" in order to solve the pain point of classmates experiencing buffering while watching live football matches in dormitories. He simply took a leave of absence from school and founded PPLive, the predecessor of PPTV.

In that golden age of the Internet when broadband and distribution infrastructure were extremely scarce, PPTV once had 450 million global users and became the country's second-largest video platform after Youku. In 2014, PPTV was ultimately acquired by Suning Group.

In 2018, Yao Xin joined hands again with his old partner and HUST junior Wang Wenyu to establish Paiou Cloud Computing (Shanghai) Co., Ltd. (hereinafter referred to as Paiou Cloud, one of PaiXiang Future's consolidated affiliated entities in China). Now they are together leading PaiXiang Future in its sprint toward the capital market.

The prospectus shows that before the IPO, Yao Xin and his spouse Lyu Shanshan collectively controlled approximately 48.05% of the voting rights through several entities; Wang Wenyu indirectly held 10.17%; other investors include Millennium, IMO Ventures, CCV Fund, Huaye Tiancheng, Panlin, Feidian, and others.

2. Revenue Doubled in Two Years, but Why Does It Lose More the More It Sells?

At present, PaiXiang Future's business model adopts the charging model of cloud computing, that is, billing by computing power usage or usage duration. Its main paying customers include large Internet companies and small and medium-sized AI startup companies.

From 2023 to 2025 (hereinafter referred to as the reporting period), the company's revenue rapidly grew from 358 million yuan to 770 million yuan, with a compound annual growth rate of 46.6%.

Among this, edge cloud computing service revenue (including edge node services and edge CDN) was 358 million yuan, 548 million yuan, and 651 million yuan, respectively, consistently accounting for more than 84% of total revenue.

Meanwhile, AI cloud computing service revenue surged from 265,000 yuan in 2023 to 119 million yuan, growing more than 10 times year-on-year in 2025 and rising to 15.5% of the total, becoming the second growth curve heavily emphasized in the prospectus.

(Image / Prospectus)

Revenue grew rapidly, but gross margin declined significantly. During the reporting period, its overall gross margin slid from 17.7% in 2023 to 12.3% in 2024, and then to 9.4% in 2025, nearly halving.

(Image / Prospectus)

Behind this, the company has remained mired in losses. During the reporting period, its net losses were 189 million yuan, 294 million yuan, and 223 million yuan, respectively. If non-cash items such as fair value changes of convertible redeemable preferred shares are excluded, the adjusted net loss still expanded from 37 million yuan to 105 million yuan, falling into a situation of "the more it sells, the more it loses."

(Image / Prospectus)

The root cause of the losses mainly lies in the company's rapidly growing AI cloud computing business.

In 2025, its AI cloud computing gross margin was -10.7%, compared with -91.3% and -95.1% in 2023 and 2024, respectively. Although the gross margin loss has narrowed, it is still a "loss-making deal" to date, and as its revenue share grows ever higher, it has dragged down overall gross margin performance.

Entrepreneur Frontline learned that the company's edge cloud business mainly places computing power closest to users, specifically accelerating video, live streaming, and file downloads (that is, edge CDN).

The AI cloud business, meanwhile, provides GPU computing power and ready-made large model interfaces (MaaS), so customers can run AI inference without buying graphics cards or setting up environments themselves. However, this revenue mainly comes from model APIs, charging by token for customers' calls to large models.

But token prices are not set by PaiXiang Future; they are the official model pricing. According to a report by Aifenxi, Paiou Cloud's pricing page is consistent with the official API document pricing of DeepSeek, GLM, Kimi, and others.

(Image / Aifenxi (top), Prospectus (bottom))

The reason is that pricing power rests with model manufacturers. Token APIs are homogeneous commodities, and the cost and threshold for users to switch token suppliers are extremely low. Therefore, even if large model manufacturers have insufficient inference computing power supply, it is difficult for PaiXiang Future to charge a premium.

On the cost side, most of PaiXiang Future's GPU computing power is leased. In 2025, GPU resource costs were 117 million yuan, almost catching up with overall AI cloud revenue (119 million yuan). It can be said that for every additional token the company provides, it must pay an additional GPU rental fee.

(Image / Prospectus)

From this point of view, although the company likens itself to the "Didi of computing power," Didi has pricing power over its commission, whereas PaiXiang Future has model manufacturers with pricing power upstream and customers who can switch with one click downstream, squeezing it from both ends.

However, demand is indeed very strong. The company disclosed that average daily token consumption has long exceeded one trillion, indicating that broad market demand truly exists. But strong demand does not equal profitability, and the company's profit space has not yet opened up.

In the prospectus, the company also warned: "As usage of the company's products and services increases, customers may request lower pricing, and some customers may choose to use other cloud service providers for business, compliance, or other reasons. To maintain or increase revenue, the company may need to invest more in sales and marketing, all of which will adversely affect the company's gross margin."

It can be seen that the company's low gross margin is not a temporary pain, but a natural limitation of this "rent cards—resell tokens" model. In addition, the computing power intermediary track has low barriers and severe homogenization, and it will not be easy for PaiXiang Future to break through.

3. Listing Countdown: Only Two Months Left

It is worth noting that the company's operating lifeline rests on a very small number of major customers.

From 2023 to 2025, revenue from the company's top five customers accounted for 92.5%, 89.5%, and 79.0% of total revenue, respectively, indicating heavy reliance on major customers.

(Image / Prospectus)

Facing major customers, PaiXiang Future has a relatively long collection cycle, causing the company's operating cash flow to remain negative.

The prospectus shows that the company usually grants edge cloud computing service customers a credit period of one to three months. From 2023 to 2025, trade receivables were 90 million yuan, 150 million yuan, and 168 million yuan, respectively.

During the same period, trade receivables turnover days were 83 days, 79 days, and 75 days, respectively. During the reporting period, 93.6%, 88.1%, and 74.9% of trade receivables were amounts due from the top five customers, respectively.

There is usually a time lag between collecting usage-based fees from customers and paying suppliers, resulting in potential operating cash flow mismatches. During the reporting period, its net operating cash flow was -35 million yuan, -41 million yuan, and -136 million yuan, respectively.

(Image / Prospectus)

The company's operations rely heavily on financing. From 2023 to 2025, its cumulative equity financing amounted to $60 million, $60 million, and $106 million, respectively, while interest-bearing bank borrowings were 44 million yuan, 74 million yuan, and 81 million yuan, respectively, rising year by year.

(Image / Prospectus)

In this Hong Kong IPO, the company plans to use the proceeds to enhance its technical capabilities in distributed cloud computing services and upgrade AI cloud computing services; expand overseas, especially in mature markets and emerging markets with untapped potential; make strategic investments and acquisitions; and for working capital and general corporate purposes.

In addition to working capital pressure, another reason PaiXiang Future is eager to list is the "redemption pressure" hanging over its head.

The prospectus reveals that PaiXiang Future is backed by investment institutions including BlueRun Ventures, Creation Partners, and Huaye Tiancheng, as well as early major investors such as Wang Xiaochuan of Baichuan Intelligence, former Microsoft Azure China president Shen Yuanqing, and Tang Binsen of Genki Forest.

From a valuation of $46 million after the angel round in 2021, to $469 million following the Series B financing in 2025, and then to the Series B-2 round just completed in late May and early June 2026, PaiXiang Future's post-investment valuation has climbed to $576 million.

While the capital raised fueled business expansion, it could also convert into debt due for repayment. The very investment consortium that once endorsed the company now constitutes precisely the group holding redemption rights.

From 2023 to 2025, the company recorded fair value losses on convertible redeemable preferred shares of 145 million yuan, 222 million yuan, and 88 million yuan, respectively.

(Image / Prospectus)

Under the agreement signed when Pre-IPO investors subscribed, if a "qualified listing" is not completed by the agreed deadline—originally seven years after investment, later shortened to five years during the Series B round in April 2025—or if the VIE structure becomes invalid or is terminated, investors are entitled to sell their shares back to the company at 150% of the original investment amount plus accrued dividends.

In June 2025, the company signed a supplementary agreement with investors. Once the listing application is submitted, the redemption right is "suspended from exercise"; however, should the Hong Kong Stock Exchange reject the application, should the company voluntarily abandon the listing, or should the listing fail to complete before December 31, 2026, the redemption right will be reinstated.

PaiXiang Future first submitted its listing application in June 2025, after which the prospectus lapsed, and it did not resubmit until June 10, 2026. Now only a little over two months remain before this listing assessment deadline.

(Image / Prospectus)

At the heart of the boom, the token factory business is indeed thriving—but once it enters the capital market, the new narrative it spins must ultimately return to the ledger of "real money."

*Note: The cover image in this article is from 699pic.com, used under the VRF agreement.