The method

A saving that cannot be reconciled to an invoice is not a saving, and a record that cannot be audited is not a record. Here is the arithmetic, the baseline it is measured against, and the point at which we stop claiming it.

An index cannot test your allocation

Three professional indexes now publish compute prices, and one of their administrators sells a benchmark tool. What none of them can do is run your own work on the allocation you are about to pay for, under an acceptance condition the supplier has agreed to. An index aggregates to one printable number; a test tool is only as good as what it is pointed at. In one study (Slynko et al., GPGPU ’26: eleven providers, 3,534 GPU instances), throughput on nominally identical H100s ranged 34.5% across providers on the PCIe sample and 10.4% on SXM.

So we measure it, using an existing benchmark suite where one fits your workload. The test runs on your actual allocation before any work starts, and a failed test has an agreed consequence — repair, reallocation, or cancellation on agreed terms — because a supplier sells hardware and uptime, not a warranty on your training throughput. That is the answer to the reasonable question of why you would not subscribe to an index instead: the index answers a question you did not ask. We mark your position against all three indexes, and we show the basis between them and your own contracts, because the two live H100 indexes disagree with each other week to week.

Which rate is the baseline

The three baselines on the problem page — stay put, your best alternative without us, the plan with us — each need a rate, and the rate is not your list price. Your blended effective rate is the reporting rate: what the whole fleet costs today, and what a saving gets reported against. The marginal rate at your incumbent, for this specific work in this specific window, is the decision rate: what you would pay if you did nothing. They are different numbers, and mixing them is how savings claims in this sector get inflated.

The baseline is written before the decision

Before a purchase is decided we store what would otherwise have happened: the requirement, the quotes considered, the timestamp, the constraint that bound, the approver, and the alternative agreed as your best route without us. It is append-only, in storage you own, and it cannot be edited afterwards. Later, the invoice is matched to it line by line. Where a saving has no feasible alternative recorded against it, we show it and leave it out of the number.

A record is not a saving

Holding your contracts in one place prevents the liability and the write-off; it does not print a saving, and we do not count one for it. The one exception is counted only on reconciled actuals: a commitment that was sized to measured demand, against the shortfall or the overflow it replaced. Everything else the record does for finance — the position, the drawdown, the reconciliation, the authorizations — is stated, not priced.

Every number carries its status

Four labels, used everywhere on this site, so you always know what you are looking at.

Where it fails

About a third of the plausible range falls below our own 12% screening rule. We publish the whole grid rather than the passing part, because a desk that cannot tell you when not to hire it is a sales motion. Read it as a curve: the price gap and how much of the movable work you actually move trade off against each other, and a single make-or-break number would be wrong in both directions.