Worldwide information technology (IT) spending will hit $6.31 trillion in 2026, up 13.5% from last year, according to Gartner’s latest forecast. Most of that growth is chasing artificial intelligence (AI). New research from RAND, MIT and Gartner’s own surveys shows most of the money still isn’t proving itself.
McKinsey has spent two years telling boards that technology’s real payoff shows up after deployment, through continuous investment in data, integration and modernization. McKinsey’s own surveys show why that theory faces such a hard test right now. Nearly two-thirds of companies say they haven’t even started scaling AI across the enterprise, let alone collected on the payoff consultants keep promising.
A Record $6.31 Trillion Bet on AI
Gartner, the research and advisory firm, has revised its 2026 spending forecast upward three times in eight months. Each revision points the same direction.
October’s forecast put the year’s total at $6.08 trillion. By February that had climbed to $6.15 trillion. The April update landed at $6.31 trillion, a 13.5% jump from 2025.
| Forecast Date | Projected 2026 Total | Year-over-Year Growth |
|---|---|---|
| October 22, 2025 | $6.08 trillion | 9.8% |
| February 3, 2026 | $6.15 trillion | 10.8% |
| April 22, 2026 | $6.31 trillion | 13.5% |
Data center spending is doing the heaviest lifting. Gartner expects it to grow 55.8% this year, the fastest of any category, and projects data center systems spending will surpass $788 billion in 2026 as hyperscale cloud providers race to add AI-optimized servers.
The same scramble is playing out inside individual companies. Amazon and Microsoft have each pledged major AI investments in India’s cloud market this year, part of the broader race for compute capacity that is inflating Gartner’s numbers.
Where the Payoff Disappears
None of that money is contingent on proof that it works. Three studies published in the past year show why the gap matters, each measuring it a different way.
RAND Corporation, a nonprofit policy think tank, found more than 80% of enterprise AI projects fail to deliver the business value they promised, roughly twice the failure rate of conventional IT projects. Broken down further, 33.8% of projects get abandoned before reaching production, 28.4% reach production but fail to deliver expected value, and 18.1% run without ever recovering the investment. Only 19.7% deliver on their business case.
MIT’s Project NANDA initiative, a research group at the Massachusetts Institute of Technology, put a number on the sharpest version of that gap. Its report, “The GenAI Divide: State of AI in Business 2025,” found about 5% of generative AI pilots achieve rapid revenue gains. The rest show no measurable impact on profit and loss (P&L) at all. The findings drew on 150 leadership interviews, a survey of 350 employees and an analysis of 300 public AI deployments.
It’s because they pick one pain point, execute well, and partner smartly with companies who use their tools.
Aditya Challapally, the report’s lead author and a research contributor to Project NANDA, said that about the small share of companies pulling ahead. Startups run by people barely out of college, he added, have taken revenue from zero to $20 million in a year by staying narrow.
Gartner’s own infrastructure survey, published in April 2026, lands closer to RAND’s numbers than MIT’s. Polling 782 infrastructure and operations (I&O) leaders, Gartner found just 28% of AI infrastructure projects deliver their promised return. One in five fails outright, and 57% of I&O leaders say they’ve seen at least one project fail within their own ranks.
S&P Global Market Intelligence tracked the trend from the other direction: 42% of companies abandoned most of their AI initiatives in 2025, up sharply from 17% the year before.
Why Do So Many AI Projects Stall After They Launch?
Most AI projects stall after launch because of data quality gaps, mounting technical debt and unclear ownership. Research from HFS Research, IBM and Cisco all points to the same pattern: systems that perform well in a contained pilot break down once they meet messy, disconnected production data.
Executives at Global 2000-scale organizations, surveyed by HFS Research and Unqork in September 2025, show worried AI tools would create new technical debt at a 43% rate. Fifty-nine percent flagged security vulnerabilities, and half pointed to integration complexity. Views on the long-term trend split further: 55% expected AI to eventually shrink technical debt overall, while 45% expected it to grow.
“Our research shows that enterprises are waking up to the economic reality of transformation,” said Hansa Iyengar, an HFS Research practice leader, in a company release. The real cost, she said, “isn’t in buying software.” It’s in maintaining and integrating it.
A 2025 study from the IBM Institute for Business Value found something similar. Companies that ignored what researchers call capability debt in their AI initiatives saw project returns drop by 18 to 29 percent, with timelines expanding by as much as 22%.
Data integration compounds the same problem. Companies with strong data integration report 10.3 times the return on investment (ROI) of those with poor data connectivity, which averages 3.7 times, according to a 2024 analysis from Integrate.io, a data integration research firm.
The Five Percent Do It Differently
A Harvard Business Review analysis published in March 2026, based on 847 enterprise AI deployments across 14 industries, found a small set of practices separates the companies that see real returns from the rest.
- Define success metrics before deployment. Companies that measure business outcomes before rolling out an AI tool are four times more likely to achieve ROI than those that deploy first and measure later, the analysis found.
- Name one executive who owns the outcome. AI initiatives with a named executive accountable for financial results, rather than a committee or the IT department, succeed at three times the rate of those without one.
- Redesign the workflow before choosing a tool. Companies reporting significant financial returns are twice as likely to have rebuilt the underlying workflow before selecting an AI tool, according to separate McKinsey research.
- Buy a narrow tool instead of building a platform. MIT’s Project NANDA found externally sourced AI tools reach production about 67% of the time, compared with 33% for systems built in-house.
Every one of these practices points the same direction: measurement and ownership decided before the deployment date, not after it.
CFOs Start Asking for the Math
Chief financial officers (CFOs), largely absent from AI budget decisions in 2024 and 2025, are now central to them. The first wave of enterprise AI spending moved fast with little financial oversight. That has changed in 2026.
AI Business Weekly reported that Forrester found enterprises are postponing 25% of planned AI spending into 2027 as financial scrutiny tightens. A separate Gartner survey found fewer than one-third of corporate decision-makers could identify specific financial outcomes tied to their AI investments.
Not every project is getting cut equally. Applications with a direct line to profit, like fraud detection, customer service automation and supply chain optimization, are keeping their funding. Projects that promise only diffuse productivity gains, without a clear tie to the P&L, are the ones getting shelved.
Professional services firm PwC’s 2026 Global CEO Survey found 56% of chief executive officers (CEOs) could not point to any measurable business impact from their AI spending, whether in higher revenue or lower costs.
AI Business Weekly also reported specific examples of the pullback. Uber capped spending on agentic coding tools at $1,500 a month per employee after the company burned through its entire 2026 AI budget of $3.4 billion in four months. Microsoft, the same report said, pulled internal licenses for Anthropic’s Claude Code after per-engineer bills climbed to $500 to $2,000 a month, shifting engineers to its own GitHub Copilot CLI tool instead.
The Split Widens Into 2027
Gartner expects the divide to keep widening. The firm’s 2025 forecast said 60% of AI projects lacking properly prepared, AI-ready data would be abandoned through the end of 2026, a prediction multiple analysts describe as already playing out.
Oxford Economics, a firm that tracks global technology spending, expects the reordering to keep accelerating for a decade. Its analysts project AI’s share of enterprise technology budgets will climb from under 4% today to nearly 23% by 2035, as the AI-specific slice of spending grows climbing from $340 billion toward $3 trillion by 2035.
Enterprises are on pace to spend $6.31 trillion this year without proof that most of it works. Oxford Economics expects AI alone to claim close to a quarter of that budget within a decade. The bet is already placed, years before most companies can show it paid off.
Frequently Asked Questions
Why Do Failure Rate Estimates Range From 60% to 95%?
Estimates differ because each study measures something different. RAND Corporation’s figure of more than 80% covers all enterprise AI projects and whether they deliver any promised business value. MIT’s Project NANDA figure of 95% looks specifically at whether generative AI pilots show up in P&L statements. Gartner’s narrower survey of I&O leaders found 28% of AI infrastructure projects meet expectations, with one in five failing outright. Critics of the higher figures note that some pilots are built as experiments and were never meant to reach production, which makes the higher failure numbers look worse than the underlying reality.
How Much of Gartner’s 2026 Forecast Goes to Generative AI?
Generative AI model spending is projected to grow 80.8% in 2026, according to Gartner, pushing GenAI’s share of the overall software market up by 1.8 percentage points this year. That growth rate has held steady even as Gartner has revised its total 2026 spending forecast higher three separate times.
Will Agentic AI Projects Fare Better Than Generative AI Pilots?
Not necessarily. Gartner expects 40% of enterprise applications to include task-specific agents by the end of 2026, but the same firm forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027 as costs and governance requirements catch up with early enthusiasm.
Why Does AI Expose Technical Debt That Stayed Hidden Before?
AI removes the workarounds that let old shortcuts stay invisible. Cisco’s engineering team has described how AI-optimized workloads expose fragmented systems, messy data and manual patches that companies had relied on for years, turning what used to be a background maintenance cost into a structural barrier that slows AI projects before they reach production.
Does Company Size Change the Odds of AI Success?
Most of the hard data on AI failure rates comes from large enterprises; RAND’s research sample skews heavily toward big companies. Smaller firms spend far less on AI relative to revenue, an average of 0.35% of revenue among German Mittelstand companies, according to consulting firm Horvath, and tend to fail less dramatically, but they run into the same three problems: weak data foundations, unclear ownership and projects that drift from their original goal.
