Author | Wang Han, Editor | Mo Ying
In 2026, the popularity of Agents is unprecedented. On September 21, at the AICC 2026 conference, IDC released a set of figures: in 2026, there were approximately 79.4 million active Agents globally, which will reach 2.216 billion by 2030. The growth in Token consumption during the same period is even more astonishing: from 0.026 Peta (quadrillion) in 2026 to 152,667 Peta in 2030. While the number of Agents has grown by about 28 times, Token consumption has surged by millions of times.
With such rapid growth on the demand side, can the computing power supply keep up? Two cases this year are highly typical: the first is Kimi K3. This model, with a total of 2.8 trillion parameters, has weight files reaching 1.56TB. The GPU memory of ordinary AI servers cannot even hold the weights, and the official recommendation is at least 64 GPUs or even larger scales to run it. It points to the "upward" dilemma: as Agents require stronger reasoning, longer context, and more multi-agent collaboration, the higher the capability ceiling, the more the deployment threshold shifts from a single machine to a cluster. The second is DeepSeek.
In the first half of 2026, it cut the API price of V4-Pro by 75%, stating the adjustment would be long-term; however, in August, it announced an overall price increase, directly citing "insufficient computing power." It points to the "broad" dilemma: when Agents massively invoke cheap models, the demand for cost-effective computing power is rapidly amplified. If the capacity is insufficient, low-price supply becomes unsustainable. How to solve these two problems? At the AICC 2026 conference, Inspur provided its answer: upward, Capability AI Compute; broad, Capacity AI Compute. At the conference, Inspur officially released the YuanNao SD200 Ultra Super Node AI Server and YuanNao HC2000 Diversified Computing Power Unit. Advancing in both "upward and broad" directions, through system synergy, Inspur is transitioning from providing computing power to producing intelligence.
01.Upward Super Node Breaks Through Intelligence Ceiling
To understand where the upward bottleneck comes from, we must first see what changes have occurred on the demand side of AI. In the past, people's demand for AI was to answer questions, and computing power invocation was single and temporary. In the Agent era, behind a single user intent, there may be dozens, hundreds, or even more inferences, involving the collaborative work of multiple Agents for several hours or even days.
The change in application forms has directly raised the requirements for computing power systems. At the same time, the scaling up of frontier models has also entered an acceleration phase during this period. Taking Kimi K3 as an example, its parameter count has reached 2.8 trillion, and the weight files are approximately 1.56TB. The memory of ordinary servers cannot even hold the weights, let alone a million-Token context. The official recommendation is to scale up to 64 GPUs or even larger to make the model truly run. Not to mention more frontier models that may scale to the ten-trillion parameter level in the future.
Nowadays, efficient communication between chips is required, memory needs unified addressing, and data movement needs to be minimized as much as possible. Simply stacking more servers together no longer works. The significance of the super node lies in breaking through the computing boundaries of traditional single machines, incorporating more chips and memory into a highly collaborative computing domain, and providing a larger "effective computing space" for frontier intelligence. The YuanNao SD200 Ultra Super Node AI Server released by Inspur is designed precisely around this goal. The YuanNao SD200 Ultra uses an open system design to tightly couple 128 domestic AI chips, achieving unified memory addressing and ultra-low latency communication, and utilizes innovative technologies such as symmetric direct connection and operator optimization to reduce data movement and waiting during model execution. Based on this system, the SD200 Ultra can run the 2.8-trillion-parameter Kimi K3 on a single machine, with Token generation latency breaking through 5.85 milliseconds for the first time. At the same time, as the model parameter scale continues to increase, it also supports frontier models with a scale of 10 trillion parameters, providing support for the continuous expansion of model scale and context length.
From a technical implementation perspective, the SD200 Ultra is based on 128 domestic AI chips, adopts the 3D HyperMesh architecture, and provides 8TB bandwidth and 64TB memory, enabling frontier models to fully utilize the computing and storage resources of multiple chips. In terms of communication, native memory semantic communication reduces cross-chip communication latency to 0.69 microseconds; in terms of inference optimization, Inspur has built a super operator library system, fusing numerous basic operators in KDA, Gated MLA, and MoE into large operators combining computation and communication. The number of operators is reduced to 1/11 of the original, significantly reducing the number of Kernel Launches and HBM access overhead, while hiding communication latency and improving the overall execution efficiency of the GPU.
Today, AI is already capable of solving the Navier-Stokes equations, which have plagued the academic community for decades. What Capability AI computing needs to do is to ensure these frontier capabilities are not just demonstrations in the laboratory, but practical and reproducible productivity.
02.Broad Making Intelligence Supply "Sufficient and Affordable"
If Capability solves "how strong intelligence can be," then Capacity solves "how much intelligence we can actually have." As the number and depth of Agent usage increase, Token demand will continue to grow rapidly. Improving the Token output per unit of computing resource has become the key to scaled supply. However, increasing Token output is not easy. Inference is not a single computing mode.
Prefill and Decode have different computing characteristics, and Attention and MoE have different requirements for computing and memory access. The development of model architectures such as multimodal and MTP will bring more types of computing loads. Simply stacking more servers will not only fail to bring linear capability growth but may also lead to long-term idleness of some computing power and continuous tension in other resources due to resource waste, imbalance, and task queuing. This means that Capacity AI computing needs a different approach: letting different computing powers handle the loads best suited for them, and then organizing them through a unified software stack. Inspur's answer is the YuanNao HC2000 Diversified Computing Power Unit. Aimed at continuous, large-scale inference demands, based on the DirectCom 2.0 ultra-fast architecture, and using native liquid-cooled high-density whole racks as the carrier, it organizes computing resources according to loads to achieve diversified computing power synergy.
Its core software stack, MCIS, establishes a synergy mechanism for diversified computing power: continuously measuring the capabilities and loads of different computing resources, dynamically allocating computing power according to task requirements, and adjusting resource configurations in real-time during operation, allowing different types and scales of computing power to undertake more suitable inference tasks.
At the hardware level, the HC2000 adopts a full liquid-cooling design and high-current power supply, with a single-rack power supply capacity reaching over 300 kilowatts, a maximum capacity of 256 AI cards, and a 4-fold leap in computing power density.
More importantly, it achieves precise matching between computing power and computing characteristics. Through different combinations, diversified chips perform their respective duties, and are then organized through the software stack to jointly complete inference synergy, improving the Token output efficiency of the entire computing power system. Based on domestic computing power, through MCIS and adaptation with a certain model, in typical Agent task scenarios and under the same SLO constraints, it achieves a 10-fold increase in Token production capacity with the same investment compared to inference systems built with a single computing power.
Liu Jun, Chief AI Strategy Officer of Inspur, once said: "The value of intelligence depends not only on how high the peak is, but more on how many people can access it." The logic of "broad" is to rapidly bring advanced inference capabilities to a wider range of users through efficiency improvements and cost reductions, transforming intelligence from a scarce resource into a fundamental capability accessible to the entire society.
03.From Capability to Capacity Why is System Synergy Needed?
The SD200 Ultra and HC2000, one going upward and the other going broad. They seem to be doing different things, but at the bottom level, they share the same judgment: when intelligence begins to be "produced," the competition in the computing power industry is no longer single-point competition, but system competition. Liu Jun summarizes this approach as "penetrating innovation".
The first level is the penetration from chips and server systems to models. In the past, innovation meant each layer of the industrial chain doing its own job—chips, systems, frameworks, algorithms, operators—and then aligning vertically; but in the Agent AI era, customized or specialized optimization is required for specific models.
The second level is the penetration from a single machine to a system. Capacity requires end-to-end innovation, taking out every link in the chain to look at, and matching the most suitable computing resources according to the characteristics of different computing loads. Prefill can use high-computing-power chips paired with LPDDR particles that have better supply and cost; for the Decode stage, AI chips with large-capacity HBM that balance memory capacity and bandwidth can be selected. On this basis, in the FFN computing link, 3D DRAM stacked chips can be selected to further shorten the data movement path. Abandoning the strategy of using the same general-purpose behemoth chip to do everything, and instead building diversified chips and diversified computing power systems.
The third level is the penetration from technology to ecology. Previously, understanding "diversified computing power" meant server manufacturers supporting AI chips from different companies; but today, a new synergy method has emerged: multiple chips in the same business, with Prefill using Company A's chips and Decode using Company B's chips, forming a computing power system. Industrial collaboration has shifted from a zero-sum game of "you win, I lose" to open win-win. From a longer-term perspective, Capability and Capacity are not two parallel lines, but a cycle.
Every time a new intelligence capability emerges, it is often expensive and scarce at first. With algorithmic progress, system optimization, and engineering maturity, these capabilities will gradually enter more models, applications, and users at a lower cost. Capability is responsible for creating new possibilities, and Capacity is responsible for making these new possibilities accessible to more people. At the same time, larger-scale intelligence popularization will bring new applications, new data, and new demands, driving the next round of Capability breakthroughs. System synergy is precisely the underlying mechanism that keeps this cycle running continuously.
04.Conclusion: In the Agent Era the Computing Industry Moves from Providing Computing Power to Producing Intelligence
Entering the Agent era, Token factories cannot just ask how many GPUs and CPUs they have; they should also ask: how strong intelligence can these computing powers support? How much intelligence can be produced at how low a cost? Inspur uses the SD200 Ultra to break through the intelligence ceiling upward, and the HC2000 to promote scaled intelligence supply broadly. One is responsible for making AI more powerful, and the other for making AI more popular. Connecting the two is system-level synergistic innovation and an open ecosystem. Today's Capability is precisely the capability that tomorrow's Capacity needs to supply at scale. In the Agent era, the mission of the computing industry has therefore become broader than ever before.