The first sentence is always “we can’t afford the GPUs.” The second one is usually about bijli. Ask any founder in Karachi who’s tried to train a model.
China has the opposite problem. It has the chips — what it doesn’t have is a way to get them to the right place at the right time. Demand shows up in the wrong city, at the wrong hour, wanting a different card from the one that’s free. One company queues while another’s racks run cold.
So it’s testing a “compute bank.” Nothing physically moves. A data centre with spare capacity parks the access on a shared platform; whoever needs it withdraws, burns it, pays for what they burned. Next door is a “compute supermarket” — you shop by location and by billing unit, card-hours or core-hours or per token, like picking a mobile package.
The first 10 partners were named at a conference in Beijing on 2–4 September. That’s the easy part. Nobody has agreed what one unit of compute is, or why the same chip should cost the same at 9pm and 4am, or who eats it when you order one card and get handed another.
If it works, compute stops being a machine you buy and becomes a thing you top up. Which is a familiar idea in a country where most people met the internet through a data bundle, not a broadband contract.
