There is a tempting comparison doing the rounds: AI is "the new electricity", tokens are metered like kilowatt-hours, and both get sold like a utility. So why doesn't Eskom, or City Power, or any grid operator simply start selling tokens? They already supply the power that every data centre runs on.
The answer is more interesting than "nobody has thought of it yet". Start from what each business actually does, and you will find the idea falls over as soon as it meets physics, accounting, or the meter on the wall.
Start with the meter
Every utility, at its heart, is a metering business. You can only sell what you can count, and you can only count what you can see at a boundary you control.
A power utility controls exactly one boundary: the point where its cables enter your property. There it installs a meter and counts the electrons passing through. It does not know, and cannot know, what happens on the other side. Those electrons might be running a GPU cluster, an aircon, a kettle, or a website nobody visits. To the grid, all of it looks the same: a load drawing current.
A token, on the other hand, only exists on the far side of that boundary. It is created when a specific piece of software, running on a specific chip, executes a specific model and produces a unit of output. The only party that can watch a token being made is the one running the software. That is why the token counter sits inside the cloud provider's API gateway, not inside a transformer on a pole in your street.
So the first-principles problem is straightforward: you cannot bill for something that happens in a place you cannot see.
Follow the value chain
It helps to lay out the whole chain, from fuel to answer:
- Coal, gas, sun or wind becomes heat or motion.
- Heat or motion becomes electricity.
- Electricity travels across the grid to a building.
- Inside the building, electricity powers chips.
- The chips run software (a model) that turns inputs into outputs.
- The outputs are packaged as tokens and sold.
A utility owns steps 1 to 3. A cloud provider owns steps 4 to 6. The handover between step 3 and step 4 is not a technicality. It is the point where one kind of asset ends and a completely different kind begins.
Look at what changes across that handover. On the utility side, the product is a pure commodity. A kilowatt-hour from a wind farm in the Eastern Cape is identical to one from a coal station in Mpumalanga. The customer doesn't care and can't tell. On the cloud side, the product is deliberately differentiated. A token from one model is not interchangeable with a token from another, and the entire competitive game is about making yours better. Commodity businesses and differentiated-product businesses are run, priced and valued in completely different ways.
Follow the money
Now look at what actually costs each business money.
A utility's costs are dominated by fuel, plant and cables. Its assets last for decades. A transformer installed today may still be in service in 2060. Capital is spent slowly, regulated heavily by bodies like NERSA, and recovered over very long periods. The logic of the business is: build once, maintain forever, charge a regulated tariff.
A cloud AI provider's costs are dominated by GPUs and engineers. A GPU bought today is economically out of date in three to five years, sometimes sooner. Models get retrained or replaced within months. The logic of that business is: buy the newest hardware constantly, hire scarce talent, and stay ahead of competitors doing exactly the same thing.
Ask a utility to sell tokens and you are asking a company built for forty-year depreciation cycles to compete in a market that runs on three-year ones. Its balance sheet, its regulator and its shareholders are all set up for the wrong tempo.
The "just buy the GPUs" trap
The obvious comeback: fine, let the utility buy the GPUs, build the data centres and hire the engineers. Then it can sell tokens.
It could. But think through what that means. It would have spent billions of rand on chips, built facilities that look nothing like a substation, written or licensed AI software, and set up an API business with usage-based billing. At that point, how much of its revenue, risk and headcount is still "utility"?
That is not a utility selling tokens. It is a cloud company that happens to own some power stations. The label changed because the business changed. Vertical integration is a real strategy, and some energy companies are experimenting with building next to data centres. But moving up the stack is a transformation, not a new product line.
A cleaner comparison
The water board doesn't sell coffee.
Water is a utility. Coffee needs water. A coffee shop's costs include its water bill. But nobody expects Rand Water to start charging per flat white, because the flat white is made from water plus beans plus a machine plus a barista, all of which sit on the customer's side of the pipe. Water is an input. Coffee is a product built from that input.
Electricity is an input. Tokens are a product built from that input. Same structure, same conclusion.
What the comparison gets right, and wrong
The "AI as electricity" idea is not useless. It captures something real about how tokens are sold: metered, on demand, priced per unit, with the customer not much bothered about the machinery behind it. That is the consumer experience of a utility.
But the comparison describes how tokens are bought, not how they are made. On the production side, tokens look nothing like power. They come from fast-depreciating hardware, differentiated software and a layer of computation the grid cannot see.
Utilities are not missing a trick here. They are supplying the ingredient to the companies that are making the product. And as AI demand grows, that ingredient may become the most valuable and most constrained input in the whole chain, which is a very good place to be, without ever counting a single token.
