What we don't claim

“GPU-hours are not interchangeable.”

Correct, and it is the real hole in the energy analogy. A megawatt-hour is identical from any generator; an H100-hour is not. Interconnect, storage speed, driver versions, and noisy neighbours all move what you actually get. How much it matters depends on the work: for a batch sweep, barely; for a communication-bound distributed train, a great deal. So we carry network topology as a hard constraint where the work demands it, and we run a real interconnect test across the machines at provisioning rather than assuming.

“That test is not free, and one benchmark is not the truth.”

Also correct. It costs minutes of paid time on every allocation and it measures one moment. It is priced into the model as a deduction, not waved away, and repeated on a schedule rather than once. One delivery of 4,088 GPUs had about 10% of its machines fail to boot; a cluster, as NVIDIA put it, “can pass every health check and still fail to run an AI workload.”

“Ramp, Flexera, and my cloud provider already do this.”

Ramp governs token spend, for free, and does it well; the record ingests what Ramp holds. Flexera, Apptio, CloudZero, and Vantage see the hyperscaler bill and automate hyperscaler commitments, and Vantage now meters GPU usage on CoreWeave and Nebius clusters at rates you type in. None of them ingests a GPU cloud’s contract, commitment, or invoice, authorizes a project before it spends, or measures what an allocation delivered on your work. Your cloud provider’s own tools stop at your cloud provider. None of them holds your contracts across suppliers or decides where work should run on price, deadline, and the contract. Where they are right, the record says so and leaves them in place.

“AWS says I cannot buy Capacity Blocks through a third party.”

It does, since 1 October 2026, and we read the clause. For AWS Capacity Blocks the desk prepares the order — the requirement, the quantity, the window, the budget line — and you submit it under your own identity; the record captures the confirmation. We ask every supplier the same question in writing before the desk places anything in its own hands, and we never treat an API as a permission.

“Why sell to finance when my FinOps team reports to the CTO?”

Because finance pays and holds the contracts, and the record enters through what finance already owns. But the data lives with technology — 78% of FinOps practices report to the CTO or CIO and 8% to the CFO, in the FinOps Foundation’s 2026 survey — so the people who run your scheduler are in the first meeting, and their job-log access is the second signature, not an afterthought.

“A record is not a saving.”

Agreed, and the method page says so. The record prevents the liability and the write-off and gives finance the accruals, drawdown, and true-up liability the month-end needs; it does not print a saving. The saving is the desk’s, it is reported only against a baseline agreed first, and commitment avoided is counted only on reconciled actuals.

“You have not priced data transfer properly.”

The 2024 egress waivers are exit-only and do not apply to ongoing work. We model your negotiated rate, and the term that matters is recurring checkpoint transfer rather than the one-off corpus move everyone quotes.

“This will wreck our commitment economics.”

Overflow above an agreement usually still earns your negotiated discount, and a spending commitment is not prepaid capacity. Both narrow the prize and both are in the model, the record holds the commitment as the contract it is, and the decision record prices it as it stands before anything is recommended.

“Moving a running job between clouds.”

Not a feature and not claimed. The literature’s most-cited system reports 28–228 seconds to move a job inside one datacenter with a shared filesystem. Across providers, over a wide-area link, nobody has measured it because nobody does it. We place work before it starts and let it finish where it started.

“If we already run SkyPilot, what do you add?”

Less than a router’s vendor would guess. Run it well against a stable set of suppliers and you have the routing half solved. What it does not do is formulate the ask, run a quote request against suppliers you have qualified, hold a forward view built from your own purchases, hold your contracts, or reconcile an invoice to a baseline stored at the moment of decision.

“Aren’t you just an AI wrapper?”

The opposite, and the distinction is the whole product. At the decision point we run a solver, not a language model: the same inputs produce the same schedule, and every decision is reproducible from stored inputs. The language model has two bounded jobs — parsing an unstructured quote reply or a contract clause into a typed field, and drafting an explanation of a decision already made — and a parsed quote is treated as untrusted input and bounds-checked before it can enter the optimizer. A saving that cannot be reproduced cannot be audited, and a desk that cannot be audited has no business being paid by the buyer.

“Why not just hedge it?”

With what? No listed compute contract has traded; the CFTC is still reviewing the first one and wrote in August that it “preliminarily believes that compute may not yet exhibit certain of the characteristics of commodities.” The one live complex holds about $266,000 of open interest. And the accounting treatment of a compute hedge is unresolved. Your hedge today is the commitment, sized to measured demand; when a contract trades, it is an overlay on a physical purchase, gated, and never the headline.

“Timing gains are mostly noise.”

Largely true. The value is in the gap between suppliers and in the commitment, not in the clock, and we will never lead a proposal with timing.

“Andromeda already does this, and sends one invoice.”

Most of the buying, on standardized contracts, with one invoice — and the record it keeps is Andromeda’s. For a buyer who wants one counterparty and does not need the baseline, that is the right product, and our readiness answer will say so. It is the wrong product when finance has to defend the purchase, when your work cannot leave your own contracts, or when the spread is your margin. Andromeda is the strongest evidence we have that the job is real.

“The market is tight. Nobody will quote.”

On-demand GPU capacity sold out across chip types in April 2026, one-year H100 contracts rose about 40% in five months, and Compute Exchange says it cannot fill most of its requests. A desk that finds cheaper capacity is worth most when capacity is loose. In a tight market the value moves to sizing the commitment, administering it, and qualifying the suppliers who will quote — and an unfilled request is an outcome in our record, not a footnote.

“Your own example shows $945 of incremental value.”

It does, and it is on the front page beside the $6,143, because hiding it would be worse. Whether real purchases look like that campaign is what the first watches find out. If they do, the desk is smaller than this site describes, the record is still the record, and we will say so.

“Isn’t this a FinOps tool, or an ETRM, or a procurement suite?”

It overlaps each at one edge and replaces none. The FinOps platforms see the hyperscaler bill and automate hyperscaler commitments, and one now meters GPU usage on two GPU clouds; none holds a GPU cloud’s contract or invoice or the budget that authorizes the spend. The energy trading and risk systems have no compute module and have announced none. Coupa and Zip hold the purchase-to-pay workflow and the intake; Kyriba holds the cash; SAP holds the ledger. The record ingests from them and exports to them — the position, the accruals, the drawdown, the chargeback — the way a utility’s trading system sits beside its ERP. What none of them has is the contracts across suppliers, the budget that authorizes compute before it is spent, the delivered performance, or a desk paid only by you.

“Why should I believe any of this?”

You should not, yet. Run the model on your own numbers, including the inputs where it tells you not to hire us. Read the method and the sources. Then bring us one purchase you have already made, and check the answer against an invoice you already hold. That is the whole sequence, and it is deliberately short.