The first thing many AI startups worry about is the cost of GPUs. The second is electricity. For a founder trying to build an AI product in Karachi, getting enough computing power can feel just as difficult as finding an affordable and reliable power supply.
.jpg "Could Computing Power Be Shared Like Money?")
But China is facing a very different challenge.
The country has invested heavily in AI infrastructure and has access to a large pool of computing hardware. The problem is not always a shortage of chips. Instead, it is how to make better use of the computing power that already exists.
Imagine a company has hundreds of powerful AI servers sitting mostly idle at certain times of the day. At the same moment, another company may urgently need those same resources to train a model or run an AI service. Moving physical machines from one data centre to another is obviously impractical.
That is where the idea of a “compute bank” comes in.
Instead of transferring hardware, companies could share access to unused computing capacity through a common digital platform. A data centre with spare resources could make that capacity available, while another business could use it when needed. The customer would pay according to how much computing power it actually consumes.
The concept is surprisingly similar to money in a bank. You do not need to own every resource yourself. You simply access what you need, use it, and pay for it.
China is also exploring the idea of a “compute supermarket.” The concept would allow customers to compare computing resources based on factors such as location, hardware type, availability and pricing. Instead of buying an expensive machine outright, a company could select a computing package that fits its project.
The pricing could eventually work in different ways. Businesses might pay for GPU-hours, CPU-hours, processing capacity or even the number of AI tokens generated. In other words, computing could become something businesses consume as a service, rather than something they must permanently own.
Several early partners were reportedly introduced during a technology conference in Beijing held from September 2 to 4. But announcing participants is only the beginning. The bigger challenge is creating common rules for this new marketplace.
For example, what exactly should count as one unit of computing power? Is an hour on one type of GPU equivalent to an hour on another? Should computing cost the same during peak demand and overnight? And what happens when a customer requests a particular processor but receives access to a different type of hardware?
These questions matter because AI workloads can be extremely sensitive to hardware performance.
If China can solve these problems, the impact could extend beyond its own AI industry. Smaller companies could gain access to powerful infrastructure without spending huge amounts on servers. Universities, researchers and startups could potentially rent computing power only when they need it.
That could change the way people think about AI infrastructure.
Today, computing power is often treated like a machine: you purchase it, install it and maintain it. Tomorrow, it could behave more like electricity, mobile data or money — available through a shared network whenever demand appears.
For countries and startups struggling with expensive GPUs and limited infrastructure, that model could be particularly interesting.
The future of AI may not depend only on building more computers. It may depend on making every available computer work harder and sharing its unused power more intelligently.
