For most of the cloud era, the FinOps job was to stop you from buying too much. Capacity was assumed to be infinite, instantly available, and yours the moment you clicked. The whole discipline grew up around a single problem: elastic supply meant it was trivially easy to leave money running. In the second quarter of 2026, the assumption underneath that job quietly broke.
Amazon told investors on July 31 that AWS contracted commitments, the revenue sitting in signed contracts that has not yet been recognized, jumped from $364 billion to $496 billion in a single quarter. That is a $132 billion increase in ninety days. Two days earlier, Microsoft closed its fiscal year reporting an $678 billion commercial backlog, up 84% year over year, and CFO Amy Hood told analysts plainly that “demand continues to surpass supply,” with no timeline offered for when that gap closes. Azure grew 43% and crossed $100 billion in annual revenue for the first time. Google, for its part, raised its 2026 capital spending guidance to between $195 and $205 billion, up from $180 to $190 billion, and cited accelerated capacity delivery to meet demand it cannot currently serve.
Read those numbers together and a different picture emerges than the one FinOps was built around. The constraint is no longer your budget. The constraint is the provider’s ability to hand you a machine.
What the earnings actually said, in plain terms
Strip out the investor framing and three facts stand out from the late-July hyperscaler results.
First, growth is reaccelerating, not cooling. AWS grew 37% year over year, its fastest rate in eighteen quarters. Azure reaccelerated to 43% from 40% the prior quarter. This is not a market where providers are hungry for your commitment and willing to deal.
Second, the backlogs are historic and they are growing faster than revenue. A combined figure well north of a trillion dollars in contracted-but-undelivered cloud sits on the books at Amazon and Microsoft alone. When backlog grows faster than the ability to fulfill it, the provider is effectively running a waitlist.
Third, the money going into capacity is enormous and still not enough. Cloud providers are on track to spend close to $600 billion on capital projects in 2026 by conservative counts. Microsoft added roughly a gigawatt of capacity in a single quarter, stood up 88 new datacenters across the fiscal year, and doubled its total capacity in two years, and still describes itself as capacity constrained. Amy Hood said the new capacity was “immediately monetized.” That is the tell. When everything you build sells the instant it lights up, you are not in an oversupplied market anymore.
FinOps grew up in an oversupplied market
The core FinOps motions all assume abundance. Rightsizing assumes you can shrink an instance today and grow it back tomorrow. Autoscaling assumes the capacity you scale into is simply there. Spot and preemptible strategies assume a deep pool of spare compute the provider is desperate to sell at a discount. Even the reserved-versus-on-demand decision, the oldest commitment question in the book, was framed almost entirely as a discount tradeoff: give up flexibility, get a lower rate.
That framing held because for fifteen years it was true. The FinOps Foundation’s 2026 State of FinOps report still puts waste reduction and rate optimization near the center of the practice, and rightly so. Around $44.5 billion in cloud spend is wasted annually, and 98% of FinOps teams now manage AI spend, up from 31% just two years ago. The discipline is doing more, across more of the technology estate, than ever.
But the reports describe a demand-side world. The new earnings describe a supply-side one. When I was running IT operations two decades ago, before public cloud was a serious option for most workloads, capacity was something you negotiated and waited for. You forecast demand a year out, you signed for hardware, and if you guessed low you did not get a bigger box next week; you got a project delay. Cloud erased that discipline because it erased the scarcity. What the Q2 2026 results suggest is that for the workloads everyone actually wants right now, GPU compute and the power to run it, the scarcity is back.
A commitment now does two jobs at once
Here is the practical shift, and it is the thing most teams have not repriced.
A reserved instance or a savings plan used to buy you one thing: a lower rate in exchange for a spending promise. In a constrained market, a capacity reservation buys you a second thing that is arguably more valuable: a guarantee that the hardware will be there when you ask for it. AWS makes this explicit with its GPU offerings. EC2 Capacity Blocks exist precisely because on-demand GPU capacity cannot be assumed to be available, and AWS raised Capacity Block prices roughly 20% on July 1, following a 15% increase in January. Two hikes in six months on the same reservation product is not a discount signal. It is a scarcity signal. You are no longer paying a premium for convenience; you are paying to hold a place in line.
That changes the math on commitments in a way that the standard break-even calculator does not capture. The traditional model asks: will I use enough of this instance over the term to justify locking in the rate? The scarcity model adds a second question: if I do not lock this in, will the capacity even be available when my workload needs it, and what does an outage or a project stall cost me? For a training run on a deadline or an inference workload tied to revenue, the answer to that second question can dwarf the rate savings. The commitment is no longer just a cost decision. It is a supply-chain decision that happens to have a cost attached.
I have started framing this for the finance leaders I work with as the difference between buying insurance and buying at a discount. A reserved GPU commitment in 2026 is closer to a supply contract with a manufacturer than to a coupon. You would not evaluate a supply contract purely on unit price, and you should not evaluate a capacity commitment that way either.
What actually changes in the practice
None of this means abandon the discipline. It means adding a layer to it.
Treat capacity availability as a forecast input, not a given. When you build a cost forecast for an AI-heavy roadmap, model the risk that the capacity you need is unavailable at on-demand rates, or unavailable at all in your preferred region. The forecast that assumes instant elasticity is now the optimistic case, not the base case.
Reprice commitments to include the value of guaranteed access. The break-even analysis you use for reserved instances and savings plans needs a second column for critical workloads: what does it cost the business if this capacity is not there? For a batch job that can wait, that value is near zero and the old rate-only math still applies. For a revenue-tied workload, it can justify committing earlier and longer than a pure discount calculation would.
Move the commitment conversation upstream. In an oversupplied market, you could optimize commitments quarterly with data from the past. In a constrained one, the capacity decision has to sit alongside the capacity planning decision, which means FinOps needs to be in the room when the AI roadmap is set, not reconciling it after the fact. This is the “shift up” the FinOps Foundation has been describing, and the supply crunch makes it urgent rather than aspirational.
Bring the negotiation forward and widen it. When capacity is the scarce good, your leverage in a cloud contract negotiation shifts from rate to allocation. The question stops being only “what discount can I get” and becomes “can you guarantee me this much capacity in this region by this date, in writing.” Providers running a waitlist will not volunteer that guarantee, but the largest committed customers can still ask for it.
The part that has not changed
It would be a mistake to read all of this as license to over-commit. The oldest FinOps failure, buying capacity you do not use, is still a failure, and a constrained market does not forgive it; it just changes the direction of the risk. Over-provisioning GPU reservations you cannot fill is now expensive in two ways: you pay for the idle capacity, and you have removed it from the pool while paying a scarcity premium to hold it.
The waste story is also not going anywhere. The same memory and hardware cost pressures pushing prices up make idle resources more expensive per hour than they were a year ago, which means the return on basic hygiene, killing zombie instances, rightsizing overallocated workloads, cleaning up orphaned storage, is higher now, not lower. Scarcity raises the price of every mistake in both directions.
What changed in the second quarter of 2026 is not the fundamentals of the discipline. It is the assumption sitting underneath it. Cloud is no longer a bottomless buffet where the only sin is taking too much. For the workloads that matter most, it is a constrained resource with a line out the door, and the FinOps teams that adjust their commitment strategy for scarcity, rather than optimizing for an abundance that no longer exists, will be the ones whose roadmaps survive contact with the provider’s waitlist.
If you have not revisited your commitment model since these earnings landed, that is the first place to look. The break-even math you ran last year was solving the wrong problem.