AI’s growth engine needs a market

A business can automate part of its operations, produce more and improve its margins. The wider economy still has to find a paying use for that extra output. Between a successful deployment and the financing of the next generation lies a question that technical demonstrations rarely settle: who benefits from the improvement, who pays to use it, and how long does the inventor keep earning a return?

A working paper by Phurichai Rungcharoenkitkul, published by the Bank for International Settlements on 12 October 2026, examines that connection. In its core framework, the cost of a further advance scales with the economy, while innovation rents, income earned from a technological lead above current production costs, come from consumption goods. A shift towards investment can therefore shrink the revenue base that rewards research. The life of those rents and the markets in which they can be earned become decisive. The author presents conditional simulations and extensions under his own responsibility. ^[S01]

The argument reaches beyond the debate over displaced jobs. It raises a question about returns: a useful technology might spread while leaving too little income to fund its next advance. Equally, a profitable industry might distribute its gains in ways that do little for workers. Assessing either possibility requires following expenditure to its final use, then tracing back through ownership rights and research costs.

Machines have customers too

A “demand shortage” can be a misleading starting point. Building a data centre, buying equipment or developing software is economic activity. National accounts include investment in final demand. Corporate purchases also generate income for suppliers and workers. ^[S15]

The relevant questions concern the destination of that spending and the ability to retain income above production costs. Equipment can be in great demand yet be sold in a competitive market with narrow margins. Software that costs little to reproduce may reward its designer for longer through licensing or hard-to-replace expertise. Economic usefulness, revenue and innovation rents are different measures.

Consider a hypothetical supply chain. A laboratory sells a tool to a manufacturer, which uses it to make products for households. The immediate AI customer is a business, but consumers supply the ultimate revenue. In another chain, the tool helps produce equipment used to make still more equipment. Investment demand can support activity for a long time. Its sustainability then depends on future uses, cost savings and the possible decline in the price of machinery.

Input-output accounts distinguish intermediate sales, value added and final uses. Adding invoices across every link in a chain would artificially inflate the size of its end market. An investigation of AI revenue needs to follow those uses without repeatedly counting the same expenditure. ^[S16]

Method and scope

Conceptual diagram, with arrows drawn without a monetary scale. Core framework: rents on consumption goods and competitive investment production; extension: rents on investment goods too, appendix A.3. BEA definitions distinguish final demand, intermediate sales and value added. Investment itself is part of final demand.

Large corporate customers establish that a market exists. Assessing the durability of its income requires asking whether customers buy to serve profitable external demand, reduce an existing cost or prepare capacity whose utilization remains uncertain. A single company may be doing all three. Our analysis of world trade driven by AI orders follows this measurement problem through supplier chains and trade statistics.

A smaller share can mean more spending

The distinction between a proportion and an amount is crucial. Consumption that grows more slowly than output loses share, even while its volume keeps rising. Treating a decline in consumption relative to output as an automatic disappearance of customers substitutes a dramatic shortcut for the economic mechanism.

In the fictional example below, an economy initially produces 100 units and consumes 70. Its consumption share subsequently falls to 60%. With output of 110, purchases amount to 66 units and have fallen. With output of 140, they reach 84 units, an increase of 20%. Prices are held constant. These are separate illustrative alternatives, not forecasts.

Method and scope

Consumption = output × consumption share. Start: 100 × 0.70 = 70. Independent alternatives: 110 × 0.60 = 66; 140 × 0.60 = 84, or +20% against 70. Preserving consumption requires 70 / 0.60 = 116.666… units of output, a +16.666…% increase. Fictional units, constant prices, with no assumed time to realization.

Keeping consumption at its initial 70 units with a 60% share would require output of about 116.7 units, a 16.7% increase. That calculation makes the expansion needed to preserve spending explicit. Its timing and distribution across households would require further information.

Research profitability introduces another comparison. Sales may rise while the resources required for the next discovery rise faster. A team pays for experiments, compute, data, engineers and verification. A growing market becomes less attractive when the unavoidable cost of staying competitive grows more quickly. The useful signal is the return on an additional discovery, rather than total revenue growth alone.

Market size has an empirical history in innovation research. Acemoglu and Linn studied the pharmaceutical industry using demographic shifts in potential demand. Their work supports asking the question. Applying it to AI requires a fresh assessment of competition, intellectual property and commercialization; the pharmaceutical elasticities retain their original sectoral scope. ^[S10]

How long a lead keeps paying

A discovery is financed against future income. A business that expects to retain its advantage for several years can justify spending more than one whose lead will quickly be copied, overtaken or commoditized. The life of the advantage matters alongside initial sales.

Wendy Li and Bronwyn Hall’s research on the depreciation of R&D capital estimates the economic erosion of research returns. The evidence is historical and spans several industries, with estimation assumptions specific to those businesses. The physical life of a GPU and the life of rents earned by a current AI model would each require different evidence. ^[S09]

Our illustration isolates the effect in an entirely fictional project, expressed in real units. Research costs 60 units upfront and initially earns a rent flow of 20 units a year. That flow gradually erodes; every case uses the same 5% discount rate. Depending on how quickly the rent disappears, the present value of future income comfortably exceeds the research cost, barely covers it or falls short. The method note gives the continuous-time formula and the illustrative parameters selected by l0g.

Method and scope

Continuous time, fictional real units. Initial flow π₀ = 20/year, upfront cost K = 60, discount rate r = 0.05/year, erosion d = 0.10; 0.25; 0.40/year. The discounted flow is 20 exp[−(r+d)t]. Its infinite-horizon present value is 20/(r+d): 133.3; 66.7; 44.4 units; net present value after cost: +73.3; +6.7; −15.6. The curves show only the first fifteen years. Parameters selected by l0g; Li and Hall’s historical paper informs the erosion concept without calibrating these examples.

Competition may lower prices and spread productivity gains among users. It can also shorten the developer’s window for recovering costs. Longer-lasting rents may support research while sheltering incumbents from new entrants. Assessing that tension requires examining innovation incentives alongside the conditions for entering the market.

Research spending itself deserves equal attention. Tools that make experiments cheaper, identify dead ends earlier or improve validation can increase a project’s return without raising the customer’s price. A larger market and a lower invention cost are two possible ways to ease the same financing constraint.

The assumption that can change the outcome

The BIS paper’s core framework locates rents in consumption goods, while investment goods are produced competitively. The author also considers extending rents to business spending and, separately, automating research. In that research extension, limited savings can prolong automation before it stalls; sufficiently strong and sustained gains can keep progress going. These variants are part of the paper and clarify the conditions behind the bottleneck. ^[S01]

The economic implications are substantial. Can a software layer or a manufacturing technology retain market power when customers invest heavily? Can users bring the technology in-house or switch supplier? Do entrants have access to the data and computing resources required to compete? Those questions direct an investigation towards contracts, switching costs and competition, rather than the count of technically automatable tasks alone.

Recent studies make different modeling choices. The paper by Korinek and coauthors, released by the Anthropic Institute in September 2026, explores capability paths between 2026 and 2030. Its scenarios reflect the assumptions of that corporate-institute paper. Acemoglu’s framework starts more directly from exposed tasks and achievable cost savings. The treatment of research itself separates these exercises before any headline numbers are compared. ^{[S02] [S03]}

Jones and Tonetti’s May 2026 version emphasizes another constraint: slow-improving complementary tasks can hold back aggregate gains. A very fast tool cannot transform an entire activity while an indispensable step remains slow. This supply-side channel can coexist with a constraint on innovation returns, but the two are distinct. ^[S04]

The next figure therefore keeps four scenarios from the same paper, at the same horizon. It avoids a dramatic league table assembled from models with different calendars, baselines and variables.

Method and scope

Source: BIS WP 1381, table 2, printed page 33, PDF page 35. Horizon: twenty years after activation, with no calendar year or probability assigned. Annual growth: 2.5; 2.3; 3.9; 5.5%. Wage gaps against a no-AI trajectory: +7.4; −14.5; −43.9; −54.2%. The wage comparison uses that counterfactual path. Growth at year twenty differs from the model’s eventual trend. The 19.8% share in the original table concerns tasks, a different measure from labor’s income share.

Wages are compared with a no-AI trajectory, so a negative gap records a shortfall against that reference path; wages could still rise from their initial level. Output growth at year 20 describes the pace along the simulated transition, while the eventual trend belongs to its long-run state. Each value retains the model’s horizon and assumptions. ^[S01]

Their value lies in illustrating a possible separation between production and labor income. Moll, Rachel and Restrepo develop this distributional dimension of automation through wages and wealth returns. ^[S12] Our investigation of AI, wages and the power to walk away examines field evidence and bargaining mechanisms. It complements the theoretical paths presented here.

Evidence that survives the move to scale

A measured improvement in a task is useful evidence. Extrapolation requires knowing the task’s share in the wider activity, the quality checks it needs and the complementary resources it consumes. The aggregate result also depends on actual adoption and the ability to sell or use the time released.

Brynjolfsson, Li and Raymond’s study, published in 2025, follows 5,172 support agents at one firm and finds an average 15% increase in issues resolved per hour, with heterogeneous effects across workers. The assistant uses GPT-3, and the main rollout took place in fall 2020 and winter 2021. The result concerns a particular occupation, tool and organization. Extrapolating it to an economy’s wage bill would require much broader evidence. ^[S05]

METR studied 16 experienced developers working on 246 real tasks between February and June 2025. With the tools available then, AI access increased average completion time by 19%. The outcome is elapsed work time, whereas the support study measures hourly output. Their percentages therefore use different denominators. ^[S06]

The February 2026 update is essential. METR explains that the follow-up suffers from participation and task-selection problems, including developers’ reluctance to work without AI, and difficulties timing parallel work. The researchers think tools have become more helpful while considering the new estimate unreliable. The initial trial therefore needs to be read alongside the limitations of the follow-up. ^[S07]

Method and scope

Support: QJE study published in 2025, 5,172 agents, GPT-3 assistant rolled out mainly in fall 2020/winter 2021, +15% in issues resolved per hour on average. Developers: METR trial randomized across 246 tasks, 16 experienced participants, February–June 2025, +19% in time. Follow-up begun in August 2025 and published on 24 February 2026: participation, task selection and parallel work limit current estimation. The two results retain their respective units, populations and protocols.

In another survey, published on 11 May 2026, METR records perceived gains among 349 technical workers surveyed between February and April. These self-reports concern the value and speed of work. Convenience sampling and respondents’ own estimates of work without AI give them a descriptive scope, distinct from an experimentally measured causal effect. ^[S17]

A further trap is declaring failure before complementary investments have had time to pay off. Research on the productivity J-curve emphasizes organizational change, training and intangible assets. Their costs can appear before the resulting gains. That lag offers an alternative explanation for disappointing initial returns. The eventual realization of those gains remains the decisive test. ^[S11]

The most consequential question for sustained acceleration may concern research itself. A proposal published by METR on 14 October 2026 describes how an assessor might investigate acceleration and separate AI’s contribution from compute, data and human resources. That measurement agenda aims to establish automation’s contribution and the conditions for a possible feedback loop in research. ^[S08]

The resulting economic test would be demanding: how much verified progress is achieved for a given quantity of real inputs? Writing experiment code faster is insufficient when validation takes longer, failed trials multiply or compute consumption grows more quickly. Conversely, robust advances achieved with fewer resources would weaken the proposition that invention costs must keep pace with the expanding economy.

Rates, future income and asset values

The debate eventually reaches financial prices. Less attractive prospective returns can alter investment choices and the conditions that balance saving against spending. An IMF working paper published in October 2025 studies how digital capital and its relationship with labor affect the natural rate of interest. The results depend on production assumptions. ^[S13]

The natural real interest rate, or r*, is a short-term real interest rate consistent with an economy operating sustainably at its potential and stable inflation. It is estimated using models. The required return on a share contains additional components, including risk; each measure answers a different question. ^[S14]

The valuation example below makes the ambiguity tangible. A fictional asset pays 100 real units next year, after which its cash flows grow indefinitely at 3%. At an 8% required return, it is worth 2,000 units. Reducing that required return to 7%, with everything else unchanged, raises the value to 2,500. But if cash-flow growth simultaneously drops to 1%, the value falls to about 1,667.

Method and scope

Perpetual-growth model with discrete annual rates: P = D₁/(r−g), for r > g. D₁ = 100 real units in year 1. Initial case: r = 8%, g = 3%, P = 2,000. Then r = 7%, g = 3%, P = 2,500. Finally r = 7%, g = 1%, P = 1,666.7. Successive changes: +500 then −833.3; overall change about −16.7%. This asset’s required return r includes its risk; r* is an estimated natural real rate. Every parameter is fictional.

This arithmetic illustration uses entirely fictional parameters. It shows how weaker future income can outweigh the benefit of cheaper financing. A higher risk premium or larger reinvestment requirements would change the result further. Moving from the mechanism to a financial diagnosis would require company-specific accounts, payment schedules and contractual obligations.

Redistribution calls for similar care. Redistributing some capital income may support particular end markets. The aggregate effect depends on financing, how recipients use the money and supply responses. Connecting an income-distribution change to total consumption therefore requires specifying those behaviors. It may also serve a distributional purpose regardless of its effect on research. ^[S12] In the BIS framework, changing the identity of consumers leaves the bottleneck in place when the total consumption share stays unchanged. ^[S01]

The next test is the cost of discovery

The investigation separates questions that are often collapsed together. A tool can perform a useful task. A business can extract a net benefit. The inventor can retain a sufficient reward. The economy can convert those results into sustained income growth. Each transition requires additional evidence, rather than a more ambitious technical promise.

Testing the demand constraint would require revenue from profitable customer uses, the persistence of margins after competitors arrive and the full cost of validated advances. Testing the optimistic response would require showing that cheaper invention or wider uses actually offset the erosion of rents. These investigations should proceed together.

As of 16 October 2026, the reviewed research offers conditional mechanisms and localized findings. Economy-wide exhaustion of AI’s end markets remains a hypothesis to test. The BIS paper directs attention towards progress’s capacity to finance its own continuation. The decisive variable may be less the size of the next investment than the income, or saving of real resources, generated by the last one.

Sources and method

The numerical l0g examples are fictional. The six notes below the figures state their assumptions, units and formulas; the consumption and valuation cases can be recalculated directly from the displayed inputs. Figure 3 uses continuous rates, while figure 6 uses discrete annual rates. Figure 4 transcribes table 2 of the BIS paper, checked visually, with no rerun of the full model. Field studies cover distinct periods, tools and populations. The references below identify the relevant versions, periods and limitations. Documentary and arithmetic checks completed on 16 October 2026.

1. S01. Transformative AI and demand bottlenecks : Phurichai Rungcharoenkitkul, WP 1381. Published 12 October 2026; manuscript dated 10 October. §§ 2.4–2.5, table 2 p. 33, appendices A.3/C.2. Full text consulted. Conditional simulations attributed to the author; full model not rerun.
2. S02. The Simple Macroeconomics of AI : Working Paper 32487. May 2024 working paper, published in Economic Policy in 2025. NBER PDF consulted. Task exposure and cost-saving framework over ten years; productivity-level effects and annual growth rates use different units.
3. S03. Economic Scenarios for Transformative AI : Korinek, Jones, Sacher, Cotter & McCrory. Anthropic Institute WP 2026-02, September 2026, author-hosted copy. Full text consulted. Corporate-institute scenarios for 2026–2030, with their own assumptions.
4. S04. Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion. Version 0.5, May 2026. Full text consulted. Task complementarities and supply constraints, a distinct framework from private research returns.
5. S05. Generative AI at Work : Brynjolfsson, Li & Raymond. QJE, online 4 February 2025, May 2025 issue. Full text consulted, especially § IV.A and table II. 5,172 agents; GPT-3 rollout mainly fall 2020/winter 2021; average 15% rise in issues resolved per hour, with heterogeneous effects.
6. S06. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. 10 July 2025; fieldwork February–June 2025. Primary publication consulted: 16 experienced developers, 246 tasks, +19% completion time. Read with follow-up S07; tools, tasks and outcome are specific to this trial.
7. S07. We are Changing our Developer Productivity Experiment Design. 24 February 2026; follow-up begun August 2025. Primary publication consulted. Participation, task selection and timing of parallel work limit estimation of a current effect.
8. S08. How an assessor could investigate evidence of acceleration from the automation of AI R&D. 14 October 2026. Methodological proposal consulted. An agenda for attributing research acceleration to AI and other inputs; an open empirical question.
9. S09. Depreciation of Business R&D Capital : BEA Working Paper 2016-5. BEA WP 2016-5, 1 August 2016. PDF consulted: Compustat 1989–2008 and BEA/NSF industry data 1987–2007. Historical economic depreciation of R&D; the l0g example uses its own fictional parameters.
10. S10. Market Size in Innovation: Theory and Evidence from the Pharmaceutical Industry. NBER WP 10038, October 2003; QJE, August 2004. Published primary abstract consulted. Pharmaceuticals and demographic changes in demand; used qualitatively in the AI discussion.
11. S11. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. NBER WP 25148, October 2018, revised January 2020; AEJ Macroeconomics, January 2021. Primary record and abstract consulted. Complementary intangible investment and delayed measured gains.
12. S12. Uneven Growth: Automation’s Impact on Income and Wealth Inequality. NBER WP 28440, February 2021; Econometrica, November 2022. NBER PDF consulted. Income and wealth distribution model, distinct from contemporary layoff statistics.
13. S13. From Servers to Rates: AI, ICT Capital, and the Natural Rate : WP 2025/224. IMF WP 2025/224, 31 October 2025. Official PDF consulted, especially pp. 6 and 15. Model estimated on US data from 1980 Q1 to 2024 Q2; natural-rate response conditional on capital–labor relationships.
14. S14. Measuring the Natural Rate of Interest. Methodological page accessed 16 October 2026. Definition of an estimated short-term real rate consistent with sustainable activity and stable inflation. No current numerical estimate quoted.
15. S15. Gross Domestic Product (GDP) : Glossary. Methodological definition accessed 16 October 2026. Investment is included in final expenditure. The l0g examples are entirely fictional stylized economies.
16. S16. What are the Annual I-O Accounts and what do they include?. Record dated 10 March 2006, accessed 16 October 2026. Intermediate uses, final demand and value added. No current industry matrix reproduced.
17. S17. Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity. 11 May 2026; survey of 349 technical workers in February–April 2026. Primary publication consulted. Convenience sample and self-reported work value and speed: descriptive scope, distinct from causal estimation.