AI is electricity. Buy it that way.

In 2017, Andrew Ng said that just as electricity transformed almost everything 100 years ago, he had a hard time thinking of an industry AI would not transform in the next several years. It reads as a metaphor. Today it also reads as a description of the supply chain.

Here is the number worth carrying around. The International Energy Agency puts data center consumption at about 415 terawatt-hours in 2024, around 1.5% of the world’s electricity, and says it is set to more than double to around 945 TWh by 2030. That is the agency’s projection, not mine.

Why software started needing power plants

My read is that a search engine mostly looks up pages that already exist, and a model writes a new answer every time it is asked. Every answer is manufactured token by token, and each one has a marginal cost.

The research behind today’s models has a cost curve in it. Kaplan and colleagues found that loss scales as a power law with model size, dataset size and training compute, with trends spanning more than seven orders of magnitude. Under a power law, each further drop in loss takes a constant multiple of the compute that bought the last one. That is arithmetic, not a finding of the paper.

Software used to be nearly free to copy. This software is made in factories, and the factories run on electricity.

What the number does not tell you

It is not the cost of your AI use. It tells you where the constraint sits. My read is that the constraint is physical: power, cooling, memory, and the grid equipment that connects them.

Take the equipment. In August 2025, Wood Mackenzie projected a supply shortage of 30% for power transformers in 2025, with lead times and prices escalating across all transformer specifications. A better model does not fix a transformer. A vendor cannot promise a date for something it does not control.

The procurement read

If AI behaves like electricity, stop buying it like software.

Software is often priced per seat and renewed once a year. Electricity is often priced per unit consumed, and peaks can cost more. A supplier’s input costs can reach the tariff, which is the schedule of prices a utility charges for each unit.

Where AI is billed by the token or by the call, that is a tariff too, and it rewards being managed like one:

• Forecast consumption, not seats. A pilot can look cheap because few people use it. Inference is the running cost, and the meter turns every time someone asks a question.
• Know your peak. Batch jobs, agents running overnight and a month-end close can all land on the same capacity at the same hour.
• Ask what passes through. A price held flat for three years on a usage-priced input is a promise worth reading closely.
• Decide in advance what you do when the price moves, in both directions. Switch, cap, renegotiate or absorb. Pick before the invoice arrives.

Most of this is ordinary energy buying. A utility contract is read for the meter, the peak and the pass-through. Ask the same questions of an AI contract. The product arrived looking like an app.

One thing to check this week

Pick one AI tool your team pays for. Pull three months of invoices or the usage dashboard and answer three questions. What is the unit: seat, token or call? What was the total? Who in the business drives the volume?

If nobody can answer the third, you are buying electricity without a meter.

Ng said he could hardly think of an industry it would not touch. The part the metaphor left out is that electricity has a supply chain, and so does this.