From Shubhi K | Product & Market Analysis

AI Infrastructure ROI: Gartner's 28% Counts IT Operations, Not GPU Buildouts

On this page

Only 28% of AI use cases in infrastructure and operations fully succeed and meet ROI expectations, according to a Gartner survey of 782 I&O leaders. Much of the coverage called those use cases "AI infrastructure projects". They are not GPU clusters. They are AI tools applied to running IT, such as auto-remediation and service desk automation. The GPU side of AI infrastructure ROI has its own, worse, public numbers.

Key takeaways

  • Gartner's 28% measures AI used inside IT operations, not GPU buildouts. The survey of 782 I&O leaders, fielded in November and December 2025, counts use cases such as auto-remediation and IT service management.
  • The real story is the 52% that land in between. With 28% fully meeting ROI and 20% failing outright, roughly half of AI use cases in I&O deliver partial value that nobody has defined.
  • GPU infrastructure has a separate and larger ROI problem. Cast AI measured average GPU utilization of 5% across the Kubernetes clusters it analysed, which means about 20 paid GPU-hours for every hour of useful work.
  • Cloud commitments turn idle GPUs into a fixed cost. Savings Plans bill a committed dollar amount per hour for 1 or 3 years, and AWS cut H100 instance prices by up to 45% in a single announcement in June 2025.

What does Gartner's 28% AI infrastructure ROI figure mean? It means that among AI use cases deployed inside infrastructure and operations teams, 28% fully succeeded and met ROI expectations, 20% failed outright, and the rest delivered partial value. It describes AI tools that run IT, not spending on GPUs or data centres, and Gartner has not published a public definition of "success".

28%Share of AI use cases in I&O that fully succeed and meet ROI expectations. Source: Gartner, April 2026.
5%Average GPU utilization across Kubernetes clusters analysed by Cast AI. Vendor telemetry, directional. Source: Cast AI, April 2026.
29%Estimated wasted cloud spend, the first increase in five years. Self-reported. Source: Flexera, March 2026.

What Gartner's AI infrastructure survey actually measured

The figure comes from a Gartner release dated 7 April 2026 titled "AI Projects in I&O Stall Ahead of Meaningful ROI Returns". The underlying survey covered 782 infrastructure and operations leaders and was fielded in November and December 2025, per The Register's report on the release. I&O is the part of IT that keeps servers, networks, cloud accounts and service desks running.

The headline finding, in The Register's wording, is that "only 28 percent of use cases fully succeed and offer return on investment". A further 20% fail outright. Separately, 57% of respondents said they had suffered at least one failure applying AI to their area. Those are three different units: use cases, use cases, and people.

Use cases, not projects

The 28% and the 20% describe use cases. The 57% describes respondents. Mixing them produces sentences like "most I&O leaders fail at AI", which the data does not say. A leader who ran four use cases with one failure counts once in the 57% and four times in the use case split.

This matters for your own reporting. If your board asks how your team compares with the 28%, count use cases with a written ROI target, not initiatives, vendors or teams. Otherwise you are comparing a different unit against Gartner's number.

The definition of success is not public

Neither the release coverage nor the trade press reports the question wording behind "fully succeed and meet ROI expectations". The CIO.com write-up gives a sample of 783 rather than 782, and no outlet reports regions, company size bands or a margin of error. My view is that the 28% is best read as a self-assessment against each organisation's own expectations, not as a measured financial return.

That reading has a consequence. A team that set a modest target and hit it counts as a success. A team that promised to halve incident volume and cut it by a third counts as partial. The figure says as much about how expectations were set as about what AI delivered.

How "AI in I&O" became "AI infrastructure projects" in the headlines

Gartner's own title says "AI projects in I&O". The Register's headline turned this into "AI infrastructure projects", and the phrase spread through syndicated coverage. To a reader in finance, "AI infrastructure" means GPUs, data centres and cloud compute. To Gartner's respondents, it meant AI used to operate infrastructure.

The two are different investments with different failure modes. An AIOps tool that fails to auto-remediate a network fault costs a licence and some engineering time. A GPU cluster that sits at low utilization costs a committed cloud bill or a depreciating asset every hour of every day. Treating one survey as evidence about the other is the most common error in the coverage, and I think it is the single most useful thing to correct.

What the 28% headline implies versus what the survey covered
QuestionWhat the headline impliedWhat the survey covered
Unit of measureAI infrastructure projectsAI use cases inside I&O teams
Typical investmentGPU clusters, data centre capacityAIOps, ITSM automation, cloud operations tooling
RespondentsNot specified782 I&O leaders, November and December 2025
Success testFinancial ROI"Fully succeed and meet ROI expectations", wording unpublished
Where the headline is rightBoth readings agree that most AI spending in IT has not yet produced the return its sponsors expected.

The 52% in the middle is the real AI infrastructure ROI story

Subtract the two published shares from 100 and you get the figure nobody quoted. If 28% of use cases fully succeed and 20% fail outright, 52% land somewhere in between. Gartner's coverage describes these as delivering partial value.

A second computed figure is more useful for planning. Of the use cases that did not fail, 28 out of 80, or about 35%, fully met expectations. Put plainly, surviving the pilot gives you roughly a one in three chance of hitting the business case, and a two in three chance of explaining a partial result to finance.

Outcome of AI use cases in infrastructure and operations Share of use cases. 52% is computed as 100 minus the two published shares. 28% full ROI 52% partial value 20% failed Strongest areas: ITSM and cloud operations Auto-remediation, self-healing Of use cases that did not fail, 28 in 80 (about 35%) fully met expectations. Source: Gartner survey of 782 I&O leaders, Nov to Dec 2025, released 7 April 2026. Chart: Zan Digital.
The red and blue blocks get the headlines. The grey block is where most teams sit, and it is the one with no agreed definition.

Where failure clusters

I&O leaders most often observed failures in auto-remediation, self-healing infrastructure and agent-led management of workflows within and between systems, per The Register. These are the use cases where AI is asked to act on production systems, not to suggest. Gartner research director Melanie Freeze said leaders "assumed AI would immediately automate complex tasks, cut costs, or fix long-standing operational issues".

Freeze also said that "the 20 percent failure rate is largely driven by AI initiatives that are either overly ambitious or poorly scoped." Among leaders who faced setbacks, The Register reports that 38% said persistent skill gaps continue to hamper AI success. The pattern links to the wider record on AI deployments that ended with negative ROI, where autonomy granted too early is a recurring cause.

Where success clusters

The strongest results came in generative AI applied to IT service management and cloud operations, with 53% of I&O leaders reporting success there. Coverage described these as mature markets with proven business value. Ticket summarisation, knowledge article drafting and cost reporting are narrow, reversible and easy to measure.

The lesson I take is blunt. AI that drafts something a human approves pays back. AI that changes production state without a human usually does not, yet. Gartner's own prescription points the same way: embed AI into existing workflows, secure executive sponsorship and write a realistic business case, plus Freeze's line that "CEOs and CFOs need to play a more active role in setting funding criteria and approving major investments."

GPU utilization is the other AI infrastructure ROI problem

If you came to the 28% looking for evidence about GPU spending, the better evidence is utilization data. It is thinner than a Gartner survey, and it comes mostly from vendors who sell optimisation. It also points to a larger gap between what is paid for and what is used.

Cast AI's 5% fleet average

Cast AI's 2026 State of Kubernetes Optimization Report, released 21 April 2026, reported that GPU utilization averaged 5% across the clusters it analysed. The data covers tens of thousands of clusters on AWS, GCP and Azure, with GPU figures running through April 2026. CPU averaged 8% and memory 20%.

Two caveats belong next to the number. The data was collected before customers enabled Cast AI's automation, so it describes organisations that had chosen to buy an optimisation product. And Cast AI's own blog notes that utilization measures compute activity, so a GPU can be fully allocated to a workload and still show low use. Co-founder Laurent Gil framed the cost plainly: "A GPU sitting idle costs dollars per hour. A CPU sitting idle costs cents."

What the survey data adds

A VentureBeat Pulse survey published 23 July 2026 asked enterprises directly. Of 100 respondents operating their own GPUs, 83 reported utilization of 50% or less, and the publication says 49% run at 25% or below. The sample was 107 enterprises, self-selected, skewed to mid-market firms, and labelled directional by its authors.

The same survey found that only 44% track compute cost and ROI rigorously, and 20% cannot quantify it yet. That second figure is the link to Gartner's result. You cannot meet an ROI expectation on infrastructure you do not meter, which is why allocating GPU cost to the teams that consume it comes before any ROI conversation.

Paid GPU-hours for every hour of useful work Multiplier equals 1 divided by utilization. Applies to any fixed price per GPU-hour. 5% (Cast AI avg)20x 10%10x 25%4x 50%2x 80%1.25x 83 of 100 GPU-operating firms in a July 2026 VentureBeat survey sit at 2x or worse. Arithmetic illustration. Sources: Cast AI, April 2026; VentureBeat Pulse, July 2026. Chart: Zan Digital.
This chart shows a relationship, not measured spend. The point is the curve: moving from 5% to 25% utilization cuts the effective cost of useful work by 80%.

The multiplier is the computed figure the survey coverage never produced. At a fixed price per GPU-hour, the effective cost of each useful hour is the list price divided by utilization. At 5% that is 20 times the rate card. At 25% it is 4 times. Any ROI model that uses the list price as its cost input is overstating return by that factor. A fuller ownership model sits in the on-premises versus cloud GPU break-even analysis.

Cloud commitments turn idle GPU capacity into fixed cost

Low utilization would be cheap if GPUs were billed only when busy. Most enterprise GPU capacity is not bought that way. It is bought on commitment, because commitment is where the discount lives.

AWS describes Savings Plans as offering low prices "in exchange for a commitment to a consistent amount of usage (measured in $/hour) for a 1- or 3- year term". EC2 Instance Savings Plans give the lowest prices and lock you to one instance family in one region. Once signed, the hourly commitment is owed whether the GPUs run at 5% or 95%. That is the mechanism that converts a utilization problem into an ROI problem.

Prices moved 44% in one announcement

On 5 June 2025 AWS announced price cuts of up to 45% on its NVIDIA GPU instances, with P5 H100 On-Demand prices down 44% from the 31 May 2025 baseline. The reductions applied to Savings Plans purchased after 4 June. A team that committed at the old rate a month earlier was paying the earlier price for capacity its competitors could now buy for far less.

The general point outlasts that one cut. GPU rental prices fall in steps, and multi-year commitments freeze your cost at the step you bought on. I would treat a 3-year GPU commitment as a bet on two numbers at once, your future utilization and the future market price, and most business cases model neither.

Flexera's 2026 State of the Cloud Report, released 18 March 2026, adds the aggregate view. Respondents estimated wasted cloud spend at 29%, the first increase in five years, and Flexera linked the rise to AI workloads. The figure is self-reported, so it is a measure of what cloud teams believe they waste, not what they do. The tooling gap behind that belief is covered in where FinOps tools fall short on AI spend.

Five causes of AI infrastructure projects missing ROI

Put the Gartner findings next to the utilization and commitment data and five causes recur. Each one is observable before money is spent, which is the practical value of naming them.

Why AI infrastructure projects miss ROI, and the test that catches each cause early
CauseEvidenceTest before you commit
Expectations set too highGartner: failures driven by projects "overly ambitious or poorly scoped"Write the ROI target as one number with a baseline and a date
Autonomy granted too earlyFailures cluster in auto-remediation and self-healingRun in recommend-only mode until accuracy is measured
Capacity bought ahead of demandCast AI 5% average GPU utilizationSize commitments to measured peak use, not forecast
Cost not meteredVentureBeat: 20% cannot quantify compute costTag every GPU-hour to an owning team before launch
Skills gap in the pilotGartner: 38% of leaders with setbacks cite skill gapsName the operator who will run it after the vendor leaves

The first two causes belong to the Gartner population. The middle two belong to GPU buyers. The fifth spans both. If your organisation is doing both kinds of AI infrastructure work, you will meet all five, and I would audit them in that order because the first is cheapest to fix. Setting kill criteria for AI projects before launch addresses the first cause directly.

Where this argument is weakest

This post separates two meanings of AI infrastructure ROI. That separation can be pushed too far, and the evidence on the GPU side is weaker than the evidence on the Gartner side.

The GPU numbers come from interested parties

Cast AI sells Kubernetes optimisation, and its 5% describes clusters belonging to companies that came looking for optimisation. That is a selection effect, and it likely pushes the average down. The VentureBeat sample is 107 self-selected firms. Flexera's 29% is an opinion survey. None of the three is a measured, representative census of enterprise GPU spending, and nothing in this post should be read as one.

There is also a measurement problem. GPU utilization measures compute activity, not whether a workload produced value. A training job at 40% utilization may be worth more than an inference service at 90%. Utilization is a cost signal, not an ROI figure, and I have used it as the former only.

Low utilization can be a rational choice

For latency-sensitive inference, spare GPU headroom is the product. A service that must answer within a fixed time cannot run its accelerators flat out. When GPU supply was scarce, holding capacity you might need was also a defensible hedge against not getting it at all. Some of what looks like waste is insurance, and the honest position is that public data cannot tell the two apart. The broader case that survey-based ROI numbers mislead in both directions is made in the comparison of AI ROI survey methods.

Frequently asked questions

What percentage of AI infrastructure projects meet ROI expectations?

Gartner's survey of 782 infrastructure and operations leaders, fielded in November and December 2025, found that 28% of AI use cases in I&O fully succeed and meet ROI expectations, and 20% fail outright. These are AI tools used to run IT operations, such as service management and auto-remediation, not GPU or data centre buildouts. Gartner has not published the question wording that defines success.

What did the Gartner I&O AI survey find?

Released on 7 April 2026, it found 28% of AI use cases in I&O fully met ROI expectations and 20% failed. 57% of leaders reported at least one failure. Failures clustered in auto-remediation, self-healing infrastructure and agent-led workflow management. Success was strongest in IT service management and cloud operations, where 53% of leaders reported success with generative AI.

Why do AI projects in infrastructure and operations fail?

Gartner attributes most failures to unrealistic expectations and poor scoping. Leaders assumed AI would immediately automate complex tasks, cut costs or fix long-standing operational issues. Among leaders who faced setbacks, 38% cited persistent skill gaps. Projects run as side efforts without a business case tend to stall, and use cases that let AI act on production systems fail more often than ones where a human approves.

What is average GPU utilization in enterprises?

Cast AI's 2026 report measured an average of 5% across the Kubernetes clusters it analysed, with GPU data through April 2026. A July 2026 VentureBeat survey found 83 of 100 GPU-operating enterprises at 50% utilization or less. Both sources are directional. Cast AI's sample is its own prospective customers, and VentureBeat's is 107 self-selected firms.

How does GPU utilization affect AI infrastructure ROI?

At a fixed price per GPU-hour, the effective cost of useful work equals the price divided by utilization. At 5% utilization you pay for 20 GPU-hours for every hour of useful compute. At 50% you pay for 2. Because most enterprise GPU capacity is bought on 1 or 3 year commitments, idle hours are billed whether or not the work exists to fill them.

Are cloud GPU savings plans worth it for AI workloads?

Only if you can forecast your utilization and accept price risk. A Savings Plan commits you to an hourly spend for 1 or 3 years in exchange for a discount. GPU prices fall in steps, and AWS cut H100 instance prices by up to 45% in June 2025. Commit to the capacity you already use at a steady rate, and buy growth on demand until it proves stable.

Where to start this week

Pull one number from your own estate: average utilization of every GPU you pay for, by owning team, over the last 30 days. If your monitoring cannot produce it, that gap is your first ROI finding, and it is worth more than any industry benchmark.

Then take your list of AI use cases in IT operations and sort each one into Gartner's three buckets, full, partial or failed, against the target written before launch. Any use case with no written target goes in a fourth bucket. In my experience that bucket is the largest, and it is the one to fix before the next budget cycle.

Related analysis

For the ownership math behind GPU decisions, read the GPU break-even utilization model. For deciding when to stop funding a pilot, read kill criteria for AI projects.

References

  1. Gartner, Gartner Says AI Projects in I&O Stall Ahead of Meaningful ROI Returns, 7 April 2026. Primary source for the 28% and 20% figures. The page blocked automated retrieval, so figures were confirmed through the coverage below.
  2. The Register, Only 28% of AI infrastructure projects fully pay off, survey finds, 7 April 2026. Used for sample, field dates, failure areas, the 57%, 38% and 53% figures and the Freeze quotes.
  3. CIO.com, AI often doesn't deliver ROI for IT departments either, 7 April 2026. Used for success factors and the 783 sample discrepancy.
  4. Cast AI, 2026 State of Kubernetes Optimization Report press release, 21 April 2026. Used for GPU, CPU and memory utilization and the Gil quote.
  5. VentureBeat, The AI compute gap, 23 July 2026. Used for survey utilization bands and cost tracking figures, n=107, directional.
  6. Flexera, Flexera Finds Cloud Value is Rising While AI Waste Grows, 18 March 2026. Used for the 29% wasted spend estimate.
  7. AWS News Blog, Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances, 5 June 2025. Used for price cuts and Savings Plan terms.

Weakest point in the source base: Gartner's question wording and success definition are unpublished, and every GPU utilization figure here comes from a vendor or a self-selected sample. Figures last verified 11 November 2026.

SK
Sanskriti Khandelwal
Writes for Zan Digital about AI product economics, B2B software markets and what the numbers behind vendor claims actually say.

Related reading