Most of your 2027 infrastructure bill will be decided in the next 13 weeks, in meetings that look like budget administration.
Most of your 2027 infrastructure bill will be decided in the next 13 weeks, in meetings that look like budget administration. A Savings Plan signed in the second week of November, a storage contract that quietly auto-renews on 1 January, a GPU reservation priced off last spring's rates: none of these feel like strategy when they happen. Together they set the floor for everything you spend next year.
This year the stakes are higher than usual. Unit prices that finance teams have assumed will fall are rising instead. A regulatory deadline that changes the economics of leaving a cloud provider lands two weeks into Q1. And the largest vendor event of the year sits right in the middle of peak contract season. This checklist covers the six decisions worth making, in order, before the year closes.
For a decade, cloud budgeting rested on one comfortable assumption: the price of a unit of compute or storage goes down over time. That assumption broke in 2026. AWS raised EC2 Capacity Block prices by roughly 15 percent on 4 January 2026, lifting the eight-GPU p5e.48xlarge from $34.61 to $39.80 an hour in most regions, according to The Register. Trade press reported a second increase of roughly 20 percent on 1 July. Enterprise SSD contract prices, driven by AI data center demand, were forecast by TrendForce to rise 53 to 58 percent in Q1 2026 alone.
Waste is moving the wrong way too. Flexera's 2026 State of the Cloud Report puts estimated wasted spend on IaaS and PaaS at 29 percent, the first increase after five years of decline, and attributes it to AI workloads and new platform services. Meanwhile the FinOps Foundation's State of FinOps 2026 found that 98 percent of respondents now manage AI spend, up from 31 percent two years earlier. AI is no longer a side budget. It is the part of the budget growing fastest and governed least.
Gartner's July 2026 forecast captures the pressure at market level: IaaS spending is expected to reach $287 billion in 2026, up 29.3 percent, and Gartner warns that technology budgets are being strained by inflation, supply shortages, and rising hardware and memory costs. Put simply, the reader who plans 2027 with 2024 assumptions will overspend before January ends.
| Input | 2026 movement | Source | Budget implication |
|---|---|---|---|
| EC2 Capacity Blocks (GPU) | +15% Jan, about +20% Jul | The Register; trade press | Re-price reserved GPU lines |
| H100 on-demand median | $3.39, up about 12% YoY | GetDeploying, 27 Sep 2026 | Do not assume GPU deflation |
| NAND Flash contract prices | +10% to +15%, Q3 2026 | TrendForce forecast | Pull storage refresh forward |
| IaaS market spend | $287B, up 29.3% | Gartner, Jul 2026 | Expect vendor pricing power |
The first decision is to stop assuming deflation. Our standing thesis at DataStorage.com is that a cheaper token is not a cheaper bill, because falling unit prices drive more usage. In 2026 even the unit is not getting cheaper.
Take the NVIDIA H100, a GPU launched in 2022 that depreciation schedules say should be getting cheaper to rent. GetDeploying's tracker put the median on-demand H100 price at $3.39 per GPU-hour across 44 providers on 27 September 2026, up about 12 percent over 12 months. For an eight-GPU node running all month (730 hours), that is roughly $19,800, or about $2,100 more per month than the same node cost a year earlier. Trackers disagree on the level: the Mercatus GPU Index put the volume-weighted H100 average at $3.84 on 13 September, while Silicon Data's neocloud index sat nearer $2.53. The spread itself is a planning input. Your number depends heavily on which providers you can actually buy from.
Build three scenarios for every line item priced in GPU-hours, gigabytes stored, or tokens: flat, plus 10 percent, and plus 20 percent on unit price, each multiplied by your realistic usage growth. Present the middle case as the plan and the high case as the risk. Finance will push back on the high case. That conversation is far cheaper in October than in March.
Savings Plans, Reserved Instances, and committed use discounts are where most teams lock in next year's spend. They are also where timing mistakes are most expensive, because a one-year or three-year commitment cannot be unwound when a better price appears a week later.
AWS re:Invent runs from 30 November to 4 December 2026 in Las Vegas. Recent editions have brought new instance families, new chips, and pricing changes that alter which commitments make sense. Signing a large compute commitment in mid-November, days before the keynote, means pricing next year against a catalog that may be out of date by December. Hold large new commitments until the announcements are public, then buy in December. Amazon's fiscal year ends on 31 December, and year-end is when account teams have the strongest reason to negotiate.
Coverage matters as much as timing. Flexera's 2026 data shows fewer than half of organizations use a single commitment discount type with any one provider, which means most teams juggle overlapping instruments with different expiry dates. Before buying anything, list every commitment by expiry month. The ones expiring in Q1 2027 are the ones to renegotiate now.
Freeze large new compute commitments from 1 November until re:Invent closes. Renew only what is expiring. Target coverage against your steady baseline, not your peak, and leave AI training capacity on shorter terms, for reasons covered in Decision 5.
The EU Data Act has applied since 12 September 2025, and its cloud switching chapter has a hard date inside your planning window. From 12 January 2027, providers of data processing services may not charge customers any switching charges, and data egress charges incurred when switching are defined as switching charges. Until that date, providers may only charge the direct costs of the switch. The rules apply to any provider serving customers in the EU, wherever the provider is based.
Two details matter for budgeting. First, ordinary operational egress, the traffic of serving data to your users, is not abolished. The Act targets the cost of leaving, not the cost of running. Second, early termination penalties on fixed-term contracts are not switching charges, so a three-year deal signed today can still cost you to exit.
That makes the renewal calendar the real lever. A contract that auto-renews on 1 January for another multi-year term locks in its current terms just before your negotiating position improves. As we discussed with Backblaze CEO Gleb Budman on the DataStorage.com Podcast, egress fees work as the toll booth of the cloud: ingress is free, and leaving is taxed. The Data Act removes the toll for switching. It does not remove the lock-in you sign up for.
Pull every storage, platform, and SaaS contract with a renewal or notice date between now and March 2027. For each, record the auto-renewal clause, the notice period, the exportable data categories, and any early termination penalty. Where you have EU operations, ask the provider for its written switching procedure and run a small export test rather than trusting a documentation page. Where a renewal falls before 12 January, negotiate a short extension rather than a multi-year term.
Storage is the line item most teams treat as a rounding error, and in 2026 it is anything but. TrendForce forecast NAND Flash contract prices to rise another 10 to 15 percent in Q3 2026 after record increases in the first half, and its 21 September update reported that North American cloud providers had raised their Q4 enterprise SSD demand forecasts, giving suppliers more pricing power. Anyone refreshing on-premises flash arrays, or buying storage-dense servers for AI, is buying into a rising market.
Cloud object storage prices have held steadier, but providers absorb rising hardware costs only for so long. More important, storage placement now decides GPU flexibility. Enterprises increasingly run AI work on whichever provider has capacity: AWS one quarter, CoreWeave or Nebius the next. Every move drags data along, and egress on training data can erase the GPU savings that motivated the move. Zero-egress, S3-compatible storage from providers such as Backblaze B2, Wasabi, or Cloudflare R2 is an architecture decision for AI teams, not a procurement footnote.
If an on-premises storage refresh is on the 2027 roadmap, get quotes now and decide whether to pull the purchase into Q4. In the cloud, tier cold data before renewal so the committed volume reflects what actually needs to be hot. For AI datasets, decide where the canonical copy lives before choosing next year's GPU providers, not after.
Conventional advice says to lock in long GPU contracts when prices are rising. That advice fits a stable market, and the GPU market is not one. Newer generations such as the B200 and B300 keep arriving, capacity from new entrants keeps coming online, and prices for older GPUs have held up because demand spilled down the stack. A three-year H100 commitment signed at today's rates could look expensive within 12 months, or cheap. Nobody knows which.
The better Q4 decision is a split portfolio: reserve the steady inference baseline you can forecast with confidence, keep training and experimental capacity on terms of 12 months or less, and keep at least two qualified providers so you can move when prices shift. As Russ Artzt, co-founder of CA Technologies, argued on the DataStorage.com Podcast, many engineers still do not realize neoclouds exist as an alternative to the hyperscalers. Q4 is the right time to qualify one.
Vet those providers properly. As Sunny Smith, founder and CTO of Massed Compute, put it on Episode 7 of our podcast, the first question to ask any provider is: do you own the GPU? Many GPU cloud brands resell capacity they do not operate, and support quality drops with every intermediary.
Split GPU demand into baseline, burst, and training. Commit only the baseline. Benchmark current rates across providers before any negotiation, and ask every shortlisted provider whether it owns its hardware and where your data will sit relative to its GPUs.
The FinOps Foundation reports that the most requested capability among practitioners is granular monitoring of AI spend: tokens, LLM requests, and GPU utilization. Many organizations also report being asked to self-fund AI investment through efficiency gains elsewhere. Both point to the same Q4 decision: AI spend needs owners, tags, and budgets before the new year, not after the first surprise invoice.
The rise in Flexera's waste figure to 29 percent is the cost of skipping this step. AI workloads bill in bursts, often on accounts and API keys set up by product teams outside the FinOps process. By the time the monthly bill arrives, the tokens are spent.
Before 1 January, require every AI API key and GPU project to map to a cost center. Set per-team monthly budgets with alerts at 50, 80, and 100 percent. Choose one unit metric per AI product, such as cost per resolved ticket or cost per thousand requests, and report it alongside total spend from January onward.
The six decisions above map onto a simple calendar. The order matters: the budget baseline informs every commitment, and the contract audit has to finish before re:Invent so that December purchasing is not rushed.
| Window | Owner | Decision | Done when |
|---|---|---|---|
| October | FinOps + Finance | Re-baseline budget; build flat, +10% and +20% scenarios; inventory commitments and contracts by expiry | Three-scenario budget approved |
| November | Procurement + Platform | Audit renewals to March 2027; send non-renewal notices; freeze new commitments until 4 Dec | Renewal register complete |
| December | FinOps + Engineering | Buy commitments against the post-re:Invent catalog; sign short extensions; switch on AI spend tagging | Coverage at baseline, AI keys mapped |
| 1 to 12 January | Legal + Procurement | Confirm EU switching terms and export procedures | Written switching terms on file |
October is for the baseline: re-price every line item on 2026 rates, build the three scenarios, and inventory every commitment and contract by expiry date. November is for the audit and the freeze: finish the renewal review, send non-renewal notices where terms are unfavorable, and hold new commitments until re:Invent closes on 4 December. December is for buying: place commitments against the post-re:Invent catalog, sign short extensions where contracts would otherwise roll into multi-year terms, and switch on AI spend tagging and budgets. The first two weeks of January are for EU contracts: confirm switching terms ahead of 12 January 2027.
Year end is when next year's bill gets written. Most teams simply do not notice they are holding the pen.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
