NVIDIA Finds 88% Revenue Impact, McKinsey Finds 37% EBIT Impact. Here's Why

NVIDIA's 2026 State of AI report says 88% of respondents saw AI increase annual revenue. McKinsey's 2026 survey says only 37% attribute any EBIT impact to AI. Both are enterprise AI ROI survey results from the same twelve months, and both are honest readings of different questions. The gap comes mostly from one choice: where on the income statement each survey chose to look.

This post is written for the founder or director who gets handed one of these numbers in a board deck or a vendor pitch, because that reader converts on knowing which benchmark to trust and which to discount.

The short answer. NVIDIA's 88% and McKinsey's 37% measure different things. NVIDIA counted any revenue increase in any part of the business, from an opt-in audience of its own contacts. McKinsey counted a positive contribution to whole-company EBIT, from a GDP-weighted global sample. The unit of measurement, local revenue against enterprise profit, produces most of the gap.

Two enterprise AI ROI surveys, two opposite headlines

NVIDIA published its annual State of AI findings on 9 March 2026. The surveys were fielded from August to December 2025 and drew over 3,200 respondents. They covered financial services, retail, healthcare, telecommunications and manufacturing.

McKinsey published the tenth edition of its global survey in August 2026, titled The state of AI in 2026: On the road to ROI. Its central profit finding is that 37% of respondents say AI made a positive contribution to their organisation's EBIT. That share was essentially flat on the year before.

Put those two numbers on one slide and they look like a contradiction. One says nearly 9 in 10 companies are earning more from AI. The other says nearly 2 in 3 cannot find it in their profit line.

Both reports get quoted far beyond their original audience. The 88% suits a vendor making the case for more AI spending. The 37% suits a sceptic, a consultant selling transformation, or a CFO pushing back on a renewal.

My position is simple. Neither number is wrong. Quoting either one without the question behind it is the error, and it is an error both camps make.

The rest of this post takes each survey apart on three questions: who was asked, what counted as impact, and what threshold counted as a positive answer. Those three questions work on any AI ROI statistic you will meet this year.

Who each AI ROI survey asked

Sample design is the first place two surveys can diverge. A survey answers the question "what do these people think", and the identity of "these people" changes the answer before a single question is asked.

NVIDIA figures from its 9 March 2026 blog. McKinsey figures from ICTbusiness and FM Magazine coverage of the August 2026 report. Neither source publishes the full questionnaire.

NVIDIA's sample is an opt-in audience

NVIDIA is unusually candid about this. Its blog states that respondents "comprise people who've opted in to receive communications from NVIDIA and have invested in or are curious about AI for their business". That is a population already leaning toward AI.

The role mix tilts the same way. 40% of respondents are AI practitioners, the people who build and run the systems. A practitioner asked whether their work increased revenue is not a neutral witness, and nobody should expect them to be.

Company size matters too. 34% of NVIDIA respondents work at companies with fewer than 100 employees. In a 60-person firm, "some parts of the business" can mean one product line, and a single AI feature can move it visibly.

McKinsey's sample is weighted toward the global economy

McKinsey's 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 annual revenue.

Weighting by GDP means answers from large economies count for more. It does not make the sample random, and the coverage I reviewed does not describe how participants were recruited. It does mean the result is built to say something about the economy, not about one vendor's audience.

How each survey defined AI impact

Sample differences are real, and they are the explanation most people reach for first. I think they are the smaller half of the story. The larger half is the wording of the question itself.

NVIDIA: revenue, in some or all parts of the business

NVIDIA's blog reports that "88% of respondents said AI has had an impact on increasing annual revenue, in some or all parts of the business". Three features of that phrasing matter.

First, it is revenue, the top line. Revenue says nothing about what it cost to earn. Second, "some or all parts" means a gain in one team, one product or one region qualifies. Third, it asks about an impact on increasing revenue, which is a judgement about causation made by the respondent, with no counterfactual.

The blog also reports that 87% said AI helped reduce annual costs. Again, there is no requirement that the saving exceed the cost of the AI that produced it.

McKinsey: EBIT, at the level of the organisation

EBIT is earnings before interest and taxes, the standard measure of operating profit. It is revenue minus every operating cost, including the licences, compute, staff and consultants an AI programme consumes.

McKinsey's coverage reports that "only 37 percent of respondents say AI has made a positive contribution to their organization's EBIT". The unit is the organisation, and the measure is net.

That single change of wording does two things at once. It moves the question from gross to net, and from a part of the company to the whole of it. Each shift alone would shrink the share of yes answers. Together they shrink it a lot.

One detail in NVIDIA's own release deserves attention. The same blog reports that 64% of respondents are actively using AI, 28% are still assessing it, and 8% are not using it with no plans to start.

88% reporting a revenue impact is higher than the 64% actively using AI. Either the revenue question was put to a subset, or some respondents still in assessment reported revenue gains. The blog does not say which.

I am not suggesting anything improper. I am saying that a headline percentage without a stated denominator cannot be compared to anything, and this one does not state its denominator.

What counted as a positive answer in each AI ROI survey

The third lever is the threshold. A survey can ask a strict question and accept a loose answer, or the reverse. Both of these surveys publish their headline at the loosest possible bar, which is "any".

NVIDIA breaks its 88% into bands. 30% reported an increase above 10%, 33% reported 5 to 10%, and 25% reported less than 5%. The three bands sum to the 88%.

So a quarter of all respondents are counted in the headline on a revenue gain of under 5%, somewhere in the business. That is a modest claim, and it is fair to report. It is not the same claim as "AI raised revenue" in the sense a reader takes from a slide.

McKinsey's 37% is also an "any" measure. In the 2025 edition, McKinsey said that 39% of respondents attributed any level of EBIT impact to AI, and that most of those put it at less than 5% of their organisation's EBIT.

The strict version is McKinsey's high-performer bar: at least 5% of EBIT attributed to AI, with the value described as significant. About 6% of respondents clear it, in line with 2025.

The cleanest like-for-like reading is the top band on each side. 30% of NVIDIA respondents report a revenue gain above 10% somewhere. About 6% of McKinsey respondents report a material, significant share of whole-company profit from AI.

Those two numbers are still not measuring the same thing. They are at least both trying to describe a large effect, rather than any effect at all. I would quote those two before I quoted the headlines.

The single difference that produces the gap

If I had to name one difference, it is the unit of measurement: local gross revenue against enterprise net profit. Sample and threshold widen the gap. The unit creates it, and it would create a gap even if both surveys had asked the same people.

The strongest evidence comes from inside McKinsey's survey. McKinsey also asks about effects by function, and the coverage reports that revenue gains are most commonly associated with AI in marketing and sales, product and service development, and software engineering. Cost reductions cluster in supply chain, service operations and manufacturing.

The 2025 edition was explicit about the split. It said that while enterprise-wide EBIT impact was limited, many respondents reported cost benefits from individual AI use cases. Ask McKinsey's respondents the local question and you get something closer to NVIDIA's answer. Ask the enterprise question and you get 37%.

That is the whole reconciliation. The two surveys are measuring two layers of the same company, and the layers disagree because costs and gains land in different places.

An illustrative P&L shows how both answers can be honest

Take a company with $100 million in revenue and $10 million of EBIT. Its sales unit brings in $20 million. An AI programme lifts that unit's revenue by 5%, or $1 million. Every number in this example is illustrative, not drawn from either survey.

Suppose the extra revenue carries a 40% contribution margin, so it adds $400,000 of gross profit. The programme costs $800,000 a year in licences, compute and one hire. Enterprise EBIT falls by $400,000.

The sales head answers NVIDIA's question with a truthful yes. The CFO answers McKinsey's question with a truthful no. Now imagine the sales head is the one filling in the survey. That is exactly the person a practitioner-heavy, opt-in sample reaches.

The example is deliberately unflattering, and plenty of real programmes clear their costs. The point is narrower. A revenue question cannot tell you whether the programme pays for itself, and a profit question can. The longer version of that reasoning, applied to individual deployments, is in the analysis of AI deployments that lose money after go-live.

A tempting alternative is that the surveys were fielded at different moments. NVIDIA's ran in late 2025 and McKinsey's in mid 2026. But McKinsey's 2025 edition, fielded closer to NVIDIA's window, found 39% reporting any EBIT impact. The profit-level number barely moved across the period. Timing is not the variable.

Where this argument is weakest

I have argued that question wording does most of the work. Here is where that case is softest, stated as plainly as I stated the argument.

Selection may explain more than I credit

NVIDIA's respondents opted in because they are interested in AI, and those firms may simply be further along. NVIDIA's own data supports that reading: a little over 40% of its executive respondents reported a revenue increase of more than 10%, higher than the 30% across all respondents.

If NVIDIA had asked an EBIT question of the same audience, the answer might still be well above 37%. I cannot rule that out, and nobody can until someone asks both questions of one sample and publishes the result.

Neither questionnaire is public in what I could verify

NVIDIA's blog describes its result rather than quoting the question. I also could not open McKinsey's full report during research, so every McKinsey figure here comes from FM Magazine, ICTbusiness and MarketScale coverage of it. Those outlets agree with each other on the numbers, but they are Tier 2 restatements of one document.

That matters for the function-level evidence especially. The coverage names which functions report revenue gains, but does not give the percentages. My claim that the local question produces NVIDIA-like answers inside McKinsey's sample rests partly on the 2025 edition's wording.

Both sponsors have something to sell

NVIDIA sells the compute that AI programmes run on, and a high revenue-impact number supports more spending. McKinsey sells transformation consulting, and a persistent gap between adoption and profit is the problem that work claims to solve.

I do not think either report is dishonest. I do think the gap between them suits both sponsors, and a reader should hold that in mind. The broader pattern of who funds published AI research is covered in the piece on original research as a content moat.

Three questions for reading any AI ROI statistic

The method above generalises. Before an AI ROI statistic goes into your board pack or your budget case, put three questions to it. Most of the time the answers are on the vendor's own methodology page, if it has one.

The same test works on a single case study, not just a survey. A vendor quoting a 40% productivity gain has answered question two with "productivity", which is neither revenue nor profit. A step-by-step method for checking those claims is in the guide to forensic reading of AI ROI case studies.

It also explains why the metrics companies report are shifting. Firms that cannot find AI in EBIT are reaching for hours saved and output per head instead, a move tracked in the analysis of the shift from ROI to productivity metrics.

A team lead deciding whether to keep a tool needs the local question. A CFO deciding next year's AI budget needs the enterprise question. Using the wrong survey for your decision is the most common way these numbers mislead.

If your decision is about the total AI line in the budget, McKinsey's frame is the relevant one. The way that frame compares with the spending forecasts is set out in the comparison of Gartner's spending forecast with McKinsey's EBIT figure.

Frequently asked questions

Where to start this week

Find the last AI statistic that made it into a deck at your company, whether it came from a vendor, a consultant or your own team. Write down its answer to the three questions. If any answer is "unknown", add that word next to the figure on the slide.

Then do the harder version on your own programme. Take your largest AI line item and ask your finance lead one question: did it raise enterprise EBIT last quarter, net of its own cost? A local revenue win is worth knowing about. It is not the same answer.

Related reading