文 | 创业最前线，作者 | 于莹，编辑 | 王亚静
当AI大模型开始收费，卖Token的“工厂”逐渐成为市场最“吸金”的AI赛道之一。
在海外，被黄仁勋称为“AI工厂的台积电”的Fireworks AI7月拿到15.05亿美元融资，估值升至175亿美元；在国内，10月12日，上海AI Infra（人工智能基础设施）企业无问芯穹确认已完成23亿元C轮融资，累计融资额达43亿元。
资本追捧之下，做Token“工厂”的企业，开始走向大众视野，硅基流动、派想未来等企业纷纷向港交所递交上市申请。
其中，派想未来的创始人是曾创立了PPTV的“80后”学霸姚欣，他在退出ppvt后，在AI算力爆发前悄悄布局了算力赛道。
派想未来是一家独立分布式云计算服务商，业务始于边缘云计算服务领域，2025年营收已经达到7.7亿元。按灼识咨询口径，公司是中国最大的独立AI云计算服务商。
不过，公司至今尚未盈利，经营现金流也呈现净流出状态。而在可转股的条件压力下，公司上市需求十分迫切。
只是，这门在资本市场陆续拿到“大结果”的Token生意，底色究竟如何呢？
1、打造token工厂，要做“算力版滴滴”
通俗来说，Token工厂就是专门规模化生产推理服务的“车间”，把电力、芯片、模型和数据，持续转化为可计量、可调度、可定价的模型调用能力。
而一般的传统云计算巨头，主要是“中心化大厂”逻辑，在全国建几个超级数据中心，通过长途光纤把算力输送给用户。这种模式，在需要超低时延、海量并发的5G和智能化时代，容易遇到物理瓶颈。
与阿里云、腾讯云等公有云平台不同，派想未来并非自建数据中心，而是通过“蜂群思维”搭建起中国最大的分布式算力网络。
公司发现，社会上存在大量分散、闲置、异构的第三方闲置算力资源，海量的算力需求与分散且尚未充分利用的算力供给之间存在重大错配。
所以，公司搭建了算力调度平台，把它们当成“私家车”，汇聚和管理异构资源的闲置算力，以最优方式向客户部署，从而加速内容传输及提升低延迟计算体验。
在接受采访时，创始人姚欣将公司比作“算力版滴滴”，称公司不采购、不持有大规模GPU服务器与数据中心，不自主搭建自有智算园区，全部算力资源向外采购整合，再通过自研分布式调度系统匹配供需。
可以说，派想未来的核心资产不是硬件，而是一套极度聪明的分布式云计算架构和统一调度平台。正是通过这种调度能力，公司把全国1340多个县市的4600多个闲置节点盘活，成功构建起了一个边缘算力网络，把算力直接下沉到网络的最边缘。
这种“轻资产、重调度”的模式，在2023年撞上了大模型爆发的超级风口。招股书显示，公司通过分布式调度后，GPU利用率在2025年一直超过75%，远高于行业40%-50%的平均水平，模型推理成本理论上可以下降超过80%。当AI时代来临的时候，这种优势被放大。
而打造派想未来，并非是姚欣的首次创业。
2004年，正在华中科技大学读研一的姚欣，为了解决同学们在宿舍看球赛直播卡顿的痛点，敏锐地捕捉到了P2P技术“人越多越快”的特性，索性休学创办了PPTV的前身PPLive。
在那个宽带和分发基础设施极度匮乏的互联网黄金年代，PPTV一度坐拥4.5亿全球用户，成了仅次于优酷的全国第二大视频平台，2014年PPTV最终被苏宁集团收购。
2018年，姚欣与当年的老搭档、华科学弟王闻宇再度联手，成立了派欧云计算（上海）有限公司（以下简称派欧云，派想未来在中国的合并联属实体之一），如今一起带领着派想未来冲刺资本市场。
招股书显示，IPO前，姚欣及其配偶吕姗姗，通过数家实体合计控制约48.05%的股份投票权；王闻宇间接持股10.17%，其他投资者包括Millennium、IMO Ventures、CCV Fund、华业天成、磐霖、沸点等。
2、营收2年翻倍，但为何越卖越亏？
目前，派想未来的商业模式采用云计算的收费模式，即按算力使用量或使用时长计费，主要付费客户包括大型互联网企业，以及中小型AI创业企业。
2023年至2025年（以下简称报告期内），公司收入从3.58亿元迅速增长至7.70亿元，复合年增长率达46.6%。
其中，边缘云计算服务收入（包括边缘节点服务、边缘CDN）分别为3.58亿元、5.48亿元、6.51亿元，长期占据总收入84%以上。
而AI云计算服务收入从2023年的26.5万元暴涨至1.19亿元，2025年同比增长超过10倍，占比提升至15.5%，成为招股书重点渲染的第二增长曲线。
营收高增，毛利率却明显下滑。报告期内，其综合毛利率从2023年的17.7%滑落到2024年的12.3%，再到2025年的9.4%，近乎腰斩。
在这背后，公司一直深陷亏损之中。报告期内，其净亏损分别为1.89亿元、2.94亿元和2.23亿元。如果剔掉可转换可赎回优先股公允价值变动等非现金项，经调整净亏损仍从0.37亿元扩大到1.05亿元，陷入了“卖得越多，亏得越多”的境地。
造成亏损的根源，主要出在公司快速增长的AI云计算业务上。
2025年，其AI云计算毛利率为-10.7%，而2023年、2024年则分别为-91.3%和-95.1%，虽然毛利率亏损收窄，但至今仍是一笔“亏本买卖”，而其收入占比越来越高，拖累了整体毛利率表现。
「创业最前线」了解到，公司边缘云业务，主要是把算力放到离用户最近的地方，专门给视频、直播、文件下载做加速（也就是边缘CDN）。
而AI云业务，则提供的是GPU算力和现成大模型接口（MaaS），客户不用自己买显卡、搭环境就能跑AI推理。但这块收入，主要来源于模型API，其为客户调大模型、按Token收费。
但Token价格并非为派想未来所定，而是模型官方定价。根据爱分析的报道，派欧云的定价页面和DeepSeek、GLM、Kimi等官方API文档定价一致。
原因在于，定价权在模型厂商手里。Token API是同质化商品，用户切换Token供给方的成本、门槛极低。因此，即使大模型厂商的推理算力供给不足，派想未来也难以溢价收费。
而在成本端，派想未来的GPU算力多是租赁而来，2025年GPU资源成本1.17亿元，几乎追平了整体的AI云收入（1.19亿元）。可以说，公司每多提供一个Token，就要多付一笔GPU租金。
从这点看，虽然公司把自身比喻成“算力版滴滴”，但滴滴的抽佣定价权在自己手里，而派想未来上游是掌握定价权的模型厂商，下游是可以一键切换的客户，两头都被挤压。
不过，需求端确实是十分紧俏，公司披露的日均Token消耗早已破万亿，说明市场确实存在广泛需求，只是需求旺盛并不等于赚钱能力，公司的利润空间尚未打开。
在招股书中，公司也提示：“随着对公司产品及服务的使用量增加，客户可能要求降低定价，部分客户可能出于商业、合规或其他原因选择使用其他云服务提供商。为此，公司为维持或增加收入，或需要在销售及营销方面投入更多的支出，这些都会对公司的毛利率产生不利影响。”
可见，公司的低毛利率并非一时阵痛，是这套“租卡—转售Token”模式的天然限制。另外，算力中介赛道门槛低、同质化严重，派想未来若想突围并不容易。
3、上市倒计时，仅剩两个月时间
值得注意的是，公司的经营命脉，压在极少数大客户身上。
2023至2025年，公司的前五大客户收入占总营收比重分别达92.5%、89.5%、79.0%，营收高度依赖大客户。
面对大客户，派想未来的回款周期较长，导致公司经营现金流持续为负。
招股书显示，公司通常向边缘云计算服务客户授予一至三个月的信贷期，2023年至2025年，贸易应收款项分别为0.9亿元、1.5亿元、1.68亿元。
同期，贸易应收账款周转天数分别为83天、79天及75天。报告期内，贸易应收款项中分别有93.6%、88.1%及74.9%为应收五大客户款项。
而从客户收取按使用量计算的费用，与向供应商付款之间通常存在时间差，从而导致潜在的经营现金流错配。报告期内，其经营现金流净额分别为-0.35亿元、-0.41亿元、-1.36亿元。
而公司运营颇为依赖融资。2023年至2025年，其累计股权融资额分别为0.6亿美元、0.6亿美元及1.06亿美元，计息银行借款分别为0.44亿元、0.74亿元及0.81亿元，逐年攀升。
此次港股IPO，公司拟将募资用于，提升公司于分布式云计算服务的技术能力及升级AI云计算服务；海外扩张，特别是在成熟市场及潜力尚未开发的新兴市场；战略投资及收购；营运资金及一般企业用途。
除了营运资金的压力，目前派想未来着急上市的原因，还有来自于悬在头顶的“赎回压力”。
招股书显示，派想未来的背后站着蓝驰创投、创世伙伴、华业天成等投资机构，以及早期重仓的百川智能王小川、前微软Azure中国区总裁申元庆、元气森林唐彬森。
从2021年天使轮后的4600万美元估值，到2025年B轮融资后的4.69亿美元，再到2026年5月底6月初刚完成的B-2轮，派想未来的投后估值已经达到5.76亿美元。
融来的钱虽然撑起了业务扩张，但也可能变成到期要还的债。当初为它背书的投资天团，此刻也正是手握赎回权的人。
2023年至2025年，公司分别录得可转换可赎回优先股公允价值亏损1.45亿元、2.22亿元及0.88亿元。
根据Pre-IPO投资人入股时约定，如果到约定期限没完成“合资格上市”（原定投后七年，2025年4月B轮时缩成五年），或者VIE结构失效、被终止，投资人就有权按原投资额的150%外加应计股息，把股份卖回给公司。
2025年6月，公司与投资人签订了补充协议，递交上市申请后，赎回权“暂停行使”，但如果港交所拒绝申请、公司主动放弃上市，或者2026年12月31日前没能完成上市，赎回权将恢复。
派想未来2025年6月首次递表，后招股书失效，2026年6月10日才重新递表，如今距离这条上市考核线只剩两个多月的时间。
风口之上，token工厂的生意的确火热，但一旦迈入资本市场，其所讲的新故事终要回到“真金白银”的账本上来。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen Zheng's first line of business, by Yu Yu, editor Wang Zing
When AILLM started charging fees, Token's “workshop” became one of the most “gold” tracks in the market.
Overseas, Fireworks AI, known by Hoang In-hoon as “Technology at the AI Plant”, received $1.505 billion in financing in July, rising to $17.5 billion; at home, on 12 October, the Shanghai AI Infra (Advanced Intellectual Infrastructure) firm confirmed the completion of the 2.3 billion C round of financing, with a cumulative amount of $4.3 billion.
With capital, the Token “factor” enterprise has begun to move towards a public view, with silica-based flows and future-thinking enterprises submitting their listing applications to Hong Kong.
Among them, Yao Xin, the founder of PPTV's “80s after” school, quietly set up the calculus track before the ACK broke out after withdrawing from the pvt.
In the future, he wants to be an independent, distributed cloud computing service provider, with operations starting in the marginal cloud computing service area, with a profit of $770 million in 2025. By calibration, the company is China ' s largest independent AI Yun computing service.
However, the company has not yet made a profit and operating cash flows have shown a net outflow. However, under the pressure of transferable terms, there is an urgent need for companies to be listed.
It's just, what's the bottom color of Token's business, which continues to get the big results in capital markets?
1. Construction of a token factory with "calculative drip"
In general, the Token plant is a “workshop” dedicated to scalable production reasoning services that continuously transforms electricity, chips, models and data into measurable, dispatchable, priceable model call capabilities.
Instead, conventional cloud computing giants, mainly the “centreization plant” logic, have built several super-data centres throughout the country to channel computing power to users through long-distance fibre-optics. This model is prone to physical bottlenecks in an age of requiring ultra-low latency, five Gs of co-generation and intelligence.
Unlike public cloud platforms such as Alibaba Cloud, Tencent Clouds, the future is not a self-built data centre, but rather the largest distributed calculator network in China through the humming mind.
The company has found that there are a large amount of dispersed, idle, insulated third-party idle computing power resources in society, and that there is a significant mismatch between large-scale computing power demands and dispersed and underutilized computing power supplies.
As a result, the company has built a platform for computing and moving them as “private vehicles” to pool and manage the idle computing power of isomer resources and deploy them to clients in the best way possible, thereby accelerating content transmission and increasing the low-delayed computing experience.
At the time of the interview, Yao Xin, the founder, compared the company to a “calculative drop”, stating that the company did not purchase, own large GPU servers and data centres, built its own smart-calculations parks on its own, consolidated all its resources externally and matched supply and demand through a self-study distributed dispatch system.
It can be said that the core assets for the future are not hardware, but rather an extremely intelligent distributed cloud computing structure and unified movement control platform. It is through this capacity that the company has managed to build a marginal computing network by bringing more than 4,600 idle nodes in more than 1340 counties and municipalities across the country to the bottom of the network.
This “light assets, heavy movement” pattern hit the super wind vent of the outbreak of LLM in 2023. The book shows that after the company's distributional movement, the GPU utilization rate remained above 75 per cent in 2025, well above the industry average of 40-50 per cent and that model reasoning costs could theoretically be reduced by more than 80 per cent. That advantage was magnified when the time of AI came.
It's not Yao Xin's first business.
In 2004, Yao Xin, who is studying at the China-China University of Science and Technology, and in an effort to address the pains of the students watching the game live in their dormitories, the P2P technology's “quicker and more people” character was captured.
In the golden age of the Internet, where broadband and distribution infrastructure were extremely scarce, PPTV, once home to 450 million global users, became the second largest video platform in the country after the coolest, and was eventually acquired by the Sunning Group in 2014.
In 2018, Yao Xin again joined his old partner, Wang Shao, the Chinese science brother, in forming Pie Ouyun Calculator (Shanghai) Ltd. (hereinafter referred to as Pie Ouyun, one of China’s combined entities), and now together leads the pie thinking about future capital markets.
According to the book, Yao Xin and his spouse, before IPO, controlled approximately 48.05 per cent of the shares through several entities; Wang Jiayu held 10.17 per cent of the shares indirectly, while other investors included Millenium, IOM Ventures, CCV Fund, China Tianqing, Rock, Boiling Point, etc.
Two years double, but why do you lose?
Currently, future business models are based on cloud-based fees, i.e., the use of power and the use of long-term fees. Major paid clients include large Internet enterprises and small and medium-sized AI start-ups.
From 2023 to 2025 (hereinafter referred to as the reporting period), corporate income grew rapidly from $358 million to $770 million, a compound annual growth rate of 46.6 per cent.
Of these, marginal cloud computing services (including marginal nodes services and marginal CDNs) account for 358 million yuan, 548 million yuan and 651 million yuan, respectively, and account for over 84 per cent of total income over the long term.
The income from AI cloud computing services rose sharply from $2.65 million in 2023 to $119 million, more than 10 times the same rate in 2025, which increased to 15.5 per cent, making it the second growth curve highlighted in the book.
There has been a sharp increase in revenue while there has been a marked decline in the Māori rate. During the reporting period, its combined Māori ratio slipped from 17.7 per cent in 2023 to 12.3 per cent in 2024, to 9.4 per cent in 2025, which is close to cutting.
Behind this, the company has been in deep loss. During the reporting period, the net losses were 189 million yuan, 294 million yuan and 223 million yuan respectively. If non-cash items, such as changes in the fair value of converted shares, are removed, the adjusted net loss continues to rise from $37 million to $105 million, falling into a situation of “more sold, more lost”.
The causes of the losses were mainly in the fast-growing AI cloud computing operations.
In 2025, its Ai cloud calculated Māori rate was 10.7 per cent, compared to -91.3 per cent in 2023 and -95.1 per cent in 2024, respectively, and although the Māori rate has been reduced, it remains a “loss trade” and its income is increasing, slowing down overall Māori performance.
The Front Line of Enterprise has learned that the company's marginal cloud operations are mainly aimed at bringing the computing power closer to the user, specifically to speed up video, live broadcasts and downloads of documents (i.e. edge CDN).
The AI Cloud business, on the other hand, provides GPU calculator and off-the-shelf LLM interface (Maas), where customers can run AI reasoning without buying GPU of their own. However, this revenue is derived mainly from model API, which transfers LLM to customers and charges Token.
Token prices, however, are not intended for the future, but rather for official model pricing. According to the love analysis, the price-fixing page of Pie Yoon and official API files such as DeepSeek, GLM and Kimi are priced consistently.
The reason is that the pricing rights are in the hands of the model manufacturer. Token API is a homogenized commodity, with a very low user-to-to-element cost and threshold. Therefore, even if the reasoning of the LLM manufacturer is not sufficiently supplied, it would be difficult for him to consider future premiums.
At the cost end of the project, the future GPU calculates mostly by leasing, with GPU resources costing $117 million in 2025, almost evening the overall AI cloud income ($119 million). It can be said that every additional token it provides, it pays an extra GPU rent.
From this point of view, while companies refer to themselves as “calculative drops”, the drop-off pricing rights are in their hands, and the future is a model manufacturer with pricing rights, and the downstream is a switchable customer, both of which are squeezed.
However, the demand side is indeed very tight, with the company's disclosure of an average daily Token consumption of trillions of dollars, suggesting that there is indeed widespread demand in the market, but that it is not equal to earning capacity, and that the company's profit space has not yet been opened.
In the equity book, the company also suggests that: “As the use of the company's products and services increases, customers may require lower pricing and some may choose to use other cloud service providers for commercial, compliance or other reasons. To this end, companies are required to invest more in sales and marketing in order to maintain or increase their revenues, which can have a negative impact on the company's gross domestic product.”
Thus, the company's low Māori rate is not a pain, but a natural limitation to this “rent card-sale token” model. In addition, the low threshold and high homogeneity of the calculator track makes it difficult for those who want to break out in the future.
We've only got two months left.
It is worth noting that the company's operating lifeline is on very few large clients.
Between 2023 and 2025, the top five clients of the company accounted for 92.5 per cent, 89.5 per cent and 79.0 per cent of the total, respectively, and were highly dependent on large clients.
Faced with large clients, the expectation of a longer payback cycle in the future has led to a continuing negative cash flow for corporate operations.
According to the equity book, companies usually grant a credit period of one to three months to customers of marginal cloud computing services, with trade receivables ranging from $90 million to $150 million and $168 million, respectively, between 2023 and 2025.
During the same period, trade receivables were revolving for 83 days, 79 days and 75 days, respectively. During the reporting period, 93.6 per cent, 88.1 per cent and 74.9 per cent of trade receivables were due from five major clients.
There is usually a time difference between the collection of user-based costs from clients and payments to suppliers, leading to a potential mismatch in operating cash flows. During the reporting period, net operating cash flows were -$35 million, -$41 million and -136 million respectively.
Companies operate with considerable reliance on financing. From 2023 to 2025, the cumulative equity financing was $60 million, $60 million and $106 million, respectively, while interest-bearing banks borrowed $44 million, $74 million and $81 million, respectively, rising each year.
IPO, the Hong Kong equity firm, intends to raise funds to upgrade the company ' s technical capacity for distributed cloud computing services and upgrade AI cloud computing services; expand overseas, particularly in mature markets and emerging markets with untapped potential; strategic investment and acquisition; and operating capital and general business use.
In addition to the pressure on operating funds, there are reasons why the future is eager to be listed, as well as “foreclosure pressure” from hanging over the top.
According to the book, the future is being followed by investment agencies such as Renchi, Founding Partners, and Tiansheng, as well as by the early silos of Mawakawa, Wang Xiaochuan, former President of the China District of Microsoft Azure, Zhen Yuanqing, and Tang Binseng.
From $46 million after the Angel Round in 2021 to $469 million after the B Round in 2025, to the B-2 round just completed at the end of May and early June 2026, the post-investment valuation for the future has reached $576 million.
The money that has been sequestered, while sustaining business expansion, may also become a debt that is due to be repaid. It was also the people who held the right of redemption at this very moment.
Between 2023 and 2025, the company recorded losses of $145 million, $222 million and $88 million, respectively, on the fair value of the converted and redeemable shares.
According to the agreement of the Pre-IPO investor when entering the stock, the investor is entitled to sell the shares back to the company at 150 per cent of the original investment, plus accrued dividends, if the “qualifying listing” was not completed by the agreed deadline (seven years after the date of the original offer and five years after the end of the B round in April 2025), or if the VIE structure has failed or is terminated.
In June 2025, the company entered into a supplementary agreement with the investor to “suspension” the right of foreclosure upon filing an application for listing, but the right of foreclosure will be restored if the application is rejected by the port of Hong Kong, the company voluntarily abandons the listing or the listing is not completed by 31 December 2026.
It was hoped that the first statement would be delivered in June 2025, after which it would expire, and that it would be resubmitted only on 10 June 2026, now more than two months before the listing.
In the wind, the token factory's business is indeed hot, but once it enters the capital market, its new story will eventually go back to the books of the “gold and true”.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.