AI Platforms & InfrastructureAI Infrastructure & LLMOps
Inference Cost Teardown
A clear breakdown of where their AI spend goes (by feature, model, provider and GPU), the waste behind it, and a cost plan ranking each fix by savings and effort, with quick wins they can apply in the first month.
Inference Cost Teardown
$18K–$30K
- Duration: 2 weeks
- Tier: Diagnose
- Format: Fixed scope
You leave withToken and GPU spend analysis with a costed optimisation plan
Book a Free SessionSound familiar?
Our AI bill has tripled in six months, but usage hasn’t. Nobody can tell me which feature is eating the money, or whether we really need the most expensive model for everything.
What you leave with
Spend breakdown
Format: Spreadsheet + charts
How we know it’s acceptableTies back to the actual invoice totals within about 5%
Infra and GPU utilisation findings
Format: Report section + spreadsheet
How we know it’s acceptableEvery idle or oversized resource is listed with its monthly cost
Findings report
Format: PDF
How we know it’s acceptableEvery finding has evidence, estimated saving and effort level
Costed optimisation plan
Format: Spreadsheet
How we know it’s acceptableEach action has owner, saving, effort and quality risk
Cost tagging and dashboard recommendations
Format: Short doc
How we know it’s acceptableClient’s engineers confirm they can implement it
Readout session
Format: 60-minute meeting
How we know it’s acceptableEngineering and finance attend and agree the top three actions
Also included
- Spend breakdown by AI feature, model, provider and environment, with cost per request, per user or per conversation.
- GPU and AI infrastructure utilisation review: idle GPUs, oversized instances, unused reservations, dev/test usage.
- Ranked waste findings, such as oversized prompts, missing caching, the wrong model tier for simple tasks, retries and loops, unbounded output lengths, and dev/test traffic on production models.
- A cost optimisation plan: each fix with estimated monthly saving, effort, quality risk and owner.
- Quick wins for the first 30 days.
- Cost tagging and dashboard recommendations so spend stays visible.
- Findings report and a 60-minute readout with engineering and finance.
How it runs
2 weeks
Days 1–3
Access and data: billing exports, usage logs and traces, map features to models.
OutData inventory and gap list.
Days 4–6
Breakdown: attribute spend per feature and per request; review GPU and infra utilisation.
OutSpend breakdown and unit costs.
Days 7–9
Findings: identify waste and test top recommendations on sample data.
OutRanked findings with savings estimates.
Day 10
Readout.
OutCost optimization plan, final report, readout session.
Is this right for you?
This is for you if
- Buyerthe CTO or Head of Engineering/AI whose AI bill has grown faster than usage or revenue.
- Also has to say yesFinance or the CFO, who usually raised the alarm about the bill, and the product owner of the AI feature, who must accept any change that could affect answer quality.
This is the wrong answer if
- AI spend is too small. Below about $8K–$10K a month of AI spend, the savings won’t repay a $18K–$30K study within a year; offer the free Agent Reliability Scorecard instead.
- The system is not in production yet. Without real traffic there is no real spend to analyse.
- The client wants us to implement the fixes. The Inference Cost Teardown diagnoses and plans; implementation is a separate sprint or Managed Services — Growth.
- The problem is general cloud cost (servers, databases, networking), not AI token or GPU spend.
- The client won’t share billing data, usage logs or traces.
What’s included, and what isn’t
In scope
- analysis of LLM API and GPU spend
- feature-level cost breakdown
- GPU and AI infrastructure utilisation review
- waste findings
- costed optimisation plan
- light quality checks of the top one or two model-switch recommendations on client sample data
- cost tagging and dashboard recommendations
- findings report and readout
Out of scope
- implementing the changes
- general cloud cost not tied to AI workloads
- buying reservations or changing infrastructure
- negotiating provider contracts
- building an evaluation harness (Eval Harness Sprint)
- ongoing monitoring (Managed Services — Essential or Growth)
Before we start
Before day one
- Read access to the cloud billing export and cost explorer.
- Read access to GPU instances/clusters and their utilisation metrics.
- Read access to AI provider billing and usage dashboards.
- Access to logs or traces of AI calls, or approval to add lightweight logging.
- Read access to the code that calls the models.
- A technical contact who knows how the AI features are built and can grant access within two days.
- A sample of real inputs for quality checks.
- The two-week clock starts only once access is complete.
What we’ll need from your team
- A technical contact for about 3–4 hours a week, a finance contact for about 1 hour (billing data and readout), and the product owner for the readout.
Pricing
What we need to give you a quote
- Last three months of AI provider billing.
- Cloud billing export for the same period (GCP/AWS/Azure).
- List of AI features and the models each uses.
- List of GPU instances or clusters, with reservation/commitment details.
- Whether resources and usage are tagged per feature or team.
- Whether usage logs or traces exist.
- Whether an evaluation set exists.
After we hand over
Who owns it after we leave
The client’s Head of Engineering owns the plan and Finance tracks savings. The plan names an owner for every action.
What we measure
- Baseline monthly AI spend and cost per request
- Estimated savings
- Actual spend 60–90 days after the client implements the plan
What comes next
- An implementation sprintto apply the fixes.
- Managed Services — Growthto keep cost down month after month.
- Private AI Deployment Sprintif self-hosting an open-weight model is cheaper.
- Managed Services — Essentialif they have no monitoring.
Questions buyers ask us
It shows totals by model, not which feature, user or prompt causes it, or what to change.
Sometimes right, often not. Without testing on your own tasks you may cut cost and quality together. We tell you where it’s safe.
Price per token falls, but usage grows faster, and the waste patterns stay.
They can, but it takes them off the product for weeks. We do it in two with a tested method.
Fair. That’s why we check spend on the first call and don’t sell this below about $8K–$10K a month of AI spend.