AI capacity, utilization and revenue: what is the difference?

Use this guide to: compare commercial claims with the costs and commitments behind them.

Start with the unit being sold

A colocation operator may sell committed rack capacity. A compute provider may sell accelerator-hours or reserved instances. An application may sell subscriptions, completed tasks or access to a workflow. Each has a different revenue denominator. The word utilization becomes useful only after that denominator is defined.

Before modelling demand, write one sentence: a named category of customer pays this price for this unit over this period. Then identify what constitutes an accepted, billable delivery. Do not borrow a utilization percentage from a different business model just because the underlying hardware looks similar.

Keep three ledgers

The first ledger is physical: installed equipment, available electrical capacity and operating hours. The second is technical: useful work completed at an acceptable level of quality and latency. The third is commercial: committed capacity, billed usage, collected revenue and credits. The same hour may appear in all three, but the quantities do not have to move together.

Equipment can draw electricity while processing tests, retrying failed jobs or serving an unpaid trial. Reserved capacity can generate contractual revenue without being fully loaded. Neither observation establishes whether the business is attractive. They explain why a single percentage cannot stand in for a revenue model.

Construct a small, explicit example

Assume a hypothetical fleet offers 10,000 accelerator-hours in a period. It records 7,000 active hours, but 1,000 of those hours are internal testing. If the contract bills eligible productive usage, the starting billable quantity is 6,000 hours, before service credits or other adjustments. Active utilization is 70%; the simplified billable ratio is 60%.

Now change the contract to a fixed reservation. Revenue may instead depend on the contracted reservation, subject to its terms. The physical fleet has not changed, yet the commercial calculation has. The point of this example is the boundary, not the percentage: every rate needs a documented definition.

Follow the cost boundary too

A contribution calculation should include the costs needed to deliver the commercial unit. Depending on the business, those can include model or compute charges, electricity, payment processing, human review and support. Hardware depreciation, financing and facility commitments may sit outside that contribution measure but still matter to cash flow and investment decisions.

Make a bridge from billed revenue to contribution, then from contribution to operating cash requirements. Show fixed commitments separately from costs that vary with use. This makes a low-demand case legible and prevents attractive variable economics from hiding a large minimum monthly commitment.

What would change the conclusion?

Look for contract terms, actual invoices and workload-level operating records. Test lower demand, a lower selling price and higher delivery cost one at a time. Then test a combined downside. An improvement in the forecast matters only if the underlying evidence has improved, or the reader can see exactly which assumption has changed.