Somewhere in your organisation there is an AI business case with a 12-month payback line in it. It was probably written in 2025, it probably justified a GPU reservation or a multi-year platform commitment, and it is probably wrong.
TEKsystems surveyed technology and business decision-makers for its State of Digital Transformation 2026 report, published in February. The share expecting a return on their transformation investment within six months fell from 42 percent in 2025 to 27 percent in 2026. In the same survey, 71 percent said they plan to increase AI spending this year. Expectations of a fast payback are collapsing while the spending keeps climbing. That combination is the AI ROI reset, and it lands squarely on infrastructure budgets.
TEKsystems is not an outlier. Every serious survey published in 2026 tells the same story from a different angle.
Deloitte set the baseline in October 2025 with a survey of 1,854 executives across Europe and the Middle East. Most respondents said a typical AI use case takes two to four years to reach satisfactory ROI, against the seven to 12 months they expect from conventional technology. Only 6 percent reported payback inside a year, and even among the most successful projects just 13 percent saw returns within 12 months.
PwC’s 29th Global CEO Survey in January 2026 found that 56 percent of CEOs had seen neither revenue growth nor cost reduction from AI over the previous 12 months. Only 12 percent reported both. PwC’s April 2026 Executive Views survey of 633 US executives then found 81 percent believe they are at least a year away from meaningful returns beyond basic efficiency.
McKinsey’s State of AI 2026, published on 25 August from a survey of 1,719 respondents in 97 countries, is the starkest. The share attributing any EBIT impact to AI is 37 percent, essentially unchanged from 2025, even though the share scaling AI across the enterprise rose from 38 to 44 percent. AI high performers, those attributing 5 percent or more of EBIT to AI, stayed flat at about 6 percent. Adoption went up. Financial impact did not move.
Gallagher’s 2026 AI Adoption and Risk Survey puts a number on the new expectation: organisations measuring AI ROI now anticipate an average of 28 months before the value outweighs the upfront cost.
The easy explanation is that AI underdelivered. The more useful explanation for an infrastructure buyer is that costs scaled faster and earlier than benefits, and most business cases modelled the opposite.
BCG’s Applied AI Index 2026, released on 30 September from a survey of 1,330 senior leaders, found corporate AI spending rose from about 1.7 percent of revenue in late 2025 to 3.3 percent now. More than 80 percent of that spend now sits outside the enterprise IT budget. When most AI cost lands in business-unit budgets, nobody owns the full bill, and payback gets measured against a fraction of the real cost.
McKinsey found that about one in five organisations is already limiting AI use because of operating costs, including tokens. This is the pattern we have tracked all year: a cheaper token is not a cheaper bill. Per-unit prices fall, usage rises faster, and agentic workloads consume far more compute per task than chat. The cost line in a 2025 business case assumed today’s unit price multiplied by yesterday’s usage.
Gartner has predicted that through 2026, organisations will abandon 60 percent of AI projects unsupported by AI-ready data. Getting data AI-ready means moving it, cleaning it, tiering it and placing it next to compute. Those are storage and network costs, and they hit in the first six months, long before a model earns anything. Data gravity, being landlocked by where your data already lives, is the constraint almost no business case priced.
This is not a bubble-bursting story, and treating it as one would be the wrong decision. BCG’s headline finding is that nearly half of companies now capture meaningful value from AI: 7.5 percent it classifies as future-built and a further 41 percent that are actively scaling and outperforming laggards. McKinsey found that 80 percent of respondents say AI has improved their individual productivity.
The read for buyers is precise. Value exists, but it compounds over two to four years, not two to four quarters. Our guest on Episode 7 of the DataStorage.com Podcast, Sunny Smith of Massed Compute, argued that enterprise production deployment has barely started, held back by talent scarcity and production-readiness cycles. A longer payback curve is exactly what that thesis predicts. The mistake is not investing in AI. The mistake is signing infrastructure commitments that assume the value arrives on the old timeline.
Be careful with the optimistic counter-data too. Several of the fastest payback claims in circulation come from parties selling the product being measured: Google Cloud has cited IDC research putting payback on its AI products at roughly eight months, and Extreme Networks reported that 57 percent of executives expect returns within weeks. Both may be true for the narrow use cases they measured. Neither is an independent benchmark for an enterprise AI programme, and neither should anchor your contract terms.
Here is where the reset turns into a procurement problem. Consider a mid-size inference deployment with a business case approved on a 12-month payback. The illustrative model below uses round assumptions so the mechanics are clear: 16 H100-class GPUs on a three-year reservation at an assumed blended $2.20 per GPU-hour, plus 400 TB of training and retrieval data on hyperscaler object storage at an assumed $21 per TB per month. Live GPU rates move weekly, so check current pricing before plugging in your own numbers.
That is $308,352 a year in compute and $100,800 a year in storage, or $409,152 in annual infrastructure spend before a single engineer is paid. The table shows how much of that committed spend is gone by the time the project pays back under each payback scenario.
| Payback reached at | Cumulative infra spend | 3-year reservation used | Verdict |
|---|---|---|---|
| 12 months | $409,152 | 33% | Earns on proven workload |
| 24 months | $818,304 | 67% | Mostly spent before value |
| 36 months | $1,227,456 | 100% | Fully spent pre break-even |
| 48 months | $1,636,608 | 100% + renewal | Renewed before payback |
At the 12-month payback the business case promised, one third of the reservation is consumed and two thirds of it is earning on a proven workload. At the two-to-four-year payback Deloitte says is typical, the entire reservation is spent before the investment breaks even, and at 48 months you have renewed the contract on faith. The reservation discount that looked like prudence becomes a three-year bet on a timeline the data no longer supports.
There is a second cost that only appears when you try to change course. If the project stalls and you want to move that 400 TB to a cheaper GPU provider or a cheaper storage tier, egress at an assumed $0.05 per GB is a $20,000 toll just to leave. Egress fees are the toll booth of the cloud, and a longer payback window means more months in which you might need to pay it. Zero-egress storage such as Backblaze B2, Wasabi or Cloudflare R2 turns that exit cost into close to zero, which is worth more when your payback horizon is uncertain.
The goal is not to slow down. It is to make sure infrastructure commitments match the payback curve the evidence supports.
Present three scenarios to finance and fund the project against the middle one. If the case only works at 12 months, it does not work. Deloitte’s data says 6 percent of organisations hit that window.
Run the first 6 to 12 months on on-demand or short-term GPU capacity, even at a premium. Convert to a reservation only once utilisation and value are measured. The supply balance in GPU markets has shifted toward buyers in 2026, which makes short contracts and multi-provider strategies more viable than they were a year ago.
Price the full data path: storage at each tier, retrieval, inter-region transfer and the cost of exiting. Keep training data and retrieval corpora on storage that does not charge you to move it, so a provider switch is a decision rather than a penalty.
With BCG finding more than 80 percent of AI spend outside IT, FinOps teams need tagging and showback that capture tokens, GPU hours, storage and SaaS AI features across business units. Payback measured against a partial bill is not payback.
Define the metric that proves value before any spend is committed, then review it at month 9. That gives you time to decide on reservation conversion before the standard 12-month renewal cycle forces the decision.
AI is not failing to pay back. It is paying back on a timeline your infrastructure contracts were never written for.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
