Reporter |
Editor | Wen Shuqi
In most computing power company stories, the newer and bigger the GPU, the easier it is to win over investors. But Cloudsky tells a completely different tale. It doesn’t manufacture end-side GPUs. Instead, it rounds up scattered GPUs — different models, many just sitting idle — and beams that computing power over high-speed networks straight to phones, PCs, cars, and smart gadgets. The result? Those devices perform as if they’ve got a top-tier GPU tucked inside.
Founded back in 2016, Cloudsky just announced a monster Series E round of over 1 billion yuan (roughly $140 million), led by the China Internet Investment Fund, with CICC Capital and others joining the ride.
In a sit-down with NUPIAO, founder and CEO Mao Xiaodong shared that over the past six years, Cloudsky has been growing revenue at a compound annual rate of 100%. The client roster now covers most of China’s biggest internet platforms and device makers. If there’s one thing keeping him up at night, it’s still supply — demand for computing power just keeps outstripping what they can deliver.
Mao’s resume reads like a who’s who of silicon valley royalty — he did GPU design at Intel, then jumped to Sony where he worked on multiple generations of PlayStation consoles and cloud gaming platforms. Before starting his own thing, he seriously considered doubling down on end-side GPUs. But the PS3 era left him with what he jokes is “GPU PTSD.” Shoving a chip into consumer electronics means wrestling with performance, size, power draw, and cost all at once — and chasing cutting-edge manufacturing processes burns cash like there’s no tomorrow.
By the PS4 generation, he started mulling a radically different approach: instead of fighting to build a better end-side GPU, why not park capable GPUs in servers close to users, and pipe the rendering power over fast networks in real time? If the latency from transmission and compute stays low enough, users can’t tell the difference — it feels exactly like the GPU lives in their device. Think of it as a “virtual end-side GPU” living on the network.
Cloud gaming and cloud PCs actually sparked a wave of hype over a decade ago, all built on pretty much the same idea.
Mao sees gaming as one of the harshest proving grounds out there. Take esports cafes, for instance — they demand 2K resolution and 300Hz refresh rates, where the gap between consecutive frames is a mere 3.3 milliseconds. That means the entire chain — from mouse clicks to video transmission to server-side rendering — has to happen in a blink. Parking a GPU in a server is the easy part. The real magic is making users feel like it’s running right there in their phone or PC. And if a network can handle hardcore gamers duking it out in real time, then it’s ready to take on smart cockpits, robots, and beyond.
Then came the 5G buildout, and opportunity knocked hard. Faster networks and lower latency finally made it practical to push real-time computing power to end devices over the wire. Cloudsky rode that wave straight into the supply chains of telecom operators and internet giants — Cloudsky brings the real-time rendering and transmission tech, operators bring the network, data centers, and customer channels, and together they roll out edge nodes.
Riding the operators’ 5G expansion, Cloudsky scaled up its node footprint and cracked the code on monetization. Cloud gaming and cloud conferencing quickly became poster children for 5G commercialization. By the company’s own numbers, it’s now running hundreds of computing nodes across China, with network coverage in more than 300 cities worldwide, over 300,000 GPUs under management, and more than 600 million end users reached.
In theory, the denser the nodes and the closer they sit to users, the lower the latency. Mao breaks it down: a node within a 50-kilometer radius can hold latency under 10 milliseconds. Shrink that coverage down to 5 kilometers, and you’re looking at around 1 millisecond.
This is a business with serious scale economics. Cloudsky has dubbed its strategy “Real-Time Intelligent Computing Mesh,” with a long-term ambition to blanket the map with computing nodes just like telecom base stations. More nodes, closer to users, means the end-side experience edges ever closer to local computing. And as customers and use cases stack up, the same network gets utilized more efficiently — which only thickens the company’s moat.
So why are operators and internet companies willing to pay up? Simple: computing resources are scarce and pricey. Here’s the backdrop — China’s market is starved for high-end GPUs, yet it’s also swimming in mid-to-low-end domestic chips and idle capacity. At the same time, the country boasts one of the world’s most developed communications networks. Figuring out how to use that network to breathe life into scattered, underused computing power and turn it into on-demand capability for end devices? That’s a business with real, urgent demand.
Mao explains that Cloudsky’s in-house StackGPU technology uses software orchestration to rope GPUs of every brand, model, and performance tier into a single shared resource pool, then slices and dices compute power to fit different devices and applications. He describes it as a “giant virtual GPU” that lives outside the device. Most everyday inference tasks get handled at the edge, while heavier lifting escalates to the central cloud — slashing overall costs and reducing dependence on any single high-end GPU.
On top of that, internet platforms can offload rendering tasks to the cloud, letting low-spec devices run heavyweight content. Device makers can tap remote computing to fill in local performance gaps and trim hardware and AI inference costs. But this is way more than just plugging GPUs into a pool. Every GPU needs to be understood, tuned, and adapted for specific workloads — that takes years of grind. Building the network, stacking heterogeneous chips, and adapting to one use case at a time — that’s the technical foundation Cloudsky has spent the last decade laying down.
Sure, the big cloud players and the three major telecom operators are all dabbling in edge computing. But Mao argues the cloud giants are more comfortable building massive centralized nodes, while Cloudsky is all about node density and scenario fit — the long game being to spread computing points as densely as cell towers. Smaller players lack the nationwide reach, while the giants balk at the cost of distributed deployment and case-by-case adaptation for niche scenarios. That slog — long, expensive, and thankless — is exactly the pocket Cloudsky has carved out for itself.
Beyond the big tech crowd, Cloudsky’s fresh growth engines are shifting from AI applications to AI hardware. The generative AI boom isn’t just feeding the appetite of internet giants — tons of small and mid-sized businesses are hungry for computing power to power content generation, short videos, and AI-powered staff that cut costs. Meanwhile, large language models are making their way into phones, wearables, toys, home appliances, and cars. But these gadgets can’t fit a power-hungry GPU, and they can’t afford the bill for endlessly pinging a central cloud model.
Take an AI phone booking a train ticket through an intelligent agent. If the token cost of that single request costs more than the ticket itself, the whole feature is pointless. That’s exactly where low-cost, low-latency edge computing steps in to save the day.
Mao lets slip that Cloudsky already covers over a million automotive-related devices, and AI toys could top ten million units.
Cloudsky is also pushing hard into overseas markets. Over the past year, it’s dipped its toes into the US, Europe, South America, and the Middle East. Each market has its own quirks — the Middle East expansion has been complicated by regional conflicts, while Europe and the US have relatively abundant computing supply, so what customers there really want is convenience and supplementary capacity. But the bottom line is this: the hunger for low-cost computing power is universal, and that’s Cloudsky’s golden ticket.
In a way, Cloudsky is betting on a new division of labor taking shape: chip companies supply the compute units, telecom operators provide the network, model companies deliver the underlying intelligence, and Cloudsky ties it all together through a real-time intelligent computing network — ferrying GPU muscle and AI capabilities to the devices that need them, and closing the “last mile” before AI actually lands in people’s hands.
Mao’s vision for the next decade? Any device, the moment it connects to the network, becomes a powerful intelligent terminal. Getting there means relentlessly expanding node coverage, iterating on technology, and adapting to more and more scenarios. The bigger the network gets, the more computing, customers, and use cases can be recycled and shared. In his view, AI computing power just might be humanity’s last great infrastructure buildout.
