The AI computing power sector has a new player racing toward a Hong Kong IPO.

On June 10, Pixiang Future Group filed its listing application with the Main Board of the Hong Kong Stock Exchange.

The company’s backstory is quite intriguing. Founder Yao Xin previously built PPTV, utilizing P2P distributed scheduling technology to activate idle bandwidth across networks. After successfully commercializing P2P streaming media, he sold it to Suning.

This time, he has shifted focus from ‘idle bandwidth’ to ‘idle computing power’ in the AI era.

The numbers are eye-catching. Over the past year, Pixiang Future’s AI cloud revenue skyrocketed from RMB 10.387 million to RMB 119 million, marking a tenfold year-on-year surge. Based on daily Token consumption in 2025, the company has become China’s largest independent AI cloud service provider.

However, the flip side is equally striking. In 2025, AI cloud business revenue hit RMB 119 million, but computing resource procurement costs reached RMB 117 million, resulting in a full-year gross margin of -10.7%.

This seemingly paradoxical situation stems from Pixiang Future’s core business model: instead of building its own AI computing infrastructure, it aggregates idle computing power from the market and delivers services through a unified scheduling network. This approach makes Pixiang Future both unique and controversial.

Today, Silicon-Based Gentleman will delve into this AI cloud company that has emerged by integrating idle computing resources.

/ 01 / AI Cloud Revenue Rockets Tenfold in a Year!

Like most AI computing power firms, Pixiang Future has experienced explosive growth.

According to its prospectus, from 2023 to 2025, Pixiang Future’s revenue reached RMB 427 million, RMB 958 million, and RMB 1.334 billion, respectively, with a compound annual growth rate (CAGR) of 76.7%.

Pixiang Future’s business is essentially distributed cloud computing. In simple terms, it integrates idle computing power nationwide through a unified scheduling platform, transforming it into standardized computing nodes closer to users.

Specifically, this involves edge cloud and AI cloud services.

Edge cloud forms the foundation of Pixiang Future’s business, primarily integrating idle servers from internet cafes, small and medium-sized enterprises, and local IDCs. It is well-suited for scenarios like live streaming and short videos. From 2023 to 2025, edge cloud revenue accounted for 99.93%, 98.14%, and 84.5% of total revenue, respectively.

However, growth in this core segment is slowing. Due to the maturation of traditional internet audio and video distribution scenarios and ongoing price wars among public cloud giants like Alibaba Cloud, Tencent Cloud, and Huawei Cloud, edge cloud revenue growth plummeted from 53% to 19%.

The real growth driver for the company is AI cloud.

This business only commenced in 2023, reaching RMB 10.387 million in scale in 2024. By 2025, revenue soared to RMB 119 million, a tenfold year-on-year surge, with its share of total revenue jumping from 1.9% to 15.5%.

Operational data reveals exponential growth. The number of AI cloud computing nodes increased from 5 in 2023 to 64 in 2025. From 2024 to 2025, daily Token consumption on the AI cloud platform surged from 27.1 billion to 271 billion.

Behind this frenetic growth lies a development path starkly different from mainstream AI cloud providers.

AI cloud providers like Coreweave and Nebius follow a capital-intensive approach, investing heavily in GPUs and data centers before leasing them out.

In contrast, Pixiang Future adopts a purely asset-light model, leasing idle GPUs and server resources from upstream suppliers, integrating them through its own scheduling system, and then delivering computing power services.

The advantages of this model are clear: minimal capital expenditures and rapid scalability.

From 2023 to 2025, Pixiang Future’s cumulative capital expenditures amounted to just RMB 87 million, less than 5% of revenue during the same period. In comparison, Coreweave’s capital expenditures in Q1 2026 alone reached USD 6.8 billion, 3.3 times its revenue for the same period.

Meanwhile, building data centers—from GPU procurement to deployment—typically takes 12-18 months, often missing the window of rapid computing power demand. Pixiang Future, through leasing, can complete computing power expansion in 1-2 months, quickly meeting sudden market demands.

This is the key reason its AI cloud business grew so rapidly within a year. However, it also brought challenges to Pixiang Future.

/ 02 / Where Did the Profits Vanish? The Profitability Dilemma of the ‘Computing Power Middleman’

The asset-light model is Pixiang Future’s most compelling narrative—and its most daunting challenge.

Despite strong apparent demand, generating profits remains elusive for Pixiang Future.

Financially, from 2023 to 2025, the company’s adjusted net loss widened from RMB 37.1 million to RMB 105 million.

The root cause lies in declining gross margins and cost pressures.

From 2023 to 2025, the company’s overall gross margin steadily declined from 17.7% to 12.3%, and then to 9.4%, dropping by 8.3 percentage points over three years—nearly halving.

Breaking it down by business, edge cloud computing’s gross margin fell from 17.8% to 13.0%, primarily due to multiple rounds of price wars initiated by public cloud giants like Alibaba Cloud, Tencent Cloud, and Huawei Cloud, which continuously squeezed the company’s pricing margins.

More concerning is the AI cloud computing business. Despite its revenue surging tenfold, this segment has yet to achieve profitability, with a gross margin of -10.7% in 2025.

The prospectus reveals a harsh reality: in 2025, the company’s computing resource procurement costs for AI cloud computing alone reached RMB 117 million, nearly equal to its RMB 119 million in revenue for the same business. Essentially, after a year of hard work, nearly all revenue went to upstream computing power suppliers, leaving little profit margin.

Why is this happening? The answer lies in Pixiang Future’s asset-light model.

Traditional self-built computing power generates profits through economies of scale. Once a GPU server is purchased, depreciation and electricity costs remain largely fixed. Increasing utilization from 50% to 80% converts most of the additional revenue into profit, as fixed costs are spread across more orders.

However, Pixiang Future ‘purchases’ computing power. When a customer buys one hour of GPU time, Pixiang Future must also purchase one hour of GPU time upstream. While higher customer usage increases revenue, procurement costs rise proportionally.

The economies of scale it can achieve come only from bulk purchasing discounts, off-peak scheduling, and software inference optimization. As orders increase, most of the gains from higher utilization are first captured by computing power resource owners.

Even more problematic is that, despite not owning GPUs, Pixiang Future cannot escape ‘idle risk.’

AI cloud services require on-demand availability; the platform cannot wait for customer requests before sourcing GPUs from the market. To ensure service stability, it must pre-lock a certain scale of computing resources. Customers may not use them, but Pixiang Future must guarantee their availability.

This creates an awkward situation: the company does not own servers and cannot fully benefit from high utilization rates, yet it must still pay for some idle resources to ensure service quality.

This explains why, despite achieving an average GPU utilization rate of 75% in 2025—far higher than the industry average of 40%-50%—this ultra-high utilization did not translate into high profitability.

It also explains why most computing power leasing platforms, once demand stabilizes, gradually shift toward a capital-intensive model and begin purchasing their own GPUs.

For example, Parallella Technology initially relied heavily on external GPU resources but rapidly increased its proportion of self-owned GPUs as orders grew. AutoDL no longer solely provides GPU leasing but has also entered server sales and data center hosting.

Beyond profitability challenges, not building its own computing power severely limits Pixiang Future’s ability to serve major clients.

Individual developers and small-to-medium clients are insensitive to computing power sources, prioritizing affordability and usability. However, large enterprises care more about cluster performance, data security, and service stability, often requiring specific GPU models and dedicated clusters.

Pixiang Future’s aggregated heterogeneous resources make it difficult to fully control underlying servers and network environments. This positions it better for scattered, elastic, price-sensitive demands but makes it hard to secure high-value, large-scale enterprise orders.

While Pixiang Future can partially compensate through software improvements, such as enhancing scheduling efficiency and optimizing model inference, this remains unproven.

Until then, Pixiang Future resembles more of a computing power distributor than a true AI cloud company with economies of scale.

Text/Yuanyuan