The long version

Market position

Six tiers sell cloud cost work; the seventh row of the table below is this studio, placed there so the comparison is on one axis. They differ less in method than in what they will put in writing before a call. None of the seventeen firms and vendor programs surveyed publishes a deal size, a client-size band or a staffing ratio.11

Scope, and the argument against it

The research argued for a narrow wedge: AWS, Kubernetes and LLM spend only, refusing Azure and GCP for the first twelve months. The scope actually sold is wider: AWS, Azure, GCP and AI spend. That decision was made deliberately and it creates a real credibility problem, because a studio claiming three clouds reads as a generalist.

The compensating control is the depth-per-estate section below. Every estate states its export of record and one named failure mode you can check against your own bill in an afternoon. Where depth is scoped rather than deep, the word used is “scoped”, and re-architecture on those estates is referred out rather than sold.

All major providers now export FOCUS-conformant billing data,6 which is what makes a single normalized schema across three estates feasible at all. GCP’s FOCUS export was still Preview in August 2026,7 MEDIUM CONFIDENCE so on GCP it is a cross-check and the detailed usage export to BigQuery is the pipeline.

How the ladder is priced

One price is published: the $900 export check, three days, credited in full against the teardown. Every rung past it is a fixed fee in USD, quoted at the teardown debrief from the teardown’s own findings and fixed at signature. Nothing is a starting point and there is no hourly rate.

Questions and artifacts

Every question this offer has to survive has a written answer somewhere on this site. This is the index of which artifact answers which question. Nothing in it is behind a call.

Depth per estate

Depth is claimed as a number of cells out of five, with the export of record named and one failure mode stated in enough detail that you can check it against your own bill. A claim you cannot check is not a claim.

The arithmetic

The full derivation, the queries and one worked synthetic month live at methodology. This is the shape of it, so you know what you are opening.

The cost basis is amortized effective cost, never unblended: unblended is wrong for unit economics by exactly your commitment discount. Shared cost is split in five ordered passes, and the order is the claim: each pass removes cost from the pool the next one is allowed to split.

Guarantee mechanics

If a teardown does not identify three times its fee in annualized savings, you pay nothing and keep every artifact: the report, the queries and the model.

A guarantee is only as good as its definitions, so both halves are published. The exclusions are the half that works against the studio.

Counts as a finding

Itemized, owned, priced

Traced to rows in your export

At your effective rate, never list

Excluded from the 3×

Already ticketed before kickoff

Pricing or contract changes

Engineering time, risk, reliability

Acceptance criteria and the cap per rung are on pricing. Savings are counted at your effective rate, never at list, which is the difference between an honest number and a flattering one.

What gets refused

Tooling has absorbed commitment laddering, basic rightsizing, Kubernetes allocation math and pre-deploy estimates.14MEDIUM CONFIDENCE Selling those by hand would be selling a worse version of software you can buy. Each row below is declared on the service page that excludes it.

What is absent

Buyers at this size run a security and vendor review. These are the findings that review will produce, written down before it starts. Mid-size and enterprise buyers routinely require seven-figure per-claim and aggregate errors-and-omissions cover;12 none is bound.

Access boundary

Every credential is created by you, scoped to read, time-boxed to the engagement and revocable by you at any time. There is no write access anywhere in the ladder, and no access to prompt or completion payloads. The exact policy documents are on access.

Further reading

The long-form depth is in the notes. Each states whether its figures are synthetic or drawn from public data before you start reading it.

• The access appendix, in full9 min · public data
• When not to hire a cost mechanic6 min · public data
• A cost model that ties out exactly is hiding something13 min · synthetic data
• What FOCUS normalizes, and the five things it does not15 min · public data
• The node is the billing unit, the namespace is the ownership unit15 min · synthetic data
• Per-tenant AI cost is a join, not a report18 min · synthetic data

Sources

Every market figure on this page is numbered here with its confidence grade and the research artifact it came from. Where the two artifacts disagree about a grade, the lower one is used and the disagreement is stated. Source 15 is the only case so far, and it is left in the list rather than quietly dropped, because the way a studio handles a figure it cannot stand behind is more informative than the figure.

If you got this far

The first step is a $900 export check: three days, read-only access you create and revoke, and a written go or no-go on whether your billing exports can support the rest of the ladder at all. It is the cheapest way to find out that the answer is no.

Write the email ·hello@costmechanic.com