The Agentic Infra Strategy Backing Infinigence: A Choice by Four Leading Companies

At the WAIC (World Artificial Intelligence Conference) held in Shanghai on July 20, MiniMax and StepFun both appeared at an event hosted by the same AI infrastructure company. MiniMax participated in a strategic partnership agreement, while StepFun delivered a keynote speech. According to reports from QbitAI, this is not an isolated collaboration but rather an indication that China’s leading large language model companies are converging on common infrastructure.

A Shared Partner for Four Major Companies

This trend began four months ago. At the Zhongguancun Forum, Zhang Peng from Zhipu and Yang Zhilin from Kimi joined a roundtable discussion alongside a co-founder and CEO of a particular company. It was revealed during the event that this company was already providing services to Kimi and Zhipu.

Kimi, Zhipu, MiniMax, and StepFun—China’s four top foundational model companies—have each established deep partnerships with the same AI infrastructure company: Infinigence. As described in reports from QbitAI, Infinigence is likened to “the CATL of the battery industry.” While car manufacturers compete with one another, they cannot avoid relying on a shared layer for batteries. Similarly, Infinigence aims to become the “common choice” in the era of large language models.

The Growing Gap Between Demand and Supply

By 2026, the AI industry reached a pivotal turning point. Inference had surpassed training as the primary consumer of computational resources in AI. While the cost of inference fell 280-fold in just two years, total AI expenditure by companies has only continued to rise.

In China, the number of daily token calls has already surpassed 140 trillion, marking a 40% year-on-year increase. Demand is skyrocketing exponentially.

On the supply side, however, the situation is starkly different. The expansion of physical computational resources remains linear, and even the addition of more server rooms or GPU cards is unlikely to close the supply-demand gap within the next 3–5 years. This gap—an exponential growth in demand versus linear growth in supply—is exactly where Infinigence positions itself.

Adding to the challenge is the difficulty of determining from the outside whether models are being deployed properly. While models may respond normally to requests, output accuracy could quietly degrade by as much as 30% without detection. Often, such issues are only discovered when users notice a decline in business metrics, by which point the model’s trustworthiness has already been compromised.

Tackling Three Core Challenges

The decision by Kimi, Zhipu, MiniMax, and StepFun to partner with Infinigence stems from three fundamental concerns: whether the model’s effectiveness deteriorates, whether the costs are justified, and whether the system remains stable during issues.

Testing Standards for Accuracy Maintenance

One of the most significant challenges in ensuring effectiveness is the issue of “hidden accuracy degradation.” A QbitAI report cited an example where a third-party supplier deployed a model, resulting in a 30% accuracy drop compared to the original. The client only discovered the issue after noticing a decline in business metrics.

To address this, Infinigence has established a series of access testing standards. Every aspect, from tool invocation consistency to inference mode precision alignment, undergoes rigorous verification. New models must pass these tests before being added to the platform. Consequently, clients reportedly notice almost no difference in experience whether they use Infinigence’s interface or the original API.

Cost Reduction Through PD Segregation Technology

On cost optimization, Infinigence introduced cross-cluster heterogeneous PD segregation at this year’s WAIC. Inference for large-scale models involves two distinct phases—Prefill and Decode—which have different computational loads and hardware requirements. By segregating these phases during deployment, different types of chips can be utilized for tasks they perform best.

However, separating these processes across server rooms introduces new challenges. Post-segregation, transferring the KV cache between heterogeneous chips via wide-area Ethernet encounters limitations in bandwidth and latency.

To address this, Infinigence adapted its Radix Cache technology, originally designed for Decode instances, to this cross-cluster architecture, reducing data transfer volume by an order of magnitude. Additionally, it pioneered the world’s first PDD architecture, dividing the traditional PD link into three phases: Prefill, RelayDecode, and MainDecode.

In cases of high-latency requests, RelayDecode delivers tokens to users first, ensuring they experience no noticeable delay. Once data transfer is complete, the process seamlessly transitions to MainDecode. According to tests, this architecture reduces the time to the first token (TTFT) by 51.5% and cuts per-token costs by 37.5%.

Automating Operations and Maintenance

As the scale of clusters grows, ensuring stability becomes increasingly challenging. Managing the inference of large-scale models involves overseeing dozens to hundreds of clusters, supporting terabytes of daily traffic distributed nationwide. When servers go down, relying on human monitoring is insufficient.

Infinigence has introduced an “Intelligent Computing Cluster Operations and Maintenance Agent System,” which addresses operational challenges end-to-end in real-world production environments. Operating 24/7, the system transitions operations from “humans detecting issues” to “issues finding humans,” and ultimately to “issues resolving themselves.” This has enhanced operational efficiency by over fivefold and improved the speed of handling major incidents by sixfold.

The “Front Store,

Back Factory, Unified Center” Architecture

Infinigence has proposed a “Front Store, Back Factory, Unified Center” architecture. This model represents AI productivity as a formula: “AI Productivity = Scale of Intelligent Resources × Token Conversion Efficiency × AI Productivity Conversion Efficiency.”

The “Unified Center” refers to the computational resource aggregation hub, or the Agentic Infra autonomous infrastructure platform. China’s chip ecosystem is inherently fragmented; by consolidating and centralizing distributed computational resources, this center aims to provide an expandable and stable computational foundation for models and applications. The center currently deploys 37,000P of computational resources and supports 16 mainstream chip types.

Infinigence has also addressed the challenges of cross-cluster reinforcement learning at this layer. With the continued evolution of post-training scaling laws, reinforcement learning is seen as the key to unlocking advanced intelligence. By optimizing the network, platform, and framework layers, Infinigence has achieved seven days of uninterrupted, stable operation for cross-domain reinforcement learning training.

The “Back Factory” is the token factory, or the Agentic MaaS large-scale model service platform. It’s the core of realizing advantages in effectiveness, cost, and stability. Comprising a complete service technology stack—from gateways and routing to inference instances—it is described as “optimized at every layer, with incremental improvements everywhere.” As of July, the daily token calls on Infinigence’s Agentic MaaS platform have grown 40-fold compared to December of the previous year.

The “Front Store” refers to the “AI Productivity Store,” or the Agentic Infra industry solution, which deploys accumulated capabilities across industries. It spans fields such as entertainment, healthcare, legal, and energy, transforming technological potentials into user-friendly, reusable implementations.

Editorial Opinion

In the short term, Infinigence’s case highlights the growing importance of a “common foundational layer” that balances model quality maintenance with cost optimization in China’s AI infrastructure market. The decision by four leading model companies to choose the same infrastructure provider suggests that the standards for quality assurance and operational efficiency in the MaaS market are shifting from mere API provision to end-to-end reliability design. Particularly, the PDD architecture’s achievements in reducing TTFT and costs will likely push competitors to develop similar technological innovations.

In the long term, Infinigence’s ambition to establish a “CATL-like” position could significantly impact China’s overall AI ecosystem. The approach of unifying a fragmented chip ecosystem into a single infrastructure has strategic importance, especially under China’s semiconductor export restrictions. However, the focus on a single infrastructure raises risks of vendor lock-in and potential stifling of competition. The viability of this model in global markets will depend on the regulatory environment and constraints on semiconductor procurement in each region.

References

• “不同模型厂同一家Agentic Infra，AGI时代的地基终于浮出水面”, by 克雷西 — 量子位, 2026-07-20T10:33:23.000Z (ARR)
• Source URL: https://www.qbitai.com/2026/07/455805.html