← Blog

Everyone is buying. Almost nobody is measuring.

In a year of procurement conversations, the question that almost never comes up is what share of the hardware will be doing work six months in.

Ammar Enaya Co-founder and CEO 2 min read

I have sat in a lot of buying conversations this year. They follow a pattern.

How many accelerators. What generation. When can they be delivered. What is the lead time on the next tranche. All sensible questions, all about acquisition.

The question I almost never hear is what proportion of that hardware will be doing useful work in month six.

Where the number usually lands

Published measurements put most AI clusters somewhere between 30 and 50 per cent utilisation, with inference silicon idle more than half the time. Those are not our figures, and they are not controversial. They have been reported often enough that they should be a standing item in any procurement review, and in most of the reviews I have been part of, they are not mentioned at all.

Half of a large capital purchase is quietly not working, and it does not appear on anyone’s objectives.

Why it stays invisible

Partly because it has no owner. Utilisation is decided by allocation policy, by the memory ceiling, by how good the kernels are, and by how much thermal headroom the room leaves you. Those four things sit in four different teams, and in a lot of organisations two of them are outsourced. Everyone can point at a number they are hitting.

Partly because the alternative is uncomfortable. If you accept that half the estate is idle, the next question is why you are ordering more of it, and that question is much harder to answer in a budget meeting than a lead time is.

A cheap thing to do about it

Before the next order goes in, ask for one number: sustained device utilisation across the estate over the last ninety days, measured at the device rather than at the job scheduler. Not peak, not average during business hours. Sustained.

If nobody can produce it, that is the finding. It is also the cheapest piece of due diligence available, and it usually changes the size of the order.