A budget built on last year's ratios will be wrong on GPU, storage, and egress, in different directions, unless you write down defensible assumptions for each instead of a flat inflation factor.
Most infrastructure teams build next year's cloud budget by taking this year's line items and applying a flat inflation factor: GPU up a bit, storage up a bit, egress unchanged because nobody ever changes egress. That method worked when the market was stable. It will not work for FY2027, because three things are moving at once and none of them move together.
GPU capacity is loosening in some places and tightening in others. Meta Compute entering the market as a fourth hyperscaler, and both Meta and SpaceX selling surplus GPU capacity, signal that the era of unconditional GPU scarcity is ending for buyers who can negotiate. At the same time, NVMe storage, the component that actually keeps a GPU fed, is in the opposite position: prices have moved sharply higher and multiple infrastructure operators expect the shortage to persist into 2027. Meanwhile egress, the fee nobody budgets for until the invoice arrives, keeps compounding as AI workloads move larger datasets between providers chasing GPU availability.
A budget built on last year's ratios will be wrong on all three lines, in different directions. This guide gives you defensible assumptions for each, and a framework for writing the number down instead of guessing at it.
Start with the structural shift. CoreWeave and Nebius, two of the largest independent neoclouds, both expanded their Meta contracts in 2026, CoreWeave's deal reportedly reaching roughly $21 billion and Nebius's roughly $27 billion. Those numbers matter less as headline totals than as a signal: the hyperscaler with the deepest pockets in AI is now also a capacity seller through Meta Compute, not just a buyer. That changes the negotiating position of everyone downstream.
For most of 2024 and 2025, the safe move was to lock in long-term GPU commitments before capacity disappeared. Going into 2027, that logic reverses for buyers with flexible workloads. With more sellers of surplus capacity entering the market, 12-month or even quarterly commitments preserve the ability to renegotiate as pricing moves in your favor. Budget planners should model a mix, not a single rate: a baseline reserved tranche for steady-state inference, and a spot or short-term tranche sized to actual peak usage rather than worst-case usage.
Not every GPU brand quoting you a rate owns the hardware behind it. Many neocloud brands are brokers reselling capacity they do not control, which affects support quality and, more relevantly for budgeting, price stability: a broker's rate moves when its underlying supplier's rate moves, often without warning. Before locking a 2027 number against any provider, confirm whether they own the fleet. A B300 class server runs in the neighborhood of $750,000, which is why true ownership is a financial commitment few players actually make, and why the "do you own the GPU" question is now a budgeting question, not just a technical one.
For workload level planning, model training and inference separately rather than blending them into one GPU line. Training spend is lumpy and project driven; inference spend scales with product usage and is the line most likely to be underbudgeted, because a cheaper per-token rate drives more usage rather than a lower bill. Teams evaluating current rates across providers before locking assumptions can pull live numbers from the GPU Price Explorer rather than budgeting off a quote that is already a quarter old.
Storage is the line item most likely to blow past its 2026 baseline, and it is the one finance teams are least prepared for, because storage has historically been the boring, stable part of the infrastructure bill.
High performance NVMe, the tier that sits adjacent to GPU clusters and actually determines whether a training run is compute bound or storage bound, has seen prices roughly triple according to infrastructure operators tracking the shortage, with scarcity expected to persist until the end of 2027. That is not a line you can budget flat. If your architecture depends on high throughput local storage for GPU adjacent workloads, model a step change here, not a percentage nudge, and revisit the assumption quarterly rather than annually.
With premium NVMe expensive and scarce, the cost of sloppy data placement goes up. Data that does not need to sit on the fastest, most expensive tier next to a GPU cluster should not be there. Budget planners should push storage architecture reviews earlier in the FY2027 planning cycle than usual, specifically to identify what can move to warm or cold tiers before the NVMe line is finalized, not after.
This is also where the GPU and storage conversation intersect directly. Enterprises shift GPU providers as availability changes, from AWS to Azure to CoreWeave and back, and every one of those moves drags data along behind it. That is a cost planners frequently leave out of the GPU line entirely and it belongs in the storage or egress line instead.
Ingress into a cloud provider is free. Getting your own data back out is not, and that asymmetry is deliberate. Hyperscalers spent a decade centralizing customer data specifically because the exit fee makes leaving expensive. For a budget planner, egress is the line most likely to be missing entirely from a first draft FY2027 number, because it is invisible until a migration or a multi-provider GPU strategy triggers it.
If your GPU plan for 2027 includes moving workloads between providers to chase better rates, and the analysis above suggests more teams will do exactly that, your egress line needs to scale with that plan. A single 500 TB egress event can run into the tens of thousands of dollars on a standard hyperscaler contract. Multiply that by however many provider switches your GPU strategy assumes, and egress stops being a rounding error.
Zero egress and low egress storage providers exist specifically to remove this variable. Backblaze B2 and Wasabi both offer S3 compatible storage with no or minimal egress fees, which makes them worth evaluating as the storage layer underneath a multi-provider GPU strategy even if you never move your primary compute off a hyperscaler. The savings do not show up as a storage line item improvement; they show up as an egress line item that never gets big enough to notice.
Teams sizing this tradeoff for FY2027 should run actual numbers rather than estimate them. The Cloud Cost Calculator compares real storage and egress costs across providers side by side, which turns a directional assumption into a specific number finance can put in a spreadsheet.
Put the three lines together and the FY2027 budget conversation looks different from FY2026's. Here is the sequence that produces a defensible number instead of a guess.
| Budget Line | 2026 Working Assumption | 2027 Recommended Assumption |
|---|---|---|
| GPU (reserved) | Long-term lock-in to guarantee availability | Shorter commitments; renegotiate as surplus capacity enters market |
| GPU (burst/spot) | Pay premium for any available capacity | Model against multiple sellers; expect more favorable spot pricing |
| NVMe / GPU-adjacent storage | Flat percentage increase over prior year | Step change; scarcity assumption held through full year |
| Egress | Flat historical average, often omitted | Explicit line tied to planned GPU provider switching |
The teams that get this right treat GPU, storage, and egress as one connected budget conversation instead of three separate line items reviewed by three separate owners. The provider switch that saves money on the GPU line is often the same event that spikes the egress line, and the storage tier that looks cheap in isolation is often the one creating the GPU adjacent bottleneck that shows up as wasted compute time. Building the FY2027 number with that connection in mind, rather than discovering it mid-year, is the actual deliverable of this planning cycle.
The FY2027 number is not last year's number plus inflation. It is three moving assumptions that finally intersect, and the budget that treats them as connected will be the one that survives the year without a mid-cycle correction.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
