AI Spending Up 49.5% to $2.7T, EBIT Impact Flat at 37%: Reading the Gap

Gartner expects worldwide AI spending to reach $2.67 trillion in 2026, up 49.5% in a single year. McKinsey's survey, fielded the same spring, found that 37% of respondents attribute any EBIT impact to AI, essentially unchanged from 2025. Put the two series side by side and the AI ROI gap is not closing. It is widening, and the reason is mostly about what each number counts.

This post is written for the founder or director who signs off the AI line in next year's budget, because the decision this gap forces is a capital allocation decision, not a tooling one.

The short answer. AI spending and AI profit impact are measured by different instruments. Gartner counts what vendors sell, so it grows with every server a hyperscaler installs. McKinsey counts what managers attribute to AI in their own earnings, so it moves only when work is redesigned. The gap closes through workflow change and cost discipline, not more spend.

What Gartner's $2.7 trillion AI spending forecast actually counts

The headline figure comes from Gartner's forecast published on 16 September 2026. It puts worldwide AI spending at $2.67 trillion in 2026, up from $1.79 trillion in 2025, and $3.64 trillion in 2027.

Gartner's forecast is a market sizing exercise. It estimates revenue across categories of AI products and services, globally. It is not a survey of how much your peers are spending, and it is not a measure of enterprise budgets alone.

That distinction is the whole first half of this post. A large part of the 2026 number is the supply chain buying from itself.

Infrastructure is more than half the total

AI infrastructure is forecast at about $1.48 trillion in 2026, up from roughly $982 billion. Gartner defines it as AI-optimised infrastructure as a service, AI-optimised servers, network fabric, processing semiconductors and devices. That is about 55% of the total.

John-David Lovelock, Distinguished VP Analyst at Gartner, called the data centre buildout "the largest infrastructure project humanity has ever undertaken". He added that hyperscalers and service providers buying AI-optimised servers will remain the largest single area of spending. Most of that capital sits on the balance sheets covered in the analysis of hyperscaler capex and its missing payback schedule.

So when someone cites $2.7 trillion as evidence that enterprises are pouring money into AI, they are mostly citing data centre construction. The enterprise-facing lines, software and services, add up to about $1.04 trillion.

Agents and assistants grow fastest from the smallest base

The angle that drew attention in September was agents. Gartner forecasts spending on AI agents and assistants to grow 77.3%, from about $16.5 billion to $29.2 billion. That is a striking rate on a line that is about 1% of the total.

Gartner also changed the category itself this year. It separated cross-functional agents and assistants from AI software and added consumer agents. Agent features embedded in existing products stay counted as software. Growth rates across a redefinition deserve a light touch.

The forecast has also been revised upward all year. In January, Gartner put 2026 at $2.52 trillion, up 44%. By September it was $2.67 trillion, up 49.5%. Spending expectations rose by roughly $150 billion in eight months, during the same period the profit series stayed put.

What McKinsey's 37% EBIT impact figure actually measures

The other series is McKinsey's annual survey, published in August 2026 as The state of AI in 2026: On the road to ROI. Its central profit finding is that 37% of respondents attribute at least some EBIT impact to AI, around the same as the previous year.

EBIT is earnings before interest and taxes, a standard measure of operating profit. The question is whether AI moved it at all, not by how much.

Who answered, and how it was weighted

The survey ran online from 4 May to 8 June 2026, with 1,719 participants in 97 countries. Responses were weighted by each respondent's country contribution to global GDP. About 36% of participants work at organisations with more than $1 billion in revenue.

That is a large and international sample. It is also self-reported. A respondent attributing EBIT impact to AI is giving an opinion about causation inside their own company. Nobody audits the answer.

High performers stayed at 6%

McKinsey defines AI high performers as organisations attributing at least 5% of EBIT to AI and describing the value as significant. They were about 6% of respondents, unchanged from 2025.

Meanwhile adoption measures rose. The share scaling AI across the enterprise went from 38% to 44%. Among large companies, the share scaling AI agents in at least one function rose from 27% to 40%, while smaller organisations stayed flat at 22%.

The pattern is consistent: more firms are deploying more AI more widely, and the share seeing it in operating profit has not shifted. That is the gap in its cleanest form. Why measured productivity has also been slow to appear in national statistics is covered in the analysis of the AI productivity paradox in GDP data.

AI spending and EBIT impact, the two series side by side

Here is the comparison the angle promised. One series counts dollars. The other counts the share of respondents reporting profit. They cannot be placed on one axis honestly, so the chart below indexes each to its own 2025 value.

The comparison is unfair in one direction. Spending runs ahead of profit by design, because infrastructure is built years before it earns. Nobody should expect a server installed in 2026 to show up in a buyer's EBIT the same year.

It is unfair in the other direction too. The software and services lines are bought by the same enterprises answering McKinsey's survey. Those lines grew by roughly a third to 60% this year, and the profit share did not budge. That part of the gap cannot be explained by construction lag.

My reading is that the infrastructure half of the gap is a timing question and the enterprise half is a management question. Conflating them is how both bulls and bears end up wrong.

The one-in-five cost constraint is the number to watch

The finding that deserves more attention than the 37% is buried further down McKinsey's results. About one in five respondents say their organisation has limited AI use because of operating costs, including tokens, computing resources and storage.

That is a cost wall appearing before most firms have booked a return. If 37% see any EBIT effect and 20% are already rationing use on cost, a meaningful share of organisations may be paying more without yet earning more.

What respondents meant by operating costs

These are run costs, not project costs. Every agent call, every long-context query and every stored embedding bills per use. Unlike a licence, the bill grows with adoption, so success at rollout raises cost before it raises profit.

This matches the pricing shift tracked in the piece on why cheaper tokens still produce rising AI bills. Unit prices fall while total consumption rises faster, so the invoice goes up. Gartner's own release notes that enterprises are asking providers for help managing costs and embedding usage tracking into workflows.

Lovelock also said that lock-in, data sovereignty and "run-away costs are not deterring buyers". Read alongside McKinsey, that is not reassuring. It means buyers are absorbing a known cost risk in exchange for a return most cannot yet show.

Why high performers hit the cost limit more often

The counterintuitive detail is who reports the constraint. In software coding agents, ICTbusiness reports that cost limitations are reported roughly three times as often among high performers as among other respondents.

That reframes the cost wall. The firms getting value are the ones using enough to feel the bill. A cost constraint can be a sign of success, which is why I would not read the one-in-five figure as a reason to pull back. I would read it as the reason to measure cost per completed task before scaling anything.

High performers are also more than twice as likely to spend over 15% of their ICT budget on AI. Separately, 28% of all respondents already spend more than 10% of enterprise ICT budget on AI, and 60% expect AI investment to increase next year.

Where this argument is weakest

The neat story is that spend rockets while profit flatlines. Here is where it falls apart.

The two series measure different populations

Gartner's total includes consumer devices, hyperscaler servers and model training. McKinsey asks people inside organisations about their own earnings. Dividing one by the other produces a ratio with no meaning, and I have deliberately not done it.

A fairer test would set enterprise software and services spend against enterprise EBIT impact. Even then, the populations differ by country weighting, company size and sector mix.

Survey EBIT attribution is a noisy, lagging instrument

Asking whether AI contributed to EBIT asks a respondent to isolate one cause among hundreds. Many cannot, and some cost savings land in budgets that never reach the question. A flat 37% may reflect measurement limits as much as business reality.

The question is also binary. A firm whose AI contribution grew from 1% to 4% of EBIT answers the same way in both years. The headline share cannot see intensity, only presence.

The MIT study is not corroboration

Many posts on this topic pair McKinsey with MIT Project NANDA's 2025 finding that most generative AI pilots showed no measurable P&L impact. That study was preliminary, built on a small interview base, and gathered in early 2025. Its limits are set out in the breakdown of what the 95% figure actually says.

Stacking a preliminary study on a self-reported survey does not make either stronger. Two weak instruments pointing the same way are a reason to investigate, not a verdict. The honest claim is narrower: no public series yet shows enterprise AI profit keeping pace with enterprise AI spend.

What closes the gap between AI spending and EBIT impact

If the enterprise half of the gap is a management question, the levers are visible in McKinsey's own data. The strongest one is workflow redesign.

Nearly three quarters of high performers say they fundamentally redesigned workflows because of AI, compared with about one quarter of other respondents. Correlation is not proof of cause. It is still the widest gap between the two groups in the coverage.

Gartner's figures point the same way from the supply side. Lovelock noted that enterprises are turning to service providers less often for business transformation and more often for smaller projects to use AI features inside incumbent software. That is the cheaper path, and the one least likely to move EBIT.

The build-versus-buy row has a figure behind it. McKinsey's survey found 32% of respondents say their organisation decided not to buy at least one software product or feature because it could be built in house with agentic coding tools. That is a direct spend-to-margin conversion most budgets do not track.

The stopping rule matters as much as the start. A portfolio that never kills a pilot carries its failures as recurring run cost, which is exactly the population feeding the one-in-five constraint. How to write those rules before launch is laid out in the guide to kill criteria for AI projects.

What the gap means for your 2027 AI budget

You will be asked to raise the AI line next year. Sixty per cent of McKinsey's respondents expect to. The useful question is not how much, but which half of the gap your money lands in.

Money spent on seats and features inside unchanged processes buys adoption. The data says adoption alone has not moved profit for most firms. Money spent rebuilding one process end to end, with a baseline recorded first, is the version the 6% appear to be buying.

I would make one structural change before approving any increase: require a cost-per-task figure and a pre-deployment baseline for every AI line over a threshold you choose. The cost modelling side of that is covered in the enterprise agent cost model, and the failure pattern it prevents is set out in the arithmetic behind negative-ROI deployments.

The macro numbers will keep rising regardless. Gartner already forecasts $3.64 trillion for 2027. None of that tells you whether your own spend is working. Only your baseline can.

Frequently asked questions

Where to start before budget season

This week, sort every AI line in your current budget into two columns. One for spend that changed how a process runs, one for spend that added a feature to a process that did not change. If the second column is larger, you are funding the half of the gap that the survey data says does not reach EBIT.

Then pick the single largest item in the second column and ask its owner one question: what did this cost per completed task last month, and what did the same task cost before? If nobody can answer, that is your first 2027 budget decision made.

Related reading