Commitment coverage on the steady state baseline was worth 14 of the 31 points cut from the run rate, which is more than rightsizing delivered
The standard advice is to rightsize first and commit later. Microsoft prices on demand at a premium precisely because most buyers never commit at all.
Prepared by Redress Compliance · August 18, 2026 · Azure cost reviews. 30 to 40 reviews led, 2024 to 2025.
Executive summary
Reserved and savings plan coverage was the single largest lever, worth 14 points of a 31 point reduction. Reservation gaps left 40 to 60 percent of steady state compute on pay as you go rates.
Rightsizing the top 25 virtual machines delivered more savings than rightsizing the next 1,200 combined, because those 25 accounted for 41 percent of the entire compute bill.
Between 25 and 45 percent of cost was unattributable at the start. Tagging and showback fixed the behavior loop, and teams began deleting their own waste once they could see it.
The renewal discussion improved rather than worsened. Discount bands got better once the vendor saw the new burn rate, which is the opposite of what most buyers fear.
Why does commitment come before rightsizing?
Because the reservation gap is bigger and faster than the sizing gap. In roughly 8 of 10 estates reviewed, coverage on the steady state baseline was the larger win, applied before any rightsizing at all.
On demand pricing for steady production workloads runs roughly 60 percent above a three year reservation, priced on the Azure pricing pages. That gap funds nothing strategic.
Idle cleanup is real and secondary
It is the visible half of the problem and the smaller one. Locking coverage on the proven baseline first, then rightsizing the variable layer, is the sequence that produced the result.
What did the first 30 days find?
Coverage below 35 percent, twenty five oversized machines, years of premium snapshots, and egress billing for a feature that had been deprecated. None of it was hidden. None of it was owned.
- Almost two thirds of compute hours were running at on demand rates.
- The database tier had been sized for a 2022 launch peak, and actual peak processor use never crossed 28 percent.
- Nightly backups had created 38 terabytes of cold snapshots on the most expensive storage tier, with no retention policy ever set.
- Cross region replication kept billing for a feature that shipped in 2023 and was then withdrawn.
| Category | Before, monthly | After, monthly | Change |
|---|---|---|---|
| Compute | $418,000 | $262,000 | Down 37 percent |
| Database | $104,000 | $71,000 | Down 32 percent |
| Storage | $87,000 | $36,000 | Down 59 percent |
| Networking | $48,000 | $29,000 | Down 40 percent |
| Other services | $43,000 | $38,000 | Down 12 percent |
| Total infrastructure | $700,000 | $436,000 | Down 38 percent |
The Azure cost optimization playbook
Reduce, reserve and rightsize, in the order that actually produces the saving.
Get the brief →What 30 to 40 Azure cost reviews showed
Across roughly 30 to 40 Azure cost reviews led between 2024 and 2025, the largest savings came from commitment and SKU choices rather than from turning resources off. Three patterns recur.
- Reservation gaps: 40 to 60 percent of steady state compute ran on pay as you go that qualified for a reservation or a savings plan.
- Sizing gaps: 20 to 35 percent of virtual machines were one to two sizes larger than utilization justified.
- Untagged spend: 25 to 45 percent of cost was unattributable at the start.
The Azure bill almost never fails for one reason. It fails for ten small reasons that add up to a quarter of the total.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call, so the commitment is tested before it is signed
- A negotiation playbook, talking points, and a two page executive brief on day one
Where is the compute bill actually concentrated?
In very few machines. Twenty five of them accounted for 41 percent of the entire compute bill, and rightsizing those alone produced more savings than touching the remaining estate.
The storage story is the same shape. Moving 38 terabytes of cold snapshots off the premium tier to cool blob storage, roughly forty times cheaper for the same data, saved 41 thousand dollars a month with no operational impact.
The renewal that prices this consumption is a separate exercise, mapped across the 2027 agreement series and the Azure best practices guide.
Start where the money is, not where the count is
An estate wide sweep feels thorough and spends the effort evenly across resources that do not carry the bill. Rank by spend before ranking by anything else.
Watch the briefing · 4:04Azure and the MACC: Where the Leverage Actually IsSizing the commitment on your own consumption rather than the vendor forecast.
What made the savings hold?
Tagging and showback, which fixed the behavior loop. Engineering teams started deleting their own waste once they could see what it cost them rather than what it cost the company.
A ratio target, not a dollar target
Spend that nobody owns is invisible in Azure Cost Management until it is tagged, which is why the first thirty days are a tagging exercise rather than a cutting one.
The target was set as a ratio rather than a dollar figure: infrastructure cost back to 9 percent of revenue, down from 14.2 percent, with no slowdown to the roadmap and no headcount reduction. Visible movement inside 90 days, the new run rate locked by month nine.
The commercial side improved with it. Discount bands got better once the vendor saw the new burn rate, and the mechanics of that conversation sit in the agreement renewal guide, the renewal playbook and its companion paper.
Which levers move that band, and how the seat mix and assistant licensing sit alongside the consumption line, are worked in the discount levers guide, the SKU comparison and the assistant licensing guide.
What the reviews measured, 2024 to 2025
Two cuts of the engagement file explain why the sequence matters more than the effort.
Workloads that qualified for a reservation or a savings plan and were left at on demand rates anyway.
Spend with no owner and no cost center, which is why the behavior loop could not close before tagging.
Neither is an engineering failure. Both are governance gaps that a standing process closes and a one time cleanup does not.
Your first five moves
- Tag everything to an owner and a cost center first, because 25 to 45 percent of cost was unattributable at the start and nothing else can be measured until it is.
- Lock commitment coverage on the proven steady state baseline, the single largest lever and worth 14 of the 31 points, since 40 to 60 percent of it was running at on demand rates.
- Rank the estate by spend and rightsize the top machines before the rest, because 25 of them carried 41 percent of the compute bill.
- Set a retention policy on backups and move cold data off the premium tier, worth 41 thousand dollars a month on 38 terabytes with no operational impact.
- Stand up showback so teams see their own spend, and set the target as a ratio rather than a dollar figure. The Microsoft practice runs the commitment model before the renewal, and the Azure playbook carries the operating model.
Frequently asked questions
Should you rightsize or commit first?
Commit first, on the proven steady state baseline, then rightsize the variable layer. In roughly 8 of 10 estates reviewed, coverage was the bigger and faster win.
How large is the reservation gap?
Between 40 and 60 percent of steady state compute ran on pay as you go that qualified for a reservation or a savings plan. On demand runs roughly 60 percent above a three year reservation.
How much did commitment coverage deliver?
Fourteen points of a thirty one point reduction on the estate worked in detail, which was more than rightsizing produced across the whole rest of the footprint.
Where is the compute bill concentrated?
In very few machines. Twenty five of them carried 41 percent of the entire compute bill, so rightsizing those alone beat rightsizing the next 1,200 combined.
What does storage tiering save?
On this estate, 41 thousand dollars a month. Thirty eight terabytes of cold backup snapshots moved off the premium tier to cool blob storage, roughly forty times cheaper for the same data.
Why is tagging the first step?
Because 25 to 45 percent of cost was unattributable at the start. Nothing can be measured or owned until spend maps to an owner and a cost center.
Does showback actually work?
Yes. Engineering teams started deleting their own waste once they could see what it cost them, which is the behavior loop a one time audit cannot close.
Does optimizing hurt the next renewal?
It did not here. Discount bands improved once the vendor saw the new burn rate, which is the opposite of what most buyers expect when they cut consumption before a renewal.
What should the target be?
A ratio rather than a dollar figure. This estate targeted infrastructure at 9 percent of revenue, down from 14.2 percent, with no roadmap slowdown and no headcount reduction.
How long does it take?
Visible movement inside 90 days and the new run rate locked by month nine, on a nine month program combining a usage audit, a discount mechanics rework, and an operating model the company runs itself afterwards.
