Verified by trillions of token calls, how does PPIO build a new-generation intelligent token factory?

The Token Economy Era Has Arrived

In 2026, the global AI industry is undergoing a power transition from training to inference. At a State Council Information Office press conference in March, the National Data Administration released a set of statistics: as of March 2026, China's daily average Token call volume has exceeded 140 trillion, representing a more than 1000-fold increase from 100 billion at the start of 2024. Within just two years, this metric has grown by four orders of magnitude.

Tokens are no longer an obscure unit of measurement found only in technical documents. They have become the most fundamental transactional unit in the AI era — comparable to electricity, water, or CPU core counts in the past, with the key distinction that their rate of expansion far outpaces that of any other foundational resource in human history.

IDC also forecasts that global annual Token consumption will surge from 0.0005 Peta Tokens in 2025 to 150,000 Peta Tokens by 2030, marking an astonishing compound annual growth rate of 3418%. By 2031, the number of active global agents will reach 350 million.

Driving this explosive growth is the structural shift taking place across the AI sector.

Prior to 2026, the AI industry's primary battleground was training, where the focus was on stacking GPUs, optimizing parameters, and chasing benchmark rankings. But after 2026, inference is emerging as the main engine of computing power consumption. IDC estimates that China's AI server market will reach 3.5 trillion yuan in 2026, with the demand structure having shifted from "training-driven" to "dual-wheel driven by training and inference," and shipments of inference servers are now approaching those of training-oriented models.

Beyond the continuous expansion of inference demand, the emergence of agent applications and the precipitous drop in inference costs are all fueling the industry's explosive growth. A growing consensus is forming across the sector: AI is evolving from "answering questions" to "performing practical tasks." At this year's WAIC, multiple agent applications and mobile devices were launched simultaneously, and every instance of an agent's autonomous planning, tool invocation, and multi-step execution consumes Tokens at an exponential rate.

This transformation is also having a disruptive impact on computing power infrastructure.

YAO Xin, Co-founder and CEO of PPIO, told 36Kr: "Traditional cloud computing services are designed for humans, where programmers spin up a virtual machine that runs for days or even months. In contrast, Agent tasks are fragmented and high-frequency. The smallest billing unit on the PPIO platform's Sandbox is already measured in seconds."

Human cloud usage follows peak and trough patterns aligned with work and sleep cycles. But Agents utilize cloud resources 24/7 without interruption.

There are also notable differences in latency requirements. Humans can tolerate latency at the second level, but when an Agent executes a complex task requiring dozens or even hundreds of repeated calls, each few hundred milliseconds of latency is amplified through cycles, making task execution efficiency unacceptably low.

These disparities mean that cloud computing architectures designed for human users cannot directly serve Agent workloads.

Against this backdrop, Token factories have become an indispensable component for regulating computing power capacity, drawing significant market attention.

According to CIC Consulting data, based on average daily Token consumption in 2025 and the first quarter of 2026, PPIO ranks first among independent AI cloud service providers in China. In April 2026, the platform's daily average Token consumption reached 1.03 trillion operations; by June, this figure further exceeded 1.2 trillion operations, representing a more than 8-fold increase compared to the same period in 2025.

Additional disclosures show that PPIO's AI cloud revenue surged from 10.387 million yuan in 2024 to 119.2 million yuan in 2025, marking a year-over-year increase of over 1000%. The platform's total registered global developers grew from 125,000 at the end of 2024 to more than 666,000 by June 2026.

Solving the Efficiency Challenge: The Smart Token Factory Has Arrived

As inference costs decline 10-fold annually, business models that only provide raw computing power are rapidly losing their premium value. The real value lies in the ability to deliver Tokens with higher efficiency, lower costs, and better user experiences.

The "Smart Token Factory" as defined by PPIO is a large-scale production and delivery system optimized around the full lifecycle of Tokens.

YAO Xin told 36Kr that while the concept of the Token factory rose to prominence at the GTC 2026 conference in March, PPIO has been working on inference services since 2023 and launched its MaaS platform back in 2024, accumulating more than three years of expertise in this domain.

"Token factories themselves are not a novelty today. What matters more is how to continuously elevate the intelligence level of Token factories," YAO Xin argues. "You can buy a batch of computing resources and deploy models, which seems to enable Token production. But the real question is: how do you produce smarter Tokens with high quality and high efficiency?"

YAO Xin has therefore proposed a core formula for the Agent era: Agent Productivity = Token Intelligence Density × Agent Loop Duration. Here, Token intelligence density determines the upper quality limit of each decision an Agent makes, while Agent Loop duration determines how long an Agent can run continuously and how complex the tasks it can complete.

Centered on this formula, PPIO's Smart Token Factory has pioneered the launch of an intelligent model gateway in China, which acts as the intelligent scheduling center for AI Agents — critical decisions are validated by hybrid models to enhance quality, while simple tasks are automatically routed to lightweight models through model scheduling, ensuring that Agents complete tasks at the lowest possible Token cost with the highest intelligent performance, and continuously boost Token intelligence density.

Regarding hybrid models, YAO Xin revealed to 36Kr that PPIO is currently testing combinations of two to three different models, allowing them to compete and deliberate with each other. In internal tests, this hybrid model approach has already outperformed GPT-5.6 in certain task execution scenarios. "While there may still be a slight gap in individual model performance, by investing a certain amount of computing power and Token consumption, we can achieve stronger final task execution capabilities." This implies that the intelligence ceiling of models is not limited to internal parameters, but can also be broken through externally via engineering methods.

This is exactly how the intelligent model gateway operates: by providing a hybrid inference system for AI applications, the intelligent model gateway turns every model call into a "consultation of experts."

Over the years, PPIO has built full-stack AI cloud capabilities from the ground up, spanning GPU clusters, inference service optimization, and in-depth understanding of application scenarios, which reflects its ultimate pursuit of efficiency.

YAO Xin cited three typical scenarios for us: human conversational interaction, Agent invocation, and AI programming, each with completely different performance requirements. Human users prioritize sub-second response, Agents require millisecond-level latency, while AI programming demands the highest accuracy with minimal errors. Performing customized optimization based on different scenarios is a key differentiator between PPIO and other platform providers.

In terms of specific product capabilities, PPIO's differentiated approach is also reflected in several dimensions:

Rapid integration, out-of-the-box usability.

Platform neutrality, unbound by any model ecosystem. YAO Xin disclosed to 36Kr that an average application developer needs to invoke 10 to 15 different types of models, and switches models every three months as they iterate. The PPIO platform supports unified access for over 200 open-source models, allowing users to switch between models by modifying just one line of code. This neutral positioning has built trust among developers and represents a key strategic choice for market penetration.

Self-developed inference acceleration engine, deeply optimized for Agent invocation patterns. Unlike the general-purpose architectures of traditional cloud computing, PPIO's entire technology stack is fundamentally designed to adapt to the fragmented, high-frequency, and continuous characteristics of Agent workloads.

It is worth noting that the industry average GPU utilization rate typically ranges between 40% and 50%, while PPIO has long maintained a rate above 75%. YAO Xin told 36Kr that the company leverages its distributed computing power scheduling capabilities to interconnect more than 5,000 computing nodes across six continents worldwide, relying on time-zone staggered scheduling between the eastern and western hemispheres to seamlessly integrate workloads across different times, locations, and request frequencies, ultimately presenting a nearly straight utilization curve in the background.

Against the backdrop of continuously rising computing power costs, the ability to fully utilize every bit of generated computing power directly determines a company's long-term profitability. PPIO's performance on this metric forms its critical competitive moat.

Agentic Cloud: The Next Infrastructure Wave of Token Consumption

The Smart Token Factory addresses the question of "how to efficiently produce Tokens today." The Agentic Cloud answers the more ambitious proposition of "who will consume Tokens tomorrow, and in what manner."

In 2026, global cloud giants have simultaneously unveiled their Agentic Cloud strategies. Google introduced its Agentic Cloud strategy at the Next'26 conference, launching Agent Engine and Agentic Data Cloud; Alibaba Cloud announced the completion of full-stack agentization upgrades; Amazon launched AgentCore; and Microsoft's Foundry Agent Service officially went commercial. The sole shared goal is to position Agents as the core users of cloud services.

At WAIC 2026, PPIO officially released its new "Agentic Cloud" positioning. According to YAO Xin, the core logic of this strategy is: The primary users of cloud computing are shifting from humans to AI Agents.

The industry trend is also moving in this direction: the autonomous reasoning, tool invocation, and multi-step workflows of AI Agents are characterized by continuous, high-frequency, and intensive Token consumption, placing new demands on cloud infrastructure.

YAO Xin also shared a set of data with 36Kr: of the 8 billion people in the world, perhaps only 20% have used AI, and only about 5% are paying users, which means there is still huge room for growth in human AI usage. At the same time, Agent invocations are growing exponentially. Taking PPIO as an example, with more than 300 employees, there are already nearly 1,000 Agents running in the background, handling development, operations, customer service, and other work in an agentized manner. One person can own 10 or even 100 Agents, each consuming Tokens 24/7.

Based on these diverse requirements, PPIO's Agentic Cloud product architecture is divided into three layers: the infrastructure layer, the model service layer, and the Agent Harness platform layer.

It is worth special note that Harness, as an engineering framework dedicated to constraining, guiding, verifying, and correcting Agent execution, covers all segments beyond the large model itself: context construction, tool orchestration, verification loops, cost control, and observability.

The Sandbox is one of the core components of Harness. Last year, PPIO launched China's first Agent Sandbox compatible with the E2B interface, providing a securely isolated runtime environment for Agent Harness.

YAO Xin mentioned that when the Sandbox was released last year, the industry's understanding of Agents was still limited to viewing them as the "hands and brains of model invocation." By the beginning of this year, with the popularity of open-source Agents like OpenClaw, the industry began to realize that the real pain point lies in the continuous execution of long-horizon and complex tasks.

According to official PPIO disclosures, the PPIO Agent Sandbox features a cold start latency of less than 200ms, with system-level security isolation that ensures each Agent task runs in an independent virtual machine environment; it supports the simultaneous creation of tens of thousands of sandboxes, and billing automatically pauses when tasks are idle, resulting in an overall usage cost that is over 90% lower than comparable products. In the year since the PPIO Agent Sandbox was launched, its business scale has grown by more than 120 times.

Today, a distinct pattern has emerged in Token pricing rankings: programming commands the highest price, followed by Agent usage, and conversational services for human users are the lowest. This ranking itself clearly indicates the future direction of Token consumption.

YAO Xin believes: "The users we serve have shifted from humans to Agents, and possibly to robots in the future."

This transformation is reshaping the business model of cloud computing. In the past, cloud computing competition was measured by the number of features and the degree of ecosystem closure. But in the Agent era, competition is defined by who can enable silicon-based life to run faster, cheaper, and more stably. PPIO's choice is: not to build a closed garden, but to embrace open source; not to pursue full-stack substitution, but to act as a complementary partner in the AI era.

The Token economy is redefining the value metric of computing power, and in this reconstruction, only the most efficient players will secure the final ticket to success.