CapEx for compute

Every serious machine I ever brought into any business I was responsible for forced the same three-way question: build the capability, buy the equipment, or lease it.

We ran the math every time. Payback. Utilization. The cost of being wrong. Nobody in my industry would dream of signing for a six-figure machine on a hunch and a vendor demo.

And yet that’s how a lot of companies are signing for compute right now.

The cloud bill used to be a utility bill

For years, cloud spend behaved like electricity. Variable. Elastic. Scale up in a busy month, scale down in a quiet one. Finance coded it to opex, watched the trend line, and moved on. That treatment was fair, because the commitment underneath it was small.

AI ended that. Training runs and inference at scale come with reserved capacity, multi-year committed spend, and contracts sized like equipment purchases, signed, too often, like software subscriptions.

What the smart CFO knows, is that a commitment that behaves like debt is a capital decision, regardless of its geography in your books.

I wrote recently about the finance seat being pulled back toward atoms. Compute is the sharpest version of it, because it hides so well. It arrives as a monthly invoice. It should arrive as a capital request.

Build, buy, or lease: the machine framework

The discipline that governed every machine purchase translates directly. Three paths, three risk profiles:

Build. Own the hardware or take dedicated capacity in a data center. Highest control, highest commitment. And the machine ages. With every piece of equipment I approved, the real question was never just the price, it was what that machine would be worth in year three, when a newer model changed what clients asked for. Compute hardware carries the same obsolescence clock, and it ticks faster than most. Build only what you are confident you will saturate.

Buy. In compute terms, this is reserved capacity and multi-year committed spend. You trade flexibility for price, and the discount is real. So is the risk: commit to capacity your demand never fills, and the savings become the most expensive line on your P&L. The lease-versus-buy math I ran on equipment applies cleanly: what utilization makes the commitment cheaper than staying flexible, and how much do I trust the forecast that gets me there?

Lease. On-demand, pay-as-you-go. You pay a premium for the right to walk away. When demand is early and unproven, that premium is worth every penny, it’s the rental machine you run before you know the service sells. Flexibility is not waste. Flexibility is what you pay when you don’t yet have the information a commitment requires.

Five questions before anyone signs

1. What utilization makes this pay? Name the number. Then name who forecast the demand behind it, and what their last forecast did against actuals.
2. What does it cost to be wrong? Exit clauses, transfer rights, the sunk portion if you walk. A reserved commitment has no resale market. Price the mistake before you price the deal.
3. What happens if demand halves or doubles? A commitment sized for the base case should survive the bad case and leave a path for the good one. If it can’t do both, it’s the wrong size.
4. Where does it land: P&L or balance sheet? The answer depends on how the deal is structured and which accounting standards you report under, and it changes your EBITDA, your covenants, and how a buyer reads you. This is a conversation to have with your accountant and your auditor before signing. Not after.
5. Who owns the decision? If engineering is signing capital-scale commitments alone, you don’t have a technology strategy. You have a governance gap.

The part the calculator can’t do

Pricing tools can compare providers in seconds. AI will happily model the crossover point between on-demand and committed spend.

What no tool can do is size a commitment against a demand forecast it has reason to doubt. That’s judgment. It’s built by living with capital decisions long enough to feel where the forecast is soft and it’s exactly what the finance seat is for.

So I’ll leave you with this: nobody would have let me sign for a machine without the math. Don’t let anyone sign for compute without it.

Run it like a machine. Because that’s what it is.