Cost Mechanic

Cloud · Kubernetes · LLM spend engineering

We take your cloud bill apart.

Diagnosis first. We instrument the bill until every dollar has an owner, a rule and a residual you can check.

Read-only access you create and revoke. A fixed $900 entry. Find less than three times the fee, pay nothing.

Book the export check

For the platform teamYou own the infrastructureUnlabelled GPU nodes, shared cluster cost, dimensions never recorded.See the attribution work For financeYou own the numberA gross margin and cost per customer that survive diligence.See the arithmetic

You own the infrastructure

The bill is one number. The estate is thousands. Attribution is the join.

How cloud and AI spend becomes an owned numberA pipeline diagram in two converging planes. The billing plane carries provider cost exports (AWS CUR 2.0, the Azure cost details export, the GCP BigQuery billing export and LLM provider usage) into a normalization step that rewrites every row into the FOCUS schema: one row per resource, per charge, per hour, carrying BilledCost, EffectiveCost, ResourceId, tags and environment. The application plane carries request logs and runtime telemetry into a sessionization step that reduces each request to a unit of work labelled with tenant, feature, model, tokens and duration. The two planes meet at a join on ResourceId to node, tags to service, and timestamp window to request; unmatched rows are flagged rather than dropped. Attribution then assigns direct cost to the resource owner and splits shared cost by a declared driver such as requests, gigabyte-seconds or tokens, keeping any residual whole. Three numbers come out: cost by owner, cost by customer, and cost by feature. Two checks hang off attribution: a reconciliation harness asserting that attributed cost plus residual equals the invoice or the build fails, and an unattributed bucket that is reported rather than spread silently.BILLING PLANEAPPLICATION PLANEAWSCUR 2.0 · hourlyAzureCost details exportLLM providersusage + cost APIGCPBigQuery exportRequest logsroute · tenant · tsRuntime telemetrypod · GPU · durationFOCUSNormalizeresource × charge × hrBilledCost / EffectiveResourceId · Tags · EnvAPPSessionizerequest → unit of worktenant · featuremodel · tokens · msKEYSJoinResourceId → nodeTags → servicets window → requestunmatched → flaggedRULESAttributedirect → resource ownershared → driver splitrequests, GB-s, tokensresidual → kept wholeOwnerteam · serviceCustomercost per tenantFeaturecost per unit shippedReconcileattributed + residual= invoice, else failUnattributedreported, neverspread silently1CUR 2.0 · line_item_unblended_cost · split_line_item2BigQuery billing export table3GKE cost allocation · shared splitFLOWCHECK · RESIDUALSCHEMATIC · FIELD NAMES, NO FIGURESHow cloud and AI spend becomes an owned numberA pipeline diagram in two converging planes. The billing plane carries provider cost exports (AWS CUR 2.0, the Azure cost details export, the GCP BigQuery billing export and LLM provider usage) into a normalization step that rewrites every row into the FOCUS schema: one row per resource, per charge, per hour, carrying BilledCost, EffectiveCost, ResourceId, tags and environment. The application plane carries request logs and runtime telemetry into a sessionization step that reduces each request to a unit of work labelled with tenant, feature, model, tokens and duration. The two planes meet at a join on ResourceId to node, tags to service, and timestamp window to request; unmatched rows are flagged rather than dropped. Attribution then assigns direct cost to the resource owner and splits shared cost by a declared driver such as requests, gigabyte-seconds or tokens, keeping any residual whole. Three numbers come out: cost by owner, cost by customer, and cost by feature. Two checks hang off attribution: a reconciliation harness asserting that attributed cost plus residual equals the invoice or the build fails, and an unattributed bucket that is reported rather than spread silently.BILLING PLANEAPPLICATION PLANEEXPORTSAWS CUR 2.0Azure cost detailsLLM usage APIGCP BigQuery exportEVENTSRequest logs · tenantRuntime · pod · GPUFOCUSNormalizeresource × charge × hrBilledCost / EffectiveResourceId · Tags · EnvAPPSessionizerequest → unit of worktenant · featuremodel · tokens · msKEYSJoinResourceId → nodeTags → servicets window → requestunmatched → flaggedRULESAttributedirect → resource ownershared → driver splitrequests, GB-s, tokensresidual → kept wholeOwnerteam · serviceCustomercost per tenantFeaturecost per unit shippedReconcileattributed + residual= invoice, else failUnattributedreported, neverspread silently1231CUR 2.0 · line_item_unblended_cost · split_line_item2BigQuery billing export table3GKE cost allocation · shared splitFLOWCHECK · RESIDUALSCHEMATIC · NO FIGURES

  • GPU-UNLABELLED

    H100 hours land in cur2 with no owner and no job.

  • SHARED-CLUSTER

    Pod requests decide the split. Without kube_state_metrics, idle headroom goes unassigned.

  • DIM-NOT-RECORDED

    FOCUS 1.1 normalizes what was captured, not what was never recorded.

You own the number

Gross margin per customer is one division. Everything hard is above the bar.

costdivided bycustomer
cost = direct + shared × split + commitment ÷ term

A number survives diligence when someone else can rerun it. We publish the split rule, the amortization window and the error bars first.

Read the arithmetic

The object, not the promise

What you receive

Every dollar carries a tenant, a rule and a residual. It reconciles to the invoice, or it fails loudly.

Instrumentation Build

Synthetic sample of the cost-per-tenant sheet. Invented figures, exact arithmetic: the line items sum to the total.
tenant90-day cost · USD
acme-prod18,412.06
globex-prod11,207.44
initech-batch7,918.50
shared · split by driver4,566.00
residual · unattributed1,102.13
reconciles to invoice43,206.13

How the work runs

Ordered · each stage gates the next

  1. Diagnose

    Three days on your exports, then a written go or no-go.

    Spend Teardown

    $9003 days

  2. Instrument

    Cost per customer, feature and environment into production, FOCUS-normalized.

    Instrumentation Build

    Fixed feeat the debrief

  3. Repair

    Routing, caching and attribution shipped behind flags, measured on your traffic.

    Inference Cost Controls

    Fixed feeat the debrief

Price

Starts at

$900

Export check · three days · read-only

The rungs past the export check
Spend teardownWhere the money goes, reconciled to invoice
Instrumentation buildCost per customer, feature, environment
Inference cost controlsRouting, caching, capacity, shipped
Margin watchWeekly variance, fixes as pull requests

Each rung is a fixed fee, priced at the teardown debrief from the teardown’s own findings. Find less than three times the fee, pay nothing.

Guarantee and terms

Guarantee, access, and what happens next

The guarantee

If the teardown finds less than three times its fee, it costs nothing.

The exact terms

Read-onlyyou create ityou revoke itThe exact policy

Before you book

Can you name the event driving this? A board deck, a credit expiry, a margin floor?

If you cannot, it is not urgent yet.

Book the export check