Who pays for an artificial-intelligence revolution, and who collects the returns? The question sounds like a contest among countries. It is first a question about three different balance sheets: the investor funding compute, the company that owns the resulting infrastructure, and the customer who uses AI to produce something of value. Their returns need not coincide.
The bill is visible. The dividend is harder to locate.
According to Stanford's AI Index 2026, privately recorded investment in AI in 2025 was approximately $285.9 billion in the United States and $12.4 billion in China. That is a comparison of recorded private investment, not of all Chinese or American spending. It excludes important public-sector financing routes and does not measure social returns.
The International Energy Agency reports that capital spending by five major technology companies exceeded $400 billion in 2025 and projected further growth in 2026. Those figures principally, but not exclusively, concern data centres. They are not a pure AI investment measure. A data centre's cost is an expenditure; neither an AI user's productivity nor the infrastructure owner's profit can be inferred from that expenditure alone.
American companies are not merely paying the bill for everyone else. They also own model businesses, distribution channels and much of the compute capacity. An Epoch AI estimate attributed roughly 71% of global AI compute, measured in H100 equivalents, to five large American hyperscalers at the end of 2025. The estimate depends on incomplete ownership data and a chosen equivalent-compute metric; it is an estimate of strategic concentration, not of profit or control over every component of the industry.
China is not one number behind America
China's lower recorded private-investment total is neither a complete measure of its funding nor proof that it cannot compete technologically. Chinese laboratories publish substantial model research, including the DeepSeek-R1 technical paper. The Stanford technical-performance analysis finds that the US-China model performance gap remained in single digits on its selected comparisons. Different measures of capability, openness and publication timing can yield different descriptions of the gap; none by itself establishes a winner in commercial deployment.
China also has an unusually large industrial automation base. The International Federation of Robotics reported that China accounted for 54% of global industrial robot installations in 2024. Robot installations are not generative-AI adoption. They do indicate the scale of industrial systems into which future AI capabilities might be integrated.
Commercial returns are a further question. Epoch AI's October 2026 analysis of Chinese AI companies describes a tension between wide adoption of open-weight models and the pricing power of the company that trained them. A model's diffusion can be a technological success while producing comparatively little revenue for its original developer.
Cheaper intelligence can help buyers and suppliers at once
Epoch AI's inference-price analysis documents very large but uneven declines in the cost of reaching selected benchmark performance thresholds. A falling price for a fixed capability does not establish that the total AI market will shrink. If lower prices unlock enough demand, aggregate spending on AI services and infrastructure can keep rising even while individual tasks become cheaper.
Equally, falling unit costs do not guarantee a profitable investment. Owners still face energy costs, network and property commitments, financing charges, and replacement of fast-aging hardware. If customers capture the lion's share of the benefit while price competition prevents suppliers from recovering capital costs, a technology can transform the economy without making every infrastructure investment successful.
The global windfall is not automatic
It is tempting to say that countries without frontier-model laboratories can simply import cheap intelligence and become the biggest winners. That is possible for particular firms and activities. But IMF research on the intelligence divide argues that the important gap may concern the ability to convert available technology into lasting productivity: human capital, organizational capacity, infrastructure and institutions remain decisive. It is a working paper and its model is not an observed global distribution of future gains.
A second IMF working paper values time currently saved by AI at $2.7 trillion annually, expressed as a labor-cost equivalent. That is a useful scale indicator. It is not realized GDP, company profit or cash return to data-centre investors. Time saved must be distinguished from additional sales, profits and broad prosperity.
An influential field study of AI assistance in customer support found a measurable improvement in cases resolved per hour, especially among less experienced workers. That is evidence of a bounded operational benefit, not evidence that the same percentage applies to factories, governments or entire national economies.
Three accounts that should not be merged
The first account is financial return: cash ultimately retained by the investors financing chips, data centres and AI companies, net of operation, depreciation and financing. The second is producer value: the additional output or cost reduction retained by businesses using AI in real workflows. The third is consumer surplus: benefits enjoyed by users without necessarily appearing as provider revenue. These quantities overlap in ordinary discussion but answer different questions.
The map is cross-border. US firms design prominent chips, Asian firms fabricate many of them, specialized suppliers provide equipment, international investors finance projects and customers worldwide use the services. Calling the entire investment bill "American" or the entire diffusion dividend "foreign" ignores both ownership and value chains.
The strongest objection
The hypothesis that AI users will capture most of the gains could fail. Scarce high-quality computing, distribution power, proprietary data, integration costs and long contracts could allow infrastructure and platform owners to preserve margins. Falling costs could stimulate more demand than supply can satisfy. Even firms that buy AI services might struggle to reorganize work and keep the savings.
A second objection cuts the other way: some of the most widely repeated figures concern investment commitments, gross revenues or benchmark prices, rather than audited net cash generation. The evidence does not yet establish either that an AI-investment bubble will collapse or that today's projects will earn attractive returns. Forecasts for 2027-2030 remain scenarios, not settled outcomes.
The test for the coming years
For each major AI investment, ask who owns the asset, what recurring cash flow it earns after operating and capital costs, who absorbs the model into productive work, and what measurable gain the end user actually receives. Track each account separately, across countries and across time.
The AI revolution may reward its builders, its financiers, its users, or some combination of all three. The most misleading answer is to treat technical success, investor success and broad prosperity as the same result.
Source and status, 12 October 2026. Adapted from the author-supplied Czech essay Kdo zaplati AI revoluci - a kdo na ni skutecne vydela?, dated 10 October 2026. The investment and compute figures are attributed estimates, not comparable measures of national welfare or guaranteed returns. IMF working papers do not express official IMF policy. The three-account distinction is an analytical framework. Public evidence review completed for publication version 1; companion evidence review is available in the public investigation. No corrections recorded.