Located in Wudaokou, Beijing, Infinigence AI's office is less than a ten-minute walk from the South Gate of Tsinghua University, surrounded by a cluster of AI companies such as Zhipu AI and ModelBest. This AI infrastructure company, spun off from the Department of Electronic Engineering at Tsinghua University, has just celebrated its third anniversary. One piece of data from them is worth noting: By the end of July 2026, the average daily Token invocation volume on Infinigence AI's Agentic MaaS platform has grown by more than 40 times compared to the end of 2025, doubling on average every two weeks since the end of January.
Recently, in an interview with Infinigence AI's CTO Li Boxun, Semiconductor Industry Insights explored why the competitive focus of the AI industry is shifting from chips to Tokens. He stated, "Tokens are transitioning from a technical unit of measurement to a real economic variable." While some are still competing on chips, Infinigence AI has already started competing on Tokens.
01 Technology Equals Profit
Over the past few years, the competitive focus of the AI industry has been very clear: who has better chips and larger models. However, Li Boxun, CTO of Infinigence AI, has observed that this logic is undergoing a fundamental shift. He abstracts the development of AI into a simple formula: scale multiplied by efficiency, and further explains: "Previously, the industry focused on how to improve model capabilities. At that time, inference tasks had not yet scaled up, and the scale of computing resources could support the demand. Now, model capabilities have leaped forward, inference demand has exploded, scale has become a bottleneck, and efficiency issues have become more prominent."
Data from the National Data Administration corroborates this trend. By March 2026, the average daily Token invocation volume in China had exceeded 140 trillion, an increase of over 40% compared to the end of 2025. IDC predicts that by 2026, the Token consumption in China's MaaS market will reach 40 quadrillion. The Token economy is moving from concept to reality.
This explains why the concept of a "Token factory" has suddenly gained traction. Given the limited scale of computing resources, whoever can produce more Tokens with fewer resources can establish a cost advantage. Over the past year, Infinigence AI has reduced inference costs by more than 10 times through full-stack technical optimization. In Li Boxun's words, "Only when you can push costs low enough can you generate profit; technology equals profit."
The AI productivity formula proposed by Infinigence AI: AI Productivity = Scale of Intelligent Resources × Token Conversion Efficiency × AI Productivity Conversion Efficiency, is precisely the quantitative expression of this logic. At the 2026 World Artificial Intelligence Conference (WAIC), Xia Lixue, Co-founder and CEO of Infinigence AI, officially released the strategic layout of "Front Store, Back Factory, and One Center." The three focal points clearly correspond to the three factors of the AI productivity formula: the computing power distribution center maximizes the scale of intelligent resources, currently accessing over 37,000P of computing power and covering 16 mainstream chips; the Token factory solves Token conversion efficiency, achieving a 10-fold reduction in inference costs; and the AI productivity store opens up the channel from Tokens to industrial value, implementing Infra solutions in industries such as healthcare, law, energy, and gaming. From chips to Tokens, Infinigence AI has completed the switch in industrial logic.
02 From Serving Humans to Serving Machines
If the reduction in Token costs is the current main thread of the industry, then the changes brought by Agents point to a more distant future. When Agents enter real production environments, the consumption pattern of Tokens changes completely. In a bank credit approval scenario, traditional AI chat consumes less than 10,000 Tokens, while after multiple rounds of tool calls by an Agent, it can surge to 470,000. Consumption shifts from "single, short-chain, predictable" to "continuous, long-chain, unpredictable," subverting the design premises of traditional cloud infrastructure.
Li Boxun believes that behind this lies a deeper paradigm shift: "Our current computer protocols and programming languages were defined by humans in the past, so that humans could write and supervise them. But when AI develops to be strong enough, the positioning of infrastructure must shift from serving humans to serving machines." The Agentic Infra proposed by Infinigence AI is born precisely for this purpose: a new generation of infrastructure built specifically for Agents. Traditional architectures are designed for human interaction, while the consumption pattern of Agents is completely different; the underlying architecture must change accordingly.
We also saw that in July, Infinigence AI released the industry's first cross-cluster heterogeneous PD separation architecture (Prefill-RelayDecode-MainDecode, PDD), which innovatively deconstructs the traditional two-level "P-D" separation into a three-level "P-RLD-MD" separation inference architecture. This can bring out the respective advantages of domestic chips and pursue the optimal solution for the combination. Actual tests show that this architecture achieves a 51.5% reduction in first Token latency and a 37.5% reduction in per-Token production cost.
As for the next explosion point after Agents, Li Boxun's judgment is: "It lies in whether new productivity can be born and bring new production value." He depicted a scenario: in the future, apps might not even exist; when a new task comes, the model directly writes an application on the spot, runs it, and finishes. The form of expression is not entirely certain, but the explosion of productivity is. There is a general judgment in the industry: from human user-driven, to human-Agent coexistence, to Agent-dominated, the global Token demand will make an exponential leap along the path from humans to machines. After 2027, the scale of Tokens autonomously consumed by machines will far exceed that of humans.
03 Building Bridges on the "Tower of Babel" of Chips
One of the core values anchored by Infinigence AI from the very beginning can be summarized with a simple formula: the efficient collaboration of M models × N chips. Chips from different manufacturers have different interfaces, architectures, and standards, each speaking its own dialect, and computing resources are scattered everywhere. What Infinigence AI does is to build a simultaneous interpretation system among these dialects, allowing different chips to understand each other, and enabling different models to be efficiently deployed on different hardware.
In 2025, Infinigence AI took the lead in the industry in achieving the commercial deployment of large-scale mixed training based on diverse heterogeneous chips. It completed the world's first mixed training of a 70-billion-parameter large model using six types of chips in a "4+2" configuration (including Huawei Ascend, Iluvatar CoreX, Metax, Moore Threads, AMD, and NVIDIA), with the computing power utilization rate of the thousand-card cluster reaching up to 97.6%. In 2026, the company further broke through in cross-domain collaborative training, completing the first domestic cross-domain communication training for large models over 4,000 kilometers in collaboration with China Telecom, connecting heterogeneous clusters in Hami, Xinjiang, and Shenzhen, Guangdong, and improving joint training performance by 165%. Currently, Infinigence AI's Agentic Infra autonomous infrastructure platform has accessed 37,000P of computing power and 16 types of chips, reducing the adaptation time for newly launched chips to 2 days and cutting the adaptation cost for domestic chips by over 70%.
What Infinigence AI does is essentially redefining the organization of computing power between chips and models. The customer list also corroborates this: the platform has served leading enterprises such as Kimi, Zhipu AI, Baichuan AI, Lenovo, Li Auto, and China Mobile, with a model accuracy alignment rate of over 99.9% and an enterprise-level high availability rate of 99.95%.
04 Tsinghua Heritage and Full-Stack Thinking
The team at Infinigence AI has a Tsinghua heritage.
Wang Yu, the initiator of the company, is a tenured professor in the Department of Electronic Engineering at Tsinghua University. He initiated the founding of DeePhi Tech, which was later acquired by Xilinx, the world's largest FPGA vendor, and has rich experience in scientific research innovation and industrialization of achievements. The founding team of Infinigence AI also almost entirely comes from the NICS-EFC Laboratory of the Department of Electronic Engineering at Tsinghua University, including CEO Xia Lixue, Chief Scientist Dai Guohao, and CTO Li Boxun. Wang Yu previously stated: "The gap between exponential demand and linear supply is growing wider and will continue to widen. This optimization goal must be achieved through cooperation; only by working together can we elevate the linear supply capacity." In his view, what Infinigence AI aims to do is not just to solve the problem of computing power scheduling, but to build infrastructure for all intelligent operations in both the digital and physical worlds.
This deep integration of academia and industry is directly reflected in Infinigence AI's talent cultivation logic. Li Boxun himself graduated from the Department of Electronic Engineering at Tsinghua University and has personal experience in talent cultivation. When asked what kind of talent the industry lacks most today, his answer was straightforward: "People with full-stack thinking." Traditional computer architecture is layered and decoupled: those who work on instruction sets, those who do chip design, and those who do algorithms have very little intersection with each other, cooperating through standard protocols and intermediate layers. But AI's depth requirements for the technology stack are completely different.
Besides full-stack capabilities, Li Boxun also mentioned another key change: AI is changing the capability structure of students in electronic information. In the past, it was necessary to specialize in solving a specific core problem; now, many problems can be solved by AI. "AI amplifies the value of the broad scope and engineering thinking of electronic information students," he said, but the new challenge lies in whether students can form an independent capability closed loop, whether they can judge the direction of tasks executed by AI, and then execute it to a closed loop. Infinigence AI itself is trying to build a bridge from academia to industry. He mentioned that Infinigence AI participated in the China Graduate Electronic Design Competition very early on and set up enterprise topics, extracting real problems seen in the industry so that more students can participate. Many of the most interesting ideas and the most advanced concepts are born in universities, so it is essential to maintain good interaction with them.
05 Conclusion
A 97.6% heterogeneous cluster computing power utilization rate, a 40-fold increase in Token invocation volume, a 10-fold optimization in inference costs... this series of numbers outlines Infinigence AI's initial coordinates on the AI infrastructure map. When the focus of the computing power competition shifts from chips to systems, and from hardware stacking to efficiency optimization, it has already secured a position in the front row of this new battlefield.
But for Infinigence AI, this is not the whole story. In China, the issue of computing power has never been just about making better chips. Here lies the world's most complex heterogeneous computing power landscape, with Huawei Ascend, Iluvatar CoreX, Metax, Moore Threads, Hygon... There are intelligent computing centers scattered across the country that urgently need to be awakened, as well as the world's richest AI application scenarios. Organizing these fragmented computing resources so that they can be scheduled, metered, and used like water and electricity is itself a proposition of the era.
When domestic chips start running large models on the Infra platform, when more and more developers invoke model capabilities on demand through Token factories, and when young students turn real industry problems into their own research topics, what Infinigence AI is doing is not just building a commercial moat for a company, but paving a path from computing power to productivity for China's AI industry. As Li Boxun said: "One of our greatest ideals is to ensure that China, in the future AI era, always possesses the world's most advanced AI infrastructure."