Companies are spending billions on AI, but how many actually make a profit?

Four Big Tech groups expect 2026 capital spending of about 720-740 billion USD. McKinsey finds 37% reporting an EBIT effect and just 6% High Performers; chip and cloud profit is not the same as project ROI.

ตู้เซิร์ฟเวอร์ภายในศูนย์ข้อมูล

Major technology companies are spending on data centres, chips and networks for AI at a scale that would have seemed unimaginable just a few years ago. For calendar 2026, the capital-expenditure guidance and estimates disclosed by Amazon, Alphabet, Microsoft and Meta add up to roughly 720-740 billion USD. This is an indicative sum of differently defined capital budgets, not expenditure exclusively on AI and not a closed, audited full-year total.

As equipment suppliers begin to show profits in their financial reports, the buyers face a harder question: employees may work faster with AI, but how much of that improvement turns into measurable operating profit? In McKinsey's survey released on 25 August 2026, 37% of respondents said AI affected their organisation's EBIT in some degree, while only 6% qualified as AI High Performers, reporting at least a 5% EBIT contribution and substantial business value. That distinction captures the AI ROI Gap between spending money and earning a demonstrable return.

How many companies really profit from AI? First define profit

Three concepts frequently get conflated. First, a company can report overall net profit while selling many products that have nothing to do with AI. Second, a specific AI project can earn a positive net return once infrastructure, cloud and model fees, staff, integration, governance and reworked processes are included. Third, AI can materially lift operating profit across an entire organisation after controlling for other influences such as pricing and the economy. Only the third definition directly measures broad organisational profitability attributable to AI.

McKinsey indicator, 2026Share of respondentsHow to interpret
AI used in at least one functionalmost 90%Already using it, not evidence of profit
Scaling AI across the enterprise44%Up from 38% in 2025
AI reported to affect EBIT37%Some effect, potentially small
AI High Performers6%At least 5% EBIT impact and significant value
AI operating cost limits adoption~20%Token and running costs are constraints

McKinsey gathered responses from 1,719 people in 97 countries during May-June 2026. Respondents described their own organisations; this was not an audit of every company's AI profit. Therefore 6% does not mean only 6% of companies generate any positive AI returns, nor does it mean the other 94% operate AI at a loss. Many may be receiving smaller benefits, changing their processes or struggling to attribute gains accurately. The 2025 McKinsey survey reported roughly 39% seeing some EBIT impact and around 6% High Performers. As the samples were not identical, the 39%-to-37% movement must not be described as a verified two-percentage-point global decline.

How AI investment escalated from 2023 to 2026

From 2023 onward, demand for generative AI drove purchases of powerful processors and reservations of more electrical capacity. In 2023-2024 much activity was experimental; in 2025-2026 it became a race to build and fill data centres. The question shifted from whether models could work to when infrastructure expenditures would generate enough revenue to cover their ongoing costs.

PeriodInvestment developmentEvidence and limitations
2023 – 2024From GenAI pilots to infrastructure planningGPU, cloud and workflow integration began
2025World corporate AI investment in Stanford AI Index ~581,690,000,000 USDIncludes private capital and M&A, not just CapEx
2025IEA estimates five top tech companies spent over 400,000,000,000 USDIncludes non-AI infrastructure
2026IEA projects another rise of ~75%Forecast, not realised full-year spending

Stanford's roughly 581.69 billion USD corporate AI investment aggregate for 2025 includes private investment and merger-and-acquisition transactions. It must not be added to Big Tech CapEx: financing, acquisitions and fixed-asset purchases are different categories that can overlap. For 2026, the four major companies disclosed these indicative figures:

Company2026 CapEx guidanceCaveat
Amazon~220,000,000,000 USDCash CapEx across the whole group, not AWS/AI only
Alphabet195,000,000,000 – 205,000,000,000 USDAnnual corporate forecast including other infrastructure
Microsoft~170,000,000,000 USDCalendar-year estimate including finance leases
Meta135,000,000,000 – 145,000,000,000 USDIncludes principal payments under finance leases

The 720-740 billion USD illustrative sum is not accounting-standardised. Amazon describes cash CapEx across all its businesses; Microsoft's expected annual amount includes finance leases, while Meta also includes lease principal payments. All four invest in conventional cloud, retail, advertising and digital services as well as AI. Rising CapEx eventually entails depreciation on GPUs and servers, electricity, leases, land and facilities. A strategic investment today can protect a business position for years yet still require sufficient future revenue and cash generation to justify its cost.

Why so many businesses remain in pilot purgatory

MIT Project NANDA's preliminary 2025 report, The GenAI Divide, was widely described as finding that 95% of organisations were not realising measurable profit-and-loss gains from generative AI deployments. Its evidence included a review of more than 300 publicly described initiatives, interviews at 52 organisations and a survey of 153 senior leaders. This exploratory, nonrepresentative sample highlights the problem of pilots that never become operational workflows. It is not a census proving that 95% of the world's companies lose money on AI and cannot replace McKinsey's 2026 results.

A working model is not the same as an integrated business process. AI may write a draft faster while employees still have to verify the output, permissions and source data. Costs of connecting databases, handling errors, protecting information and paying for model calls emerge when deployment scales. Time saved in writing can become time spent checking, leaving overall cost little changed. According to McKinsey, 80% reported gains in individual productivity, yet only 37% reported an effect on organisation-wide EBIT. The difference is not evidence that AI has no value: savings become earnings only when expenses actually fall, sellable output rises or new revenue exceeds incremental cost.

Engineer inspecting systems in a data centre

Who captures profit first among chips, clouds and applications?

During an infrastructure build-out, a supplier can recognise sales when chips and systems are delivered. Its customer may have to wait years to recover the investment through usage fees. Infrastructure earnings are therefore more visible in several published accounts, though those company earnings do not prove that every AI project is profitable.

Company and accounting periodRelated revenueReported profit evidence
NVIDIA (Q2 FY2027)Data Center 89,000,000,000 USDConsolidated operating income 63,734,000,000 USD
Broadcom (Q3 FY2026)Total revenue 29,591,000,000 USDConsolidated operating income 15,955,000,000 USD
AWS (Q3 CY2026)46,000,000,000 USDAWS operating income 17,800,000,000 USD
Google Cloud (Q3 CY2026)30,709,000,000 USDGoogle Cloud operating income 11,121,000,000 USD

In its quarter ended 26 July 2026, NVIDIA Data Center revenue grew 117% year over year. In its quarter ended 2 August 2026, Broadcom's total revenue grew 86%. These are real sales and reported company operating profits from selling technology infrastructure; however, NVIDIA and Broadcom also have businesses outside generative AI. Their consolidated operating income should not be relabelled as isolated AI profit.

Cloud services show more substantial profit evidence too. AWS's Q3 2026 operating income increased to 17.8 billion USD from 11.4 billion USD a year earlier. Google Cloud's operating income reached 11.121 billion USD versus 3.594 billion USD a year earlier on revenue of 30.709 billion USD. These segments also provide databases, storage and non-AI computing, so their profitability is that of the cloud business, not separately audited AI-only earnings. Power utilities, grid-equipment makers, cooling providers and data-centre builders may benefit from orders, but contracts, construction costs, permitting and demand risk mean there is no guarantee every energy company or project wins.

Software revenue is growing, but AI-specific profit remains elusive

Software can charge subscriptions and distribute services to many users. Margins may improve if model inference costs decline faster than revenue grows. Conversely, heavy usage by large customers can overwhelm a fixed monthly subscription price, particularly when each request involves paid compute.

Microsoft's FY2026 consolidated net income was 133.749 billion USD, and in July-September 2026 it generated 18.6 billion USD of free cash flow. Neither figure is separately audited profit from Copilot or AI services. Treating Microsoft's overall profitability as proof that its AI investment has already paid back would confuse two measures.

Meta's Q3 2026 revenue was 64.601 billion USD, with 12.988 billion USD of operating income, but free cash flow was negative 2.120 billion USD while investment accelerated. Its quarterly earnings also included a roughly 9.9 billion USD one-time litigation charge unrelated to AI. Management says AI improves content recommendations and advertising, but consolidated profit and cash flow still cannot reveal a stand-alone return on AI. Likewise, many private application firms disclose user counts, annualised revenue or funding but not verified model-training expense, inference discounts, data and personnel costs or net income; sales figures alone cannot establish profit or loss.

AI coding agents introduce another challenge: McKinsey found 32% of respondents had declined to buy at least one software product or feature because it might be built internally with agentic coding tools. Vendors of generic features will face more pressure to justify product differentiation and subscription prices.

Measuring ROI without confusing slide-deck gains with earnings

A credible AI trial starts with a baseline and an outcome defined in advance: cost per customer ticket or order, rework rate, time to resolve cases or conversion rate. Use comparable test and control groups. Counting generated pages, processed tokens or model invocations is not counting profit.

ROI = (realised financial benefit – total investment and operating cost) / total investment and operating cost. Set a measurement period and include cloud/model fees, consultants, security, data integrations, staff and error checking. Saving two staff hours does not immediately create two hours' worth of cash savings unless paid cost is actually reduced or the released capacity makes additional sales. An organisation should expand projects only after verifying a genuine cost reduction or revenue gain; otherwise revise or stop the pilot. Some support functions legitimately target better service rather than cash payback within one quarter, and should be evaluated on that declared goal.

The risk is migrating from IT budgets to electricity and capital markets

The IEA estimates that data centres consumed about 415 TWh of electricity in 2024, potentially rising to about 945 TWh by 2030. Not all that electricity would run AI. Nevertheless, power supplies, transmission grids, land, cooling and construction constrain how quickly computing capacity can expand.

There are two opposing financial effects. CapEx becomes sales and employment for chip suppliers, electrical-equipment manufacturers, contractors and data-centre operators. But buyers also take on depreciation, lease commitments and capital costs before revenue catches up. If demand develops more slowly than forecast, expensive centres may remain partly idle. The income statement alone misses part of this risk.

Amazon reported negative 42.4 billion USD in free cash flow over the 12 months through September 2026, in part because purchases of fixed assets tied primarily to AI increased, even as AWS operating income grew. A profitable business can simultaneously consume substantial net cash in investment: negative free cash flow does not equal a net accounting loss, and the positive AWS result does not prove future returns on all the new infrastructure. Higher-for-longer interest rates and faster processor obsolescence can raise financing, depreciation or replacement costs. Alternatively, profitable downstream AI use might justify the installed capacity. The available evidence cannot settle which will dominate the global market.

What numbers will show AI investment has begun to pay off?

Over the next 12-24 months, look beyond chip orders to realised revenue per unit of compute, data-centre utilisation, added depreciation and power cost, free cash flow after CapEx, margin per AI request and the share of organisations reporting a bigger effect of AI on EBIT. Avoid comparing unlike measures: an equipment seller's strong earnings do not mean all its customers have recovered their purchase costs, and a software product's subscriber growth does not demonstrate sector-wide net margins.

Key takeaway Winners already report profits from chip and cloud businesses, but credible public information is still insufficient to count every company worldwide that earns separately identifiable net profit attributable to AI. McKinsey's 2026 survey found 37% reporting some EBIT impact and 6% qualifying as high performers. The long-term winners will be determined less by the amount of CapEx or number of users than by their ability to turn AI adoption into incremental earnings and free cash flow after investment.

Main sources

McKinsey – The state of AI in 2026 | MIT Project NANDA – The GenAI Divide | Stanford AI Index 2026 | IEA – Key Questions on Energy and AI

Amazon Q3 2026 | Alphabet Q3 2026 | Microsoft FY2027 Q1 | Meta Q3 2026

NVIDIA FY2027 Q2 | Broadcom FY2026 Q3 | IEA – Energy and AI

Sources

Topics: TechnologyCorporate EarningsCapital MarketsEnergy