Since the "crayfish" OpenClaw became a viral sensation at the beginning of the year, the rise of Agent has been unstoppable.
From phased training and inference of a single model to large-scale model training, multi-model collaborative inference, and continuous operation of massive agents, AI tasks are entering a new phase of "always online".
At the 2026 World Artificial Intelligence Conference (WAIC 2026), Zheng Weimin, an academician of the Chinese Academy of Engineering and professor in the Department of Computer Science and Technology at Tsinghua University, proposed that the token consumption of a single intelligent agent can reach that of traditional dialogue applications.Hundreds to thousands of timesWith massive numbers of agents concurrently calling models, reading and writing memories, orchestrating tasks, and executing tools, the rate at which tokens are consumed is increasing rapidly.exponentialSoaring.
The surge in demand for tokens has directly raised the ceiling for the scale of AI infrastructure. Thousand-card clusters are no longer sufficient, and tens of thousands or even hundreds of thousands of cards are becoming the standard for training cutting-edge large models.
However, data centers have limited power and space. To meet the unlimited demand for tokens, the core direction is...Packing more computing power into a unit of spaceThis pushes computing power density to its limit.
As computing density increases, the power consumption per rack will inevitably skyrocket, jumping from tens of kilowatts to hundreds of kilowatts, and further increasing...MW classMoving forward.
But then a problem arises. When the power of a single cabinet reaches this level, traditional cold plate liquid cooling solutions will cause the cabinet heat dissipation to be insufficient, space to be insufficient, and interconnection to be interfered with. All three bottlenecks will narrow down in the limited space at the same time.
If the Token Factory wants to continue expanding production, computing, heat dissipation, interconnection, and power supply can no longer operate independently; they must be integrated from...System bottom layerA new solution has been found.
Native liquid cooling, thisCoupling computing, heat dissipation, interconnection, and power supply at the system level.The proposed solution began to emerge in the industry.
01.
Token demand explodes
This has brought out the triple dilemma of traditional liquid cooling.
The surge in demand for tokens has driven up the power consumption per rack, pushing deployment space to its limits. In this scenario, the shortcomings of traditional liquid cooling have been dramatically amplified. Solutions that were previously viable at low power densities are now...Three key dimensionsAt the same time, a red light came on.
The first challenge is that computing power density is limited by the thermal management of the entire system.
High computing density means integrating more chips and even entire system components such as HBM, memory, network cards, and optical modules within a unit space, which also means that more heat needs to be dissipated. Under high-density deployment, the traditional cold plate liquid cooling method of "adding plates one by one" can never keep up with the growth rate of the entire system's heat generation.
If the heat dissipation capacity cannot keep up, the chip will trigger its temperature limit and reduce its frequency, causing token production to continuously decline from its peak. Insufficient heat dissipation will slow down token production.
The second challenge is that the heat dissipation structure continues to encroach on computing space.
Currently, cold plate liquid cooling dominates the liquid-cooled data center market.93.5%While this accounts for a certain market share, the underlying logic of most solutions still involves layering fans, cold plates, hoses, and distributors. This stacking of heat dissipation structures continuously encroaches on the space that should belong to the computing unit, further compressing the upper limit of GPU capacity per cabinet.
As rack power approaches the MW level, this space conflict is amplified dramatically: the more complex the heat dissipation, the more space it occupies, and the fewer computing units can be placed, thus locking in the maximum capacity of a single rack's tokens.
The third dilemma is that the conflict between heat dissipation and interconnection is amplified at the MW level.
As the computing domain expands, more computing nodes and switching nodes are needed to form a unified system with high bandwidth and low latency. However, in traditional architectures, backplanes, cables, air ducts, cold plates, and fluid channels are planned separately, which lengthens the interconnection paths, increases the number of interfaces, and raises communication latency.
When dozens or even hundreds of GPUs need to frequently synchronize data, even minute latency can be amplified exponentially, systematically reducing the overall efficiency of multi-GPU parallel inference. Decreased interconnect efficiency also slows down token generation.
The triple dilemma points to the same conclusion:The problem with the Token Factory goes beyond "heat dissipation" itself; it is a systemic dilemma in which computing, interconnection, power supply, and heat dissipation are mutually constrained within a limited space.
To overcome these three dilemmas simultaneously, we must rethink the system design from its very beginning—planning computing, heat dissipation, interconnection, and power supply as a unified whole.
02.
Native liquid cooling:
Responding to the triple dilemma from the source of the system
So what is the right approach to "rethinking from the source"? Making the cold plate bigger, replacing the flexible tubing with thicker ones… these conventional solutions only address the symptoms. To truly address the root cause, we need a different approach—Native liquid cooling.
Traditional liquid cooling follows a "determine the computing system first, then configure the cooling system" approach, meaning that space is only found to fit the liquid cooling equipment after the computing system is finalized. Native liquid cooling is not simply a "Pro Max" version of plate-type liquid cooling; the fundamental differences lie in three dimensions:
existTime dimensionAbove, thermal design has shifted from rear-mounted to front-mounted, no longer a matter of compromise;Spatial dimensionAbove, heat dissipation has shifted from fragmented coverage to comprehensive management, no longer operating in isolation;Logical DimensionIn this context, the focus of optimization has shifted from individual heat dissipation components to the entire computing system, with computing, heat dissipation, power supply, and interconnection no longer operating in isolation.
This system reconstruction unfolds sequentially along the lines of computing power density, spatial density, and interconnection density, ultimately all pointing to...Token production capacity.
First lookcomputing power densityNative liquid cooling usesThermal architecture and computing architecture work togetherThe overall system thermal design, throughIntegrated cold plate and full-area liquid cooling reconstructionThis enables coordinated heat dissipation across the entire computing system.
It breaks away from the conventional thinking of modular local optimization in traditional cold plate liquid cooling architecture, and...GPU, CPU, memory, network card, optical module, and SSD are all integrated into a unified thermal management system.The overall rack computing density can be increased to that of traditional solutions.More than 10 times.
With improved heat dissipation, the chip does not drop in frequency under continuous high load, and token output remains stable at its peak without decaying over time.
Let's look again.spatial densityNative liquid cooling, through system architecture restructuring, fully frees up computing deployment space.Adopting integrated cold plate and minimalist engineering design,accomplishZero hoses, zero cables, zero fansReduced assembly interfaces80%Reduced flow resistance60%,release50%Device space is used for core units such as chips, memory, and interconnects.
With the same rack space, the number of computing units that can be accommodated has increased dramatically. Heat dissipation no longer takes up space, and the maximum token production capacity of a single rack has been fully unlocked.
Interconnection densityThe breakthrough is equally important. Native liquid cooling utilizes...Orthogonal direct connection architecture without backplaneThis decouples computing nodes from switching nodes and constructs a high-density, low-latency direct-connection network matrix using forward and backward shortest paths.
Based on a single GPU and Switch 4.8TB/s high-speed direct connection design, the 64-card full interconnect architecture can provide 300TB/s interconnect bandwidth and support 224G high-speed data transmission. Shorter interconnect paths mean lower communication latency and higher multi-card collaboration efficiency, allowing more computing resources to work together efficiently, and systematically increasing the overall token throughput of the cluster.
03.
The Agent era requires two types of computing power foundations.
Inspur Information uses native liquid cooling to simultaneously cover
From a triple dilemma to a breakthrough in three dimensions, native liquid cooling has proven the value of system reconstruction.
But once the architectural logic is established, the next more practical question is:How will these capabilities ultimately be translated into products? What kind of hardware support will the computing power foundation of the Agent era require?
To answer this question, we must first understand how an Agent works. From receiving instructions to completing a task, an Agent goes through multiple steps, including tool invocation, memory read/write, task orchestration, and parallel scheduling. These "orchestration and scheduling" tasks primarily run on the CPU.
In other words, the complete workflow of an Agent is a "two-pronged approach": the GPU is responsible for high-speed token generation, and the CPU is responsible for efficient Agent scheduling. Both types of computing power are indispensable; if either one fails, the response speed of the entire Agent system will be dragged down.
IT infrastructure products, solutions and service providersInspur InformationBased on the native liquid cooling architecture, corresponding high-density rack products have been launched to address these two types of needs.
Let's look at the GPU side first, from Inspur Information.GPU-native liquid-cooled rack serverIt can support kilocalories of computing power within a 48U space—achieving a single-cabinet computing power density comparable to traditional solutions.More than 10 times.
It innovatively adoptsIntegrated SDAC cold plateThe GPU, CPU, memory, network card, optical module, and other heat sources are deployed in a high-density planar configuration on both sides of a single cold plate.Double the heat dissipation efficiencyRegarding interconnection,Pioneering MDP matrix direct-connect architecture without a backplaneThe 1024 GPUs are connected to the Switch via high-speed direct connections, forming a high-density fully interconnect matrix.Supports lossless expansion of 1,000 and 10,000 calories..
The final results were: single-cabinet power coverage330kW to 1MWLarge model throughput reaches1.46 million tokens/sA single MW-level rack can support a high-yield token production line, continuously outputting at full speed without frequency reduction or speed drop.
But GPUs alone are not enough. Inspur Information simultaneously launched...CPU native liquid-cooled rack serverBased on liquid-cooled OCM architecture, single-cabinet integration384CPU, supportsMore than 40,000The Agent runs with high concurrency, ensuring no frequency drop or speed reduction under high load, and keeping the Agent's scheduling response at a low latency.
GPU racks are responsible for high-speed token production, while CPU racks are responsible for efficient agent scheduling—the two types of computing power bases work together under a unified native liquid-cooled architecture.
From token generation to agent execution, the entire process is now free of heat dissipation bottlenecks, and there is no longer any situation where a slowdown at one end causes overall queuing. The native liquid cooling system systematically supports the computing power production line, ensuring that tokens remain highly efficient from generation to execution.
04.
Conclusion: Hardware convergence
System architecture capability is the ceiling for token productivity.
Currently, the hardware gap in the AI infrastructure sector is narrowing. The core components that manufacturers can procure are almost identical, and the performance ceiling of the chips themselves is becoming increasingly transparent. If you have a GPU, others also have it; if you use HBM, others can also buy it. Simply piling on hardware is no longer enough to differentiate yourself.
As GPUs, HBMs, and network chips gradually converge, and power and space become hard constraints on AI infrastructure, the real watershed lies in who can organize these identical components into a more efficient computing token production line.
With the same computing power at the kilocalorie scale, the final token output can differ by 30% or more depending on the rack system architecture. The difference lies not in the chips themselves, but in how the chips communicate, how heat is dissipated, how power is allocated, how space is utilized, and so on. These system-level design choices ultimately determine how much of the performance of each chip can be truly unleashed.
This is where the value of system architecture capabilities lies.
It defines the physical layer of the upper limit of computing power conversion: What kind of heat dissipation solution can allow the chip to run at full frequency continuously? What kind of interconnect topology can allow multi-card collaboration without creating communication bottlenecks? What kind of space layout can fit the most computing units into a limited rack?
There are no standard answers to these questions; what is tested is the ability to design an architecture that balances and coordinates computing, heat dissipation, interconnection, and power supply.
Native liquid cooling is the physical manifestation of this system-level innovation.