YTJK Systems

Home / Insights / Architecture

Cutting cloud spend without slowing delivery

Architecture · Reduce cloud and AI cost

Problem

Cloud and AI spend rising faster than revenue, with no clear owner.

Approach

Allocate every cost, fix the biggest levers first, then automate the guardrails.

Outcome

Lower, predictable spend with unit costs leadership can track.

Most cost programmes save money once and then drift back. Lasting savings come from three things working together: every cost has an owner, the largest levers are fixed in the architecture rather than by one-off clean-ups, and the guardrails run inside delivery pipelines instead of in a monthly spreadsheet.

The business problem

Cloud and AI spend grows quarter on quarter, but nobody can say which product, customer or feature drives it. Finance sees one large invoice; engineering sees hundreds of resources; neither can connect spend to value. Blanket cuts slow delivery and the savings rarely last.

Constraints

  • Delivery speed and reliability cannot drop as a result of cost work.
  • Several teams and products share accounts, clusters and data platforms.
  • AI usage (tokens, GPUs, vector stores) is new, variable and hard to forecast.
  • Commitments (savings plans, reservations) must not lock in the wrong architecture.

Architecture

Cost control loopContinuous, owned by teams, reported to leadership.
From invoice to unit costRepeats every month
  1. AllocateTags, accounts and labels map every cost to a product and team
  2. Show backCost per product, team, customer and AI request
  3. OptimiseRight-size, schedule, commit and re-architect
  4. GuardBudgets, alerts and policy checks in pipelines
  5. ReviewUnit cost trends reviewed with product owners
Target: every pound of spend has an owner and a unit cost
Savings leversOrdered by effort; the largest lasting savings usually sit in architecture.
  1. HygieneIdle and orphaned resources, non-production schedules, storage tiersDays, low risk
  2. Right-sizingInstance, node and database sizing from real utilisationWeeks, low risk
  3. CommitmentsSavings plans and reservations on the stable baseline onlyWeeks, financial lock-in
  4. ArchitectureAutoscaling, serverless, data lifecycle, model routing and cachingMonths, largest and lasting

More effort, larger and longer-lasting savings

Key decisions and trade-offs

  • Allocation before optimisation: without ownership, savings cannot be attributed or sustained. Tagging and account structure come first.
  • Unit cost as the headline metric: cost per customer, transaction or AI request shows efficiency even when total spend rises with growth.
  • Commit only to the baseline: cover steady usage with commitments after right-sizing, never before an architecture change.
  • AI cost controls by design: route simple tasks to smaller models, cache repeated prompts and set per-team token budgets.
  • Guardrails in the pipeline: cost estimates on pull requests and policy checks for untagged or oversized resources stop drift at source.

Measures of success

  • Share of spend allocated to an owner
  • Unit cost per customer, transaction or AI request
  • Commitment coverage and utilisation
  • Forecast accuracy against actual spend
  • Delivery metrics unchanged or improved

Common pitfalls

  • One-off clean-ups with no ownership model, so spend returns within months.
  • Buying commitments before right-sizing, which locks in waste.
  • Cost targets imposed without unit metrics, which punishes growth.
  • Ignoring data transfer, logging and AI token costs, often the fastest-growing lines.