How Yuntian Changxiang Is Building a National GPU Network for 600 Million Devices

Turning Scattered GPUs Into a Virtual Supercomputer

Yuntian Changxiang has spent a decade executing a counterintuitive hardware strategy: instead of building more powerful end-user GPUs, it stitches together hundreds of thousands of existing chips – of different brands, ages and performance tiers – and delivers their combined power to everyday devices over high-speed networks. The result, according to co-founder and CEO Mao Xiaodong, is a virtual GPU that enables phones, laptops, cars and smart hardware to run demanding games, AI inference and real-time rendering as if a top-tier graphics processor were sitting right inside them.

In an exclusive interview with Jiemian News, Mao disclosed that the company's annual revenue has compounded at 100% for six consecutive years, covering most of China's leading internet platforms and device brands. The startup now operates several hundred edge computing nodes across more than 300 cities globally, manages over 300,000 GPUs and serves more than 600 million end-user devices. Demand still far outstrips supply, he said, and the firm's main headache is scrambling to add capacity fast enough.

Mao's pivot dates back to his time at Intel and Sony, where he worked on GPU design and the PlayStation family. A gruelling R&D cycle for the PS3 gave him what he calls “GPU-phobia”: cramming a high-performance chip into a consumer device forces painful trade-offs between size, thermals, cost and manufacturing complexity. By the PS4 era he was already sketching an alternative – place the GPUs in operators' edge data centres near users, racing data back and forth over 5G links so that the total delay remained imperceptible. Gaming, with its demand for 300 Hz refresh and sub-4 ms frame intervals, became the acid test. Once that hurdle was cleared, the same system could move into smart cockpits, robotics and AI wearables.

China's 5G rollout provided the missing piece. Carriers supplied fibre, indoor radio sites and subscriber channels, while Yuntian Changxiang supplied the real-time rendering, scheduling software and scene-by-scene tuning. The joint deployments turned cloud gaming and virtual meetings into flagship 5G use cases and gave the startup a factory for scale – more nodes, more reuse of the same GPU fleets across different customer workloads – which Mao describes as a “real-time intelligent computing fabric” aimed at eventually matching the density of telecom base stations.

Where Yuntian Changxiang's Real-Time Computing Fabric Fits into the Industry

StackGPU: the software that pools anything with a chip

The company's secret sauce is a proprietary layer called StackGPU, which inventories GPUs from different vendors – Nvidia, AMD, domestic Chinese alternatives, even older models – and dynamically slices and combines them for each application. Most routine AI inference stays at the edge, while complex jobs burst to central cloud. This diversity reduces dependence on any single high-end processor, a critical advantage in a market where US export controls limit access to advanced silicon. Mao likens the whole pool to one giant virtual GPU that exists outside the device.

Why carriers and internet giants keep paying

For platforms and device makers, the value is straightforward: offload rendering or AI workload to the edge and let low-cost, low-power hardware behave like a premium machine. A smartphone can run a console-grade game without a flagship chip, an electric car can display a real-time 3D cockpit with minimal on-board compute, and an AI-powered toy can understand voice commands without a dedicated NPU. Telcos, meanwhile, monetise their 5G infrastructure and gain a sticky enterprise product – a service that helped Yuntian Changxiang embed itself deep in the supply chains of China's big three carriers and the major cloud providers.

Density vs. mega-clouds: a deliberate trade-off

Headline cloud players (Alibaba, Huawei, Tencent) are also pushing into edge computing, but Mao argues they prefer massive centralised hubs. Yuntian Changxiang's bet is that node density and painstaking per-scene optimisation create a moat that is uneconomic for hyperscalers to replicate. Small competitors lack the nationwide fabric; large ones are unwilling to bear the granular deployment and tuning costs. The startup occupies a niche that required ten years of network-building, heterogeneous scheduling, and customer-specific adaptation – a long, costly ramp that now acts as a barrier.

The global picture: four markets, one shortage of affordable compute

Over the past year, Yuntian Changxiang has dipped into the US, Europe, the Middle East and South America. The Middle East push has been hampered by regional conflict; in mature Western markets supply is more abundant, so customers value convenience and burst capacity over raw chip scarcity. Yet Mao insists the core need – cheap, accessible GPU cycles – is universal, and he is positioning the firm as the “last mile” of AI infrastructure: the layer that connects silicon, network and model into an always-on intelligent terminal. That vision, however, rests on relentless expansion and the assumption that demand for commodity-grade AI compute will hold up even if chip sanctions ease.

What the Rise of Distributed GPU Networks Means for Infrastructure Players

For telecom operators and cloud providers: Yuntian Changxiang's model shows that edge-node density, not just raw capacity, is becoming the differentiator in real-time AI services. Evaluate whether your own 5G edge rollouts can be accelerated by partnering with specialist schedulers, particularly for gaming, automotive and IoT workloads. Margins on these services may improve faster than on generic cloud storage.

For device manufacturers (phones, cars, wearables): The ability to offload inference and rendering to a nearby compute pool can dramatically lower bill-of-materials costs. Test workloads with StackGPU-style orchestration to see if a cheaper local chipset paired with edge compute delivers the same user experience. The key metric is frame-to-photon latency – anything under 8 ms opens the door to cloud-native cockpits and AI toys.

For investors and boards: The startup's 100% revenue CAGR and coverage of 600 million end devices put it in a rare class of infrastructure plays. Due diligence should focus on the sustainability of carrier contracts, the fungibility of its GPU fleet if sanctions tighten further, and the unit economics of its overseas pilots – especially whether it can replicate China-scale margins in markets where local silicon is already abundant.