The Next Big Opportunity in Physical AI: How Jiuzhang Cloud Computing and Other Intelligent Computing Cloud Service Providers Are Meeting Robots' Computing Power Needs
The computing power demand for physical AI is shifting from self-built hardware to on-demand intelligent computing cloud services. Jiuzhang Cloud Computing is meeting robot companies' elastic training needs through heterogeneous chip adaptation, computing power pooling, and unified scheduling. Wang Xingxing of Unitree Technology advocates moving as many computable tasks to the cloud as possible, while AgiBot's training cluster will approach the scale of ten thousand GPUs this year and will be obtained mainly through leasing. In the first half of 2026, Jiuzhang Cloud Computing's cloud business revenue grew by about 420% year on year, and its intelligent computing cloud has already been scheduling thousands of virtual robot policy samples in parallel in world model simulation.
Body · AI translation
Over the past year, humanoid robots and quadruped robot dogs have appeared with increasing frequency in factories, warehouses, and stores. The questions the industry cares about are changing: from "what actions can a robot perform" to "how can we make it more and more accurate."
The answer lies not only in the robot itself, but also in the computing system behind it: where the data comes from, where the simulation runs, how often the model is iterated, and how computing power matches the R&D cadence. None of this can be solved by simply buying a few more servers.
NVIDIA has proposed the concept of "three computers": one trains the model, one trains and validates skills in a digital twin and simulation environment, and the third is deployed on the robot itself for execution. Platforms such as Omniverse, Isaac, and Cosmos are intended to connect training, simulation, and edge computing into a single chain. Its latest quarterly earnings show that data center business revenue grew 117% year over year, and physical AI is becoming a new growth market for computing platforms.
But simulation cannot fully replace the real world. David Reger, founder and CEO of Neura Robotics, said in an interview with McKinsey that there is still a gap between digital twins and the real environment, and robots need to be trained in real physical environments and collect multimodal data such as motion, touch, and sound. This feedback then feeds into the next round of training and simulation.
As a result, physical AI R&D has formed a continuously iterating closed loop: data collection, simulation, model training, reinforcement learning, inference, and real-machine validation. Each stage has different computing power requirements, and these change dynamically with R&D progress, making fixed-configuration computing power difficult to adapt.
The change in R&D models is reshaping how computing power is supplied and served. Compared with one-time construction and long-term ownership, cloud computing power that can be called on demand and elastically scaled has begun to become a new choice for robotics companies.
Chinese robotics companies feel this more directly.
Wang Xingxing, founder of Unitree Robotics, said at the China International Industry Fair in October this year that the robot itself is limited by space, battery, and heat dissipation, so the computing power that can be deployed is limited, and computing tasks suitable for the cloud should be moved to the cloud as much as possible. Cloud computing power can be allocated in real time and is also more flexible to manage. He also emphasized that real-machine data is an indispensable closed loop for model iteration. The smarter the robot, the more it depends on the cloud computing power behind it.
Peng Zhihui, co-founder of AgiBot, revealed in an interview that the company's training cluster this year will approach the scale of ten thousand GPUs to cope with growing computing demand. Co-founder Yao Maoqing added that computing power is mainly obtained through leasing to control asset investment and retain scaling flexibility.
The computing power ledger of leading robotics companies is shifting from "how many GPUs to buy" to "whether usable computing power can be called up at any time."
The reason is that embodied intelligence R&D has obvious phases: model training, simulation, and reinforcement learning push up computing demand, while demand may fall back during the validation stage. Self-owned hardware has high costs and is prone to idling, while temporary procurement makes it difficult to ensure R&D continuity.
The capital market has already priced this demand. According to industry estimates, CoreWeave's revenue in the first half of 2026 grew 112% year over year, and its full-year capacity has been sold out in advance; domestic AI infrastructure and intelligent computing cloud service provider Jiuzhang Cloud's cloud business revenue in the first half of 2026 grew about 420% year over year. The growth of AI neoclouds reflects the market's continuously rising willingness to pay for "callable at any time" computing resources.
However, there is another way to calculate this account: the unit price of on-demand rental may be higher than self-building, and under long-term high-load operation, the total cost may not necessarily be lower; there is also no industry consensus on whether the computing demand of physical AI can grow as rapidly as that of large models. At present, the training scale of robotics companies is still significantly smaller than that of general large-model companies. But the demand for multi-step computing is already clear, and the growth pace remains to be verified.
Shang Mingdong's view points to another dimension: whether computing resources are actually put to use. Shang Mingdong, co-founder and COO of Jiuzhang Cloud, previously said in an interview with CCTV Finance that when measuring the computing industry, one should not only look at the scale of construction, but also at whether resources are truly used and whether they can serve more industrial entities. Computing power is transforming from a scarce technical resource into a general basic means of production.
Jiuzhang Cloud's approach is to undertake distributed computing demand in layers rather than relying on a single centralized solution. At the bottom layer, through heterogeneous chip adaptation, computing power pooling, technical solutions to prevent computing power fragmentation, and unified scheduling, computing resources of different architectures run collaboratively, providing a centralized large computing power foundation for robot foundation model training tasks; the middle layer relies on cloud-native cluster management, high-performance storage, and a reinforcement learning runtime environment to support automated orchestration of multi-round training and validation tasks; the upper layer connects model training and inference consistency capabilities, extends to the edge inference layer, and dynamically allocates resources according to task requirements. What is delivered is no longer isolated computing resources, but computing services that can be scheduled, scaled, and delivered through distributed collaboration.
This is also where Jiuzhang Cloud's differentiation lies. Domestic robotics companies face a more complex computing environment, with differences among vendors' chips in architecture, drivers, and software stacks, and different resource requirements for training, simulation, and inference. Cross-architecture adaptation and unified scheduling capabilities directly affect resource utilization and overall computing costs. This work is not conspicuous, but it is the key to intelligent computing cloud services truly forming competitiveness.
The transaction method is also changing accordingly. In the past, computing power procurement mainly revolved around GPU type, quantity, and unit price; now, companies care more about whether they can call computing power by task, measure it by actual usage, scale up during R&D peaks, and release resources during the validation stage. Computing power is shifting from equipment procurement to service delivery.
This capability has entered the physical AI R&D process. In joint validation with a leading world model company, Jiuzhang Intelligent Computing Cloud supported the parallel scheduling of thousands of virtual robot policy samples in an ecosystem partner's simulation environment, accelerating reinforcement learning and safety testing; in joint validation in an automotive assembly scenario, elastic multi-stage computing power scheduling supported improved policy iteration efficiency. Facing computing demand that changes with simulation scale, model complexity, and iteration rounds, elastic scheduling and multi-stage computing power invocation help reduce idle resources and improve R&D efficiency.
Third-party research also confirms this trend. Citi Research pointed out in "Embodied Intelligence: The Rise of Physical AI" that digital twins, real-world data collection, simulation, and the computing software ecosystem that coordinates hardware are all important pillars of physical AI development. Computing infrastructure has therefore become an important growth opportunity in the industrial chain.
Taken together, the direction is clear: NVIDIA connects training, simulation, and edge computing; CoreWeave integrates cloud platforms into robot R&D; and Jiuzhang Cloud uses heterogeneous computing power scheduling and an application engineering system to integrate computing resources and provide elastic computing services. The paths differ, but the goal is the same: to make computing infrastructure adapt to the continuously changing R&D needs of physical AI.
The heat of the physical AI industry is extending from the robot itself to the computing system that supports its continuous evolution. As R&D tasks become increasingly complex, computing infrastructure must not only provide computing power, but also efficiently organize and schedule it, turning it into sustained R&D capability. The robot itself determines the capability boundary of physical AI, while computing infrastructure affects its iteration efficiency. This may be the next trend worth watching in the physical AI boom, and Jiuzhang Cloud has already entered the field ahead of time.