Compute commitments are the largest single lever most Australian businesses have on their cloud bill, and the arithmetic is not subtle. A three-year Azure reservation runs about 60 to 65 per cent off pay-as-you-go, and up to 72 per cent with Azure Hybrid Benefit on Windows, while a one-year savings plan sits in the 25 to 35 per cent range. On a $12,000-a-month Azure compute spend, the gap between buying nothing and buying sensibly is roughly $50,000 to $80,000 a year. The question in 2026 is not whether to commit. It is what shape the commitment should take, now that Microsoft has closed off reserved instances for a long list of VM families.
What changed on 1 July 2026
Microsoft's reservation documentation confirms that Azure Reserved VM Instances for select VM series are no longer available for purchase or renewal from 1 July 2026. Existing reservations run out their term, but they cannot be renewed for the affected series.
The list is long and it covers exactly the families most SME estates are built on. From 1 July 2026, Microsoft stopped new purchases and renewals of one-year RIs on Av2, Amv2, Bv1, D, Ds, Dv2, Dsv2, F, Fs, Fsv2, G, Gs, Ls and Lsv2, and stopped both one- and three-year RIs on Dv3, Dsv3, Ev3 and Esv3. If your estate is mostly B-series burstable VMs and Dv3 general purpose machines — which describes a great many Australian SME workloads that lifted and shifted in the 2019 to 2022 window — your reservation renewal path has closed.
That matters more for planning than for panic. A reservation you bought in 2024 for a Dv3 fleet still delivers its discount until it expires. What it means is that the renewal decision, whenever it lands, is no longer "renew the same thing". It is either a savings plan, or a reservation on a current-generation family, which implies a VM migration first.
The discount numbers, side by side
Microsoft's public positioning is up to 72 per cent for Azure Reservations and up to 65 per cent for Savings Plans for Compute against pay-as-you-go. Those are ceiling numbers on specific SKUs in specific regions. The typical ranges are more useful for modelling:
| Instrument | Typical discount off PAYG | Flexibility |
|---|---|---|
| 3-year reserved instance | 60–65% (up to 72% with Hybrid Benefit) | Locked to VM family and region scope |
| 1-year reserved instance | 30–42% | Locked to VM family and region scope |
| 3-year savings plan | 55–65% | Any eligible compute, any region, hourly dollar commitment |
| 1-year savings plan | 25–35% | Any eligible compute, any region, hourly dollar commitment |
Those ranges come from specialist Azure licensing analysis published in April 2026 and updated in July 2026. The pattern is consistent: at the same term, a reservation is worth roughly five to seven percentage points more than a savings plan. That is the price of flexibility, and it is a reasonable price.
The practical reading is that the reservation premium only pays off when you are genuinely confident about the VM family for the full term. Three years is a long time in an environment that is still being right-sized, still absorbing acquisitions, or still moving workloads into PaaS. A five-point discount premium evaporates the first time you strand a reservation on a machine you no longer run.
The AWS side of the same decision
The structure is close enough that a multi-cloud shop can reason about both at once. AWS documents that EC2 Instance Savings Plans deliver savings of up to 72 per cent off On-Demand, comparable to Standard Reserved Instances, with the tradeoff being commitment to an instance family in a region rather than to a specific instance type.
One AWS change worth acting on if you run multiple accounts: Savings Plans and Reserved Instances group sharing became generally available on 19 November 2025, letting commitments be shared across member accounts in AWS Organizations. If you have production, staging and a couple of project accounts each carrying their own under-utilised commitments, that is unused discount you already paid for. Consolidating sharing groups is usually the highest-return AWS billing change available to a mid-sized estate.
Set coverage below peak, not at it
The most common commitment mistake is not choosing the wrong instrument. It is buying too much of the right one.
Cloud compute usage has a floor and a peak. The floor is the compute you will still be running in eighteen months regardless of what happens — domain controllers, the ERP database, the always-on application tier. The peak includes month-end batch, seasonal load, test environments someone forgot to shut down, and the project that finishes in March.
Commitments should cover the floor and nothing above it. A reasonable target is 70 to 80 per cent of steady-state compute under commitment, with the remainder on pay-as-you-go. Uncommitted hours cost list price, which stings, but an unused savings plan commitment costs 100 per cent of list price and returns nothing. The asymmetry is brutal and it argues strongly for under-buying.
Work out the floor from at least three months of actual hourly usage, not from a spreadsheet of what is deployed. Deallocated VMs, autoscale troughs and dev environments that sleep overnight all pull the real floor well below the theoretical one.
The structure that holds up in 2026
For most Australian SME and mid-market estates, the defensible shape is:
- A one- or three-year savings plan as the baseline. Sized to the steady-state floor, term chosen by how stable the estate is. If you are mid-migration, take the one-year and accept the 25 to 35 per cent band. If the estate has been stable for two years and the workloads are not going anywhere, three years at 55 to 65 per cent is real money.
- Selective three-year reservations on top, only for workloads that will not move. Current-generation families only, given the July 2026 retirements. This is where you claim the extra five to seven points, and where Azure Hybrid Benefit pushes Windows workloads toward that 72 per cent figure if you already hold Software Assurance.
- Everything else on demand, deliberately, with a quarterly review of whether the floor has risen enough to justify topping up.
- A migration plan for anything sitting on a retired family, scheduled before the existing reservation expires rather than after.
Getting the sizing right depends on having clean usage telemetry and someone who reads it monthly, which is exactly the kind of ongoing discipline that managing Azure and AWS infrastructure properly is supposed to deliver — commitment coverage is a recurring operational decision, not a purchase you make once and file.
Chasing cheaper regions is not where the money is
There is a persistent idea that Australian regions carry a residency tax and that moving non-regulated workloads offshore unlocks meaningful savings. The pricing does not support it. A cross-region comparison observed in September 2026 puts Azure Australia Central at an average compute price of around US$2.79 per hour, about 2.8 per cent below the global average — an average-tier region, not a premium one.
At the small end the numbers are similarly unremarkable. A 2 vCPU, 8 GB general-purpose VM in Sydney runs roughly AUD $80 to $110 per month on-demand across Azure, AWS and Google, depending on SKU and operating system. Comparable mainstream SKUs do not land on a single shared hourly rate — the listed on-demand prices for an AWS t3.medium and an Azure B2s in the Sydney regions differ by provider, and move again with operating system and whatever discounting you already hold. The spread is narrow enough that provider choice on list price alone is not a strategy.
The same logic applies to the broader "Azure is more expensive than AWS" claim. Analysis of the Australian market published in 2026 argues that Azure is not consistently more expensive once existing enterprise agreement discounts and integration effort are counted. A single well-sized savings plan will move your bill further than a region migration, and it will not bring data residency questions with it.
The first move
Pull twelve months of hourly compute usage from Azure Cost Management or AWS Cost Explorer, chart it, and find the line that usage never drops below. That number, minus a buffer of ten to twenty per cent, is your commitment ceiling. Then list every VM currently running on one of the retired families — Dv3, Dsv3, Ev3, Esv3 and the one-year list above — and note when its existing reservation expires. Those two outputs decide almost everything else.
If you would rather have the answer than build it, Precision IT will model your estate against both instruments, tell you the dollar figure each option saves over one and three years, flag which workloads need migrating off retired families before their reservations lapse, and say plainly if your spend is too small or too volatile for a commitment to be worth buying. Ask for a commitment review and you will get the coverage target, the recommended mix and the migration list, not a brochure.