Reserved Instances Aren't Being Killed. They're Being Disassembled.

Part three. The first piece was RIP RIs, where I argued AWS was quietly killing Reserved Instances. The second was Cloud Vendors Finally Agreed on Commitments. AI Vendors Didn't., where I showed the three clouds converging on spend-based commitments. This one is about what happened last month- and why I've stopped expecting a funeral.

Nobody is going to announce that Reserved Instances are dead.

That's the thing I got wrong in the first piece. I was watching for an obituary, a deprecation notice, an end-of-life date, a migration deadline. What's actually happening is slower and much harder to notice: RIs are being taken apart one attribute at a time.

On July 22, Microsoft published a documentation page that most people will never read. Starting February 1, 2027, reservations purchased after that date can't be exchanged if the underlying service is covered by savings plans- Azure Virtual Machines, App Service, SQL Database, and similar. Reservations bought before that date keep the right to one final exchange.

That's it. No retirement. No end of life. You can still buy an Azure Reserved Instance on February 2, 2027. It will still work. It just won't have an exit.

1. What was actually removed

Exchange is the feature nobody talks about until they need it.

You commit to three years of a VM family. Fourteen months in, the workload moves to a different family, or a different region, or the architecture changes entirely. Exchange let you swap the reservation for a different one without a penalty. Your term reset, your money stayed committed, and the mistake cost you nothing.

Strip that away and look at what's left when a commitment goes wrong:

• Cancel it. Azure caps cancellations at $50,000 in a rolling 12-month window per billing profile or enrollment. For an enterprise commitment book, that cap is a rounding error. It also didn't move in this update.
• Trade it in for a savings plan. Always available, unchanged, and not coincidentally- exactly where Microsoft would like you to go.

There's also a detail in the mechanics worth flagging, because it surprises people. Microsoft processes an exchange as a cancellation, a refund, and a new purchase. So if you use your one final exchange after February 1, 2027, the reservation you receive is governed by the new terms and isn't exchangeable. The final exchange is real. It's also terminal.

Instance size flexibility is untouched. The cancellation policy is untouched. Reservations for services savings plans don't cover — Azure VMware Solution is Microsoft's own example- are untouched.

One capability was removed. Nothing else.

2. The honest version of Microsoft's argument

I want to be fair here, because the reasoning isn't unreasonable.

Microsoft's position is that exchange existed to give reservations flexibility, savings plans now provide that flexibility natively, and running both creates a confusing framework. Choose reservations for predictable, stable workloads. Choose savings plans for dynamic ones. Clean lines.

That's a defensible product argument. It's also not the whole trade.

Reservations still carry a deeper discount than savings plans for genuinely stable workloads. That gap is the entire reason anyone still buys them. What exchange did was let you take the deeper discount and keep a way out if you turned out to be wrong about "stable."

In other words: exchange was a free option attached to every reservation. Microsoft just stopped writing free options.

If you know with real confidence which VM families you'll be running in 2029, nothing changes for you- you were never going to exchange anyway. If you don't, the discount you're buying now comes with a risk you used to be able to hand back. Same price. Different product.

3. The pattern, in order

Put the last several years in a single column and the shape is obvious.

Not one of those is a deprecation of Reserved Instances. Every one of them takes something away from Reserved Instances.

First the new instance families stopped getting them. Then the old families lost them. Now the flexibility is going. What survives is a product that still appears on the price list, still has a purchase button, and has progressively fewer reasons to be chosen.

This is how mature products get retired inside large companies. Not with an announcement. With a documentation update every few quarters, each one small enough that nobody writes about it.

4. What it means if you manage a commitment book

Three consequences, in order of how soon they'll hit you.

Forecasting stopped being optional. When exchange existed, commitment planning tolerated error- buy, observe, adjust. From February 2027, the first decision is the only decision for anything you buy after that date. The skill that mattered was rebalancing. The skill that matters now is being right the first time, which means your utilization and drift data has to be good before you sign, not after the invoice lands.

You have an 18-month window, and it isn't a free one. Every reservation you already hold carries exactly one exchange. That's a real, expiring asset, and worth an explicit audit rather than a vague intention. Which reservations are drifting away from the workloads they were bought for? Which ones would you re-cut today if it cost nothing? Do that math on purpose, before the option decays into a decision you make under pressure. Just remember that spending the final exchange after February 1 leaves you holding something you can't exchange again.

The RI-versus-savings-plan choice became a genuine commitment decision. It used to be a discount-rate comparison with a safety net under it. Now it's the discount rate against your actual confidence in a three-year forecast. Those are different conversations, and the second one belongs in front of finance, not buried in a procurement checklist.

Here's the piece I keep coming back to, and it connects this to the AI half of the bill: in the May breakdown of AI commitment products- Bedrock Provisioned Throughput, Azure OpenAI PTUs, Vertex AI GSUs, Priority Tier, Guaranteed Capacity, not one of them has an exchange mechanism. No swaps, no re-cuts, no take-backs. The flexibility that cloud commitments spent close to a decade building in is not being built into AI commitments at all. And it's now being removed from cloud reservations.

The direction of travel is the same on both halves of your bill. Commit harder, adjust less.

That's a solvable problem, but only if you can see utilization against every commitment in one place while there's still time to act on it. It's most of why we built commitment visibility into Finout the way we did- the value isn't in the report at the end of the month, it's in knowing you're drifting while the drift is still cheap.

5. So, are RIs dead?

No. And I don't think they'll ever be declared dead.

They'll keep narrowing. Fewer eligible families. Fewer eligible services. Less flexibility. Until the honest answer to "should we buy a reservation?" is "only if this workload genuinely will not change for three years"- which, for most of what's running in your account right now, it will.

The right question isn't whether Reserved Instances are dying.

It's whether you'd still buy the ones you're holding, knowing you can't change your mind. You have until February.