The problem
The problem
Seven jobs nobody in your building holds all at once. Your team can do pieces of each; what nobody has is the standing capacity to do all seven continuously, and nobody holds the contracts in one place.
The contracts are already large
In their most recent filings, non-lab public companies disclosed multi-year cloud and infrastructure commitments, each in its filing’s own category, of about $16.7 billion (Salesforce, contractual commitments including infrastructure providers), $7.7 billion through fiscal 2032 (Palo Alto Networks, cloud), $6.8 billion (Adobe, primarily hosting and data centers), a $6.0 billion cumulative minimum through March 2031 (Snowflake, amended April 2026), at least $5.6 billion through February 2036 (Palantir, amended March 2026), $5.3 billion (DoorDash, technology-platform infrastructure, from $1.07 billion), about $4.8 billion on cloud services through 2030 (ServiceNow, from $805 million a year earlier), $4.0 billion through May 2031 (Pinterest, signed May 2026), and $3.7 billion (Atlassian, purchase obligations including cloud, marketing and other services; doubled in a year). Roughly fifteen disclose $1 billion or more and about forty $100 million or more. Where the mechanics are stated, the contract is a dollar minimum with a shortfall true-up — Snowflake: “required to pay the difference”; MongoDB: “pay for such capacity irrespective of actual usage” — and Tempus AI carries a $39.5 million liability for an Azure commitment “in excess of forecasted usage,” assumed with its Paige acquisition. Snowflake reports third-party cloud “including those related to AI inference and graphics processing units” at about 75% of its cost of product revenue in the quarter to 31 July 2026, up from 70% a year earlier. These are finance objects, held in a spreadsheet and a footnote.
See what you actually pay.
Your blended rate per GPU-hour is spread across committed discounts, on-demand overflow, interruptible capacity, and an enterprise agreement that bundles compute with storage and networking. Three professional price indexes now exist and they disagree with each other, because they measure different chips, regions, and contract shapes under one name. One of them publishes the size of the problem directly: on 8 September a regulated index printed Hopper at $2.66 blended against $2.97 listed and Blackwell at $4.37 against $5.30 — advertised premiums of 11.7% and 21.3%, from a benchmark with no stake in our argument.
Hold every commitment in one place.
Floors, ramps, expiries, prepayments, credits, and side letters live in a spreadsheet, a footnote, and the inboxes of the people who signed them. Over-commitment becomes a liability; under-commitment runs at on-demand rates; nobody sees the drawdown pace until the term is nearly over. The sell side has organized itself around getting the signature early and large: CoreWeave described “a typical five year contract,” reported about $104 billion of backlog at 30 June 2026, and said its committed contracts are take-or-pay; Nebius said roughly 70% of the deals it closed that quarter carried customer prepayments.
Authorize compute before it is spent.
Project budgets are set once a year and exhausted in a quarter. Seventy-nine percent of finance leaders at companies with a thousand or more employees reported AI cost overruns in the prior year, and the most mature cost-management shops overran most. The first notice is the invoice, because nothing reserves the money when the work is approved and nothing debits it when the work runs.
Compare two suppliers on the same terms.
The advertised rate leaves out moving the data, standing the environment up, and work rerun after an interruption. And identical hardware does not deliver identical work: in one benchmark study, throughput on nominally identical H100s ranged 34.5% across providers on the PCIe sample and 10.4% on SXM. The only way to know what you will get is to measure your own work on the actual allocation.
Look anywhere but your incumbent.
Most companies renting AI compute rent it from one vendor, on a commitment signed before anyone knew what the year’s work would look like. Qualifying a second supplier means a security review, a contract, an acceptance test, and scheduler integration — weeks of specialist work nobody is free to spend. So your whole price history is one seller’s price list, and in 2026 one supplier can mean no supplier: one Azure customer waited ninety days across five regions for any GPU quota at all. On posted rates on 4 September 2026 the first alternative to a hyperscaler was priced 1.12× to 1.72× below it; across a 53-provider census, the range was 7.1×.
Say which of your work could move, and act at the speed prices move.
Deadline tolerance is not in your job logs, and neither is tolerance for interruption. Both live in your engineers’ heads, and they are the two facts that decide the size of the prize. Meanwhile the prices move: AWS raised Capacity Block reservation prices three times in 2026, about 15% in January, 20% in July, and 15% in October; CoreWeave raised prices about 25% in July; and requoting five suppliers is a week of somebody’s month.
Reconcile the invoice, and prove any decision was right.
Invoices arrive from different suppliers in different units; the largest GPU cloud’s billing export carries usage quantities and no cost fields. And nobody wrote down what the work would otherwise have cost, so there is no baseline to check the saving against. Estimated wasted cloud spend now runs at 29%, up for the first time in five years.
Same silicon, different place
ConceptualThe price difference is the whole opportunity: the same accelerator, in the same window, priced differently by place. Delivered cost decides, so a discount that is eaten by egress, storage or a slower fabric is not a discount. Suppliers you have not qualified are not on the map.
Scroll the figure sideways →
What a decision is worth — three baselines, never one
Every saving we report is measured against three baselines, and our value is the difference between the last two.
BASELINE 1
Stay put
What the purchase costs at your incumbent, as it stands.
BASELINE 2
Your best alternative without us
A competent team, one negotiated second supplier, a router. Agreed with you before the work starts.
BASELINE 3
The plan with us
Our value is 2 minus 3. The published model prints 1 minus 3.
On a representative fleet — $60M a year of AI compute, 30% of the work able to move, and a 1.45× gap against the cheapest supplier you would accept — our open model returns $2.53M a year against the incumbent after every deduction, including our subscription and your own cost of adopting us. That is 4.2% of the bill. It is the purchasing opportunity, it is a model estimate on assumed inputs, and it is not our incremental value. That number is computed on your engagement, and we have not yet measured it on a real purchase.
| The waterfall, default fleet | $ a year |
|---|---|
| Price gap on the spend actually moved | +2,849,000 |
| Interruptible discount on the share that can take interruption | +886,000 |
| Gross benefit | 3,735,000 |
| Execution: rerun after interruption, extra checkpointing, slower interconnect, data transfer at your rate | −711,000 |
| Implementation: supplier onboarding, and your own integration and ongoing effort, first year | −139,000 |
| Our subscription, the tier for a $60M fleet | −360,000 |
| Net purchasing opportunity | 2,526,000 |
Five lines, never one: the gross benefit, the execution cost, the implementation cost, the subscription, and the net. Nothing is counted twice, and the subscription can be read against the gross.
Four value lines, reported separately
| Line | What it is | How it is stated |
|---|---|---|
| Cash saved | The price of the hours bought, against your best alternative | Modeled before; reconciled to the invoice after |
| Commitment avoided | Unused committed capacity is a write-off at a hundred cents; demand above the commitment lands on demand; a commitment sized to measured demand avoids both | Per point of the bill; you set the point; counted only on reconciled actuals |
| Output gained | Work that ran because capacity was found in the window | Your number; we do not put one on it |
| Risk reduced | Counterparty exposure capped; prepayments sized; the deadline-miss probability you set, held; budget-at-risk reported | Stated, not priced |
The lines overlap and are never added. Delivered performance is measured on your work, not estimated for your fleet: in one study (Slynko et al., “Did You Win the GPU Cloud Lottery?”, GPGPU ’26), inter-provider throughput on nominally identical H100s ranged 34.5% on the PCIe sample and 10.4% on SXM, and device throughput predicted application throughput weakly. That is a reason to test, not a number to multiply.