In 2026, Token vendors are making intensive pushes into the capital market. PPIO and SiliconFlow have successively submitted listing applications to the Hong Kong Stock Exchange. Infinigence AI has secured cumulative financing exceeding 2.2 billion RMB, while Qunjing Technology has raised over 1 billion RMB within half a year. This sector lacks neither players nor capital.
However, a review of PPIO's prospectus reveals that the company, which processes over one trillion Tokens daily and ranks first among independent AI cloud vendors in China, is still losing money on its core Token business. Analysys has analyzed the complete prospectus data of PPIO to answer a critical question: Given the surging Token consumption, why are Token vendors losing more as they sell more?
01. Revenue Surges Tenfold in One Year, Yet Still Losing Money?
Since 2024, PPIO has pivoted from CDN to AI cloud services, offering model APIs, GPU cloud services, and Agent cloud services. The model API is the current focus, providing developers with Token services for large language and multimodal models, predominantly open-source ones. Let us first examine the core data of PPIO's AI cloud business.
Figure 1: Core Data of PPIO's AI Cloud Business
Within a single year, PPIO's AI cloud revenue grew tenfold, its gross margin loss rapidly narrowed to -10.7%, and daily Token consumption surpassed one trillion. The overall growth rate is exceptionally fast, accompanied by significant improvement. These figures have triggered two completely contrasting intuitions in the market. Optimists look at the slope: moving from -95.1% to -10.7%, profitability is merely a matter of time with one more year.
Pessimists look at the absolute values: despite a tenfold annual increase, the company is still losing money, and it may fail to cross the final hurdle. To determine who is closer to the truth, we must first examine the composition of the three product lines within the AI cloud business.
Figure 2: Analysis of PPIO's AI Cloud Product Lines
Based on Analysys's estimates, PPIO's AI cloud revenue primarily stems from model APIs, accounting for over 60%, which is also the primary driver of the negative gross margin in the AI cloud segment. The other two segments are GPU cloud services, which charge hourly compute rental fees with a low but positive gross margin, thus not dragging the overall margin into the negative; and Agent cloud services, which account for a small proportion and are not the main cause of the losses.
PPIO's daily Token consumption of one trillion demonstrates that the demand is genuine. However, Token pricing is not determined by PPIO but by the official model pricing. Analysys compared PPIO's pricing page with the official API documentation of DeepSeek, GLM, and Kimi, revealing that PPIO's pricing aligns perfectly with the official rates. This is not because PPIO is unwilling to charge higher prices, but because the pricing power lies with the model vendors.
Token APIs are homogeneous commodities; the cost for users to switch Token suppliers is merely changing a single line of code. Consequently, even if Kimi faces a shortage of inference compute supply, PPIO cannot charge a premium. On the cost side, most of PPIO's GPU compute is leased. In 2025, GPU resource costs reached 117 million RMB, nearly equal to the total AI cloud revenue. Therefore, every additional Token PPIO provides incurs an extra GPU rental fee. Amid continuously rising NVIDIA rental prices, PPIO's resource costs remain persistently high.
With revenue pricing power locked and resource costs consistently in an upward trajectory, this fully explains the negative gross margin of the AI cloud business.
02. Pricing Power for Tokens Lies Not with the Sellers
Among PPIO's two primary AI cloud services that generate the bulk of its revenue, the pricing power for GPU cloud services can be retained by the vendors themselves, whereas Token pricing power rests entirely with the model vendors. GPU cloud services are inherently profitable. For instance, AutoDL operates a consumer-oriented compute cloud platform targeting universities, achieving a revenue of 210 million RMB and a net profit of 26.69 million RMB in 2025. It sells GPU hours, and users pay for GPU resource services. Based on this, it is estimated that PPIO's GPU cloud services also yield a certain profit. However, Tokens are standardized commodities; identical models producing identical outputs cannot be differentiated.
Tokens are defined by upstream model vendors, but their pricing logic is not entirely driven by short-term commercialization considerations. When pricing Tokens, model vendors do not focus on the profit and loss of a single API business line; rather, they prioritize how many developers they can reach and whether they can attract more data feedback to support the next model iteration. DeepSeek prices V4-Flash at 1 RMB per million input Tokens. While this price fails to cover inference costs, model vendors do not need to profit directly from APIs; their primary goal is to maximize user adoption.
However, for PPIO, Tokens are the core source of revenue. Although it can retain developer usage data, such data trading falls into a gray area and cannot be openly monetized. In other words, every cent of PPIO's profit must be squeezed out of inference efficiency, meaning Token generation costs must be minimized to the extreme. Furthermore, the majority of model API revenue comes from small and medium-sized developers and AI application companies. These customers are extremely price-sensitive and possess no bargaining power. This pricing structure is clearly reflected in PPIO's customer composition.
In 2025, PPIO's AI cloud services had over 91,000 paying customers, with an ARPU of merely 1,300 RMB, indicating a customer base predominantly consisting of bargain hunters. Analysys believes that the root cause of PPIO's losses lies in the industry structure. It operates in a sector where pricing power is locked by upstream players; moving forward, profitability can only be achieved by pushing efficiency to the absolute limit.
03. Is There Any Chance to Turn the Gross Margin of the Token Business Positive?
Breaking down the variables, there are four main directions for improving future operational efficiency: GPU utilization rate, inference efficiency, model revenue structure, and Agent cloud services. Below, we examine these four variables individually.
The first variable, GPU utilization rate.
It goes without saying that a higher GPU utilization rate requires fewer leased GPU resources for the same output. PPIO stated in its prospectus that it has achieved a GPU utilization rate of 75%, compared to the industry average of 40% to 50%. Reportedly, the highest optimization level currently disclosed in the industry is Meta's research supercluster achieving 83% to 85%, making PPIO's 75% already outstanding. However, every incremental point increase in GPU utilization rate brings greater marginal difficulty. The most effective way to steadily improve the GPU utilization rate is to fill idle compute capacity during non-peak hours, and there is only one method to achieve this: running batch tasks rather than real-time tasks. Yet, batch tasks command lower prices, and even with full utilization, they struggle to make a significant contribution to the gross margin. Therefore, the room for improvement in GPU utilization rate is limited.
The second variable, inference efficiency.
Inference efficiency refers to achieving higher Token throughput for models on the same compute capacity through inference engine optimization. However, optimizing inference efficiency is constrained by model architecture. The optimization techniques are publicly available, primarily focusing on KV Cache and PD (Prefill-Decode) disaggregation. There are also numerous open-source projects like SGLang and vLLM, meaning the technical barrier is not particularly high. Furthermore, the room for optimization is diminishing; the first round of optimization might yield a multi-fold increase, while the second round sees a substantial drop in effectiveness. In contrast, model vendors can collaborate deeply with GPU chip manufacturers to optimize Token throughput at the foundational chip level. Moving forward, PPIO must also delve into the chip level to potentially develop distinct inference optimization capabilities. Therefore, while optimizing inference efficiency at the software level can currently narrow losses, it is highly unlikely to turn the gross margin positive on its own.
The third variable, model revenue structure.
The prevailing strategy for model API services is to use low-priced models to attract traffic and high-priced models to generate revenue. DeepSeek is a typical low-priced model, whereas Kimi and GLM are typical high-priced models. DeepSeek V4-Flash is priced at 2 RMB per million output Tokens, while Kimi K3 output is priced at 100 RMB per million, representing a fifty-fold price difference. If traffic is highly concentrated on the low-price end, the losses from low-priced models could entirely offset the profits from high-priced ones in aggregate. It is not just PPIO; SiliconFlow faces the same revenue structure issue. In 2025, SiliconFlow's public cloud revenue for model APIs was 29.3 million RMB, with a gross margin of -119%, which is even worse than PPIO's. Looking ahead, the catalyst for altering the model revenue structure lies in model routing. Developers will no longer need to decide which model to invoke; instead, models will be intelligently scheduled via model routing to minimize Token consumption while achieving task objectives. In this way, Token vendors can exert a certain degree of control over the revenue mix between low-priced and high-priced models. However, since the underlying premise remains reducing Token consumption, the room for improvement for Token vendors is still limited.
The fourth variable, Agent cloud services.
This represents the most promising direction for gross margin improvement among the four variables. Starting in 2025, PPIO began deploying Agent infrastructure and launched an intelligent agent sandbox product. Unlike APIs, which are billed per Token, sandboxes are billed by the second and can be paused during idle periods. One year after the sandbox went live, its business scale has grown by over 120 times. APIs and Agents represent two distinct business models. APIs sell standardized Tokens, with gross margins locked by pricing power. Agents sell runtime environments—sandboxes, tool invocations, memory systems, and agent hosting—which allow for differentiated pricing. If users are willing to pay for a stably running Agent environment rather than solely for Token consumption, the profit structure could be rewritten. Yao Xin, co-founder of PPIO, publicly stated that one year after the launch of PPIO's Agent services, the business scale has grown by over 120 times. Analysys believes that Agent cloud services represent the most critical path for PPIO to break through the ceiling of razor-thin margins.
Synthesizing the four variables, the room for improvement in GPU utilization rate and model revenue structure is highly limited, and the potential for enhancing inference efficiency is narrowing. Agent cloud services remain the only variable capable of breaking through the ceiling, though it will take time.
04. How Can Token Vendors Make Big Money?
PPIO's case is not merely a financial story of a single company; it answers an industry-level question: where is the ceiling of the Token business. Judging from revenue growth and scale expansion, this business model is viable. The average daily Token invocation volume has surged from the hundred-billion level to exceeding 140 trillion by March 2026. A thousand-fold increase over two years indicates that the market track is sufficiently broad.
PPIO achieving a tenfold revenue increase in one year and improving its gross margin from -95.1% to -10.7% proves that economies of scale genuinely exist in the Token segment. However, the ceiling for gross margins is equally real. In 2025, SiliconFlow invested 54.21 million RMB in free Token vouchers for customer acquisition. Excluding this subsidy, its public cloud business would yield an estimated gross margin of around 23%. Yet, this is merely a theoretical figure. Analysys estimates that the gross margin ceiling for Token vendors will likely fall within the 10% to 15% range.
What does this range imply? It means these companies can survive and scale their revenue alongside the growth in total Token consumption, but they will struggle to escape the flawed business model of traditional enterprise services. In a business with a 15% gross margin ceiling, after deducting R&D, sales, and administrative expenses, the net profit margin remains in the single digits. The financial data of PPIO and SiliconFlow expose the structural ceiling of Token distribution and scheduling as an independent business model. If this assessment holds true, the ultimate outcome of this industry will follow one of two trajectories.
One trajectory is where Agent cloud services successfully disrupt the profit structure, with vendors upgrading from selling Tokens to selling Agent runtime environments, thereby breaking through the 25% gross margin ceiling. The other is where Agents fail to scale, Token distribution remains perpetually low-margin, and vendors ultimately resort to generating gray-market profits by selling data to model vendors.