The compute boom nobody budgeted for.
The world is pouring concrete for AI at a pace with no precedent in computing history. Your AI bill lives downstream of it.
The numbers stopped looking like an IT category some time ago. Gartner expects worldwide data centre power demand to rise 27% this year, to 132 gigawatts — and to keep climbing towards 290GW by 2030. Goldman Sachs Research expects US data centre power demand alone to double within two years. Worldwide spending on data centre systems is forecast at roughly $582 billion for 2026; it took twelve years for that figure to grow from $140 billion to $236 billion, and then three years to more than double again.
One company tells the story on its own: Nvidia’s data centre revenue has grown roughly 65-fold since 2020 — a segment now approaching the size of the entire global server market it used to sell into.
Power is the new constraint
The interesting part isn’t the size of the buildout — it’s what’s limiting it. AI capacity is now constrained by electricity availability, not chip supply or capital. When the scarce input is grid connections, compute stops behaving like software (infinitely copyable, ever cheaper) and starts behaving like an industrial commodity: allocated, queued, and priced by scarcity.
For a mid-size business that never plans to build a data centre, this still matters twice over. First, the unit cost of intelligence keeps falling — chipmakers are delivering 60–70% lower cost per token of inference every year — so anything you priced in 2024 is worth re-pricing. Second, everyone else’s usage is growing faster than the price is falling, which is why total AI bills keep going up. Cheaper units, bigger bills. That is a budgeting problem, not a technology problem.
What we’d do about it
Treat compute like any other utility on the P&L. Know what you’re consuming, what each workload returns, and which of your tools are riding this cost curve down versus quietly riding your budget up. Most businesses we audit can’t answer those three questions. That’s usually where the money is.
All figures in this article: Sources: Gartner, Goldman Sachs Research, 2026 · checked 2026-07-06