In the third quarter of 2025, the four largest technology companies in the world — Amazon, Microsoft, Google, and Meta — spent $106 billion on capital expenditure, most of it on AI infrastructure. That is a single quarter. By the time 2026 is done, Goldman Sachs projects total AI-related capital spending will track above $800 billion for the year. These are not figures from a speculative business plan. They are what has already been authorised, announced, and in many cases already spent on data centres, chips, power infrastructure, and the software layers built on top of all of it. The money is real. What is less clear is whether anything measurable is coming back.
In July 2025, MIT's NANDA initiative published a report called "The GenAI Divide: State of AI in Business 2025." It surveyed organisations that had collectively invested between $30 and $40 billion in generative AI. The finding was direct: 95% of those organisations reported no measurable return. Only 5% of integrated AI pilots had produced any impact on profit and loss that could actually be recorded. Goldman Sachs, which cited the report in its own analysis, was careful to note that this does not mean AI is failing as a technology. It means that between AI working in a lab or a demo and AI producing measurable business value at an organisation-wide level, there is a gap that most companies have not crossed.
An EY survey from 2025 found something harder to explain away: 99% of companies in their sample reported financial losses tied specifically to AI-related risks, with an average loss of $4.4 million per company. Not unrealised gains, not missed projections. Actual losses. From the same year, a Wall Street Journal survey documented a significant gap between what C-suite executives reported about AI's productivity impact and what the workers using the tools reported. The executives said it was working. The workers said it was not. Both groups were describing the same companies.
The WalkMe 2026 State of Digital Adoption report, which surveyed 3,750 executives and employees across 14 countries, put numbers on that disconnect. 81% of executives believed they had significantly improved productivity through AI. Workers in the same organisations reported losing 7.9 hours per week to digital friction — the equivalent of one full working day, every week. Across the year, that adds up to 51 working days lost to technology problems per employee, a figure that was up 42% from 2025. In the same period that companies increased digital investment by 38%, the time lost to those digital tools went to a three-year high. The more companies spent, the worse the friction got.
Part of the explanation is that employees are not using the tools being bought for them. 54% of workers bypassed AI tools and completed tasks manually at least once in the past 30 days. A further 33% reported not having used any AI at all. Only 9% said they trusted AI for complex, business-critical decisions. Among executives in the same survey, 61% said they trusted AI for those decisions — a 52-point gap on the same question, inside the same organisations. The executives buying the tools and the employees expected to use them are not operating with the same understanding of what those tools can actually do.
There is also the verification problem, which tends to get undercounted. Research from Workday in 2026 found that 37 to 40% of the time supposedly saved by AI gets spent reviewing, correcting, and verifying the AI's output. You write a report 40% faster with AI assistance, then spend 25 minutes checking whether it invented a statistic. The net saving is smaller than the headline efficiency number suggests, and in some cases it disappears entirely. An AI hiring startup that tested frontier AI agents on 480 workplace tasks commonly performed by bankers, consultants, and lawyers found that every agent tested failed to complete most of the tasks assigned to it. These were not marginal or unusual failures. They were the standard performance of frontier tools on ordinary professional work.
A study published in February 2026, drawing on business outlook surveys across the US, UK, Germany, and Australia, found that the vast majority of CEOs surveyed saw little to no impact from AI on their operations. Thousands of them. The study's conclusions resurrected a concept economists had largely moved past — the Solow Paradox, named for the observation Robert Solow made in 1987 that computers were showing up everywhere except in the productivity statistics. The same pattern, four decades later, with a different technology. Goldman Sachs's own chief economist said that AI had boosted the US economy by "basically zero" in 2025.
Companies are also abandoning AI projects at a rate that is not getting much public attention. In 2025, 42% of companies abandoned most of their AI initiatives, up from 17% the year before. Only 48% of AI projects make it past the pilot stage. Researchers studying this have started calling it "pilot purgatory" — companies run a successful limited trial, cannot scale it into their actual operations, and the project stalls without ever generating the return that justified running the pilot in the first place. McKinsey's 2025 AI survey found that the organisations reporting real financial returns were twice as likely to have redesigned their end-to-end workflows before choosing their AI tools. The technology was not the differentiator. The organisational change that preceded it was.
The trust numbers from ManpowerGroup tell a similar story from the employee side. Across nearly 14,000 workers in 19 countries, regular AI use increased by 13% in 2025, but confidence in the technology's usefulness dropped by 18% over the same period. Workers are using these tools more than before and trusting them less. That combination — higher use, lower confidence — suggests people have gotten far enough into AI tools to see what they cannot do, and that has moved their assessment downward, not upward.
J.P. Morgan ran a capex analysis that nobody in a press release cited, but that is worth sitting with. They calculated that $650 billion in annual revenue would need to be generated "into perpetuity" to deliver a 10% return on current AI infrastructure investment. They drew a parallel to the late-1990s telecoms fibre buildout, where companies laid billions of dollars of cable, the revenue did not materialise fast enough to justify the spending, and the investors who funded it absorbed enormous losses. The parallel is not exact — fibre eventually did carry the internet traffic that justified building it, just not on the timeline the investment required. Whether AI revenue materialises on the timeline that $800 billion in annual spending requires is the question nobody has an honest answer to.
The hyperscalers themselves — Amazon, Microsoft, Google, Meta — have been burning through their free cash flow from operations to fund this spending. Goldman Sachs noted that companies in the S&P 500 spent $550 billion on share buybacks in early 2025, and that flatlined in Q2 as AI capital expenditure jumped 24%. Buybacks are one of the primary ways large companies return value to shareholders. Cutting them to fund infrastructure spending that is not yet producing measurable returns is a significant bet. KPMG, in a UK-focused section of its enterprise AI report, made a statement that has been quoted in several places since: "AI no longer needs traditional return on investment to be justified." That is a remarkable position for a major accounting and advisory firm to take, and it reflects where corporate decision-making on AI actually sits right now — somewhere between genuine conviction about the technology's long-term value and fear of being the company that did not invest when everyone else did.
Nobel economist Daron Acemoglu has projected a "modest 0.5% productivity gain over the next decade" from AI. McKinsey's estimate is that AI could add $4.4 trillion to the global economy. The gap between those two figures is not a rounding error — it is the difference between a technology that meaningfully reshapes economic output and one that produces a marginal increment. Both projections come from credentialed researchers using real data. The difference in their conclusions reflects genuine uncertainty about what the productivity statistics will show when the infrastructure spending has settled, and the tools have been in use long enough to affect operations at scale.
IBM's chief human resources officer said in early 2026 that the company planned to triple its number of young hires, specifically because displacing entry-level workers through automation would create a shortage of middle managers a decade from now, endangering the company's leadership pipeline. That is a company with a significant AI practice looking at the tools it is building and deciding the downstream cost of replacing early-career workers is too high. It is also an implicit acknowledgement that the human infrastructure inside an organisation — who learns what, from whom, and in what sequence — is not something AI currently replicates. The technology can handle tasks. It has not yet handled careers.
The money being spent is not going to stop. Goldman Sachs has consistently found that consensus estimates for AI capex have underestimated actual spending by significant margins for two years running. FOMO — the fear of missing a technology shift that turns out to be fundamental — is driving decisions that pure return-on-investment analysis would not justify. That is not irrational. The internet buildout of the 1990s also ran ahead of measurable returns for years before the returns arrived. The question is whether the companies' spending now will still be positioned to capture those returns when they do arrive, or whether the debt taken on to fund the infrastructure, and the shareholder value traded away through reduced buybacks, will have already extracted a cost that the eventual productivity gains do not recover.
References
- Goldman Sachs — Why AI Companies May Invest More Than $500 Billion in 2026 (December 2025): https://www.goldmansachs.com
- Yahoo Finance / Goldman Sachs — The AI Spending Boom: $800 Billion Tracking for 2026 (May 2026): https://finance.yahoo.com
- Fortune — Is AI a Bubble? Goldman Sachs Sceptics and the MIT Labs 95% Finding (May 2026): https://fortune.com
- MagicSuite / Goldman Sachs — Goldman Sachs Warns Cloud Giants Risk All Cash Flow on AI (June 2026): https://www.magicsuite.ai
- Fortune / Yahoo Finance — Thousands of CEOs Admit AI Had No Impact on Productivity or Employment (February 2026): https://finance.yahoo.com
- AI Productivity — CEOs Report AI Has No Impact: Study Details (April 2026): https://aiproductivity.ai
- Tom's Hardware — 80% of Companies Report No Productivity Gains from AI Despite Billions in Investment (February 2026): https://www.tomshardware.com
- WalkMe / AOL — Enterprises Lose 51 Workdays Per Employee to Technology Friction Despite Record AI Investment (April 2026): https://lite.aol.com
- BuildMVPFast — AI Productivity Paradox CEO Survey 2026: Why AI Isn't Moving the Needle (February 2026): https://www.buildmvpfast.com
- EY — Survey Reveals Companies Missing Up to 40% of AI Productivity Gains Due to Talent Gaps (November 2025): https://www.ey.com/en_gl/newsroom
- Goldman Sachs / AOL — Companies Cutting Share Buybacks to Fund AI Spending (2025): https://www.aol.com
- KPMG / CIO — Enterprise Disconnect Between AI Investment and ROI (April 2026): https://www.cio.com