Who Buys What the Machines Make? AI, Robotics, Cheap Energy, and the Case for an Income Floor

15 September 2026

The question underneath most debates about AI and employment is older than AI itself, and John Maynard Keynes gave it a name in 1930: "technological unemployment... due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour" [1]. Keynes treated it as "only a temporary phase of maladjustment" — society would, he thought, always find new work for displaced hands within a generation. That assumption has held for roughly a century of prior automation waves. The question worth taking seriously now is whether it keeps holding through a wave that targets cognitive labour directly, arrives alongside genuine progress in physical robotics, and coincides with an energy cost curve that may be about to fall off a cliff. None of what follows is a confident prediction — nobody currently has one worth trusting — but the pieces are real, dated, and worth laying out together rather than debated separately.

The Affordability Paradox, Stated Plainly

TheThe mechanism the reader's question points at is a real one in economics, not a rhetorical flourish: an economy that produces goods and services with dramatically less human labour is, definitionally, more efficient at production — but every one of those goods and services still needs a buyer, and buyers overwhelmingly get their purchasing power from being paid to work. Push enough of that labour out without replacing the income it provided, and you don't get a wealthier society with more leisure; you get an economy that can make more than it can sell, because the people who'd buy it no longer have wages to spend. This is the demand side of automation that a pure productivity or GDP framing tends to leave out, and it's precisely the concern behind the "who buys what the robots make" question: efficiency in production says nothing on its own about distribution of the resulting income, and the two have historically been coupled through wages — a coupling automation weakens by construction.

The honest empirical picture as of 2026 is not yet mass unemployment — but it is a real, measured, early divergence exactly where the theory predicts one. Brynjolfsson, Chandar, and Chen's high-frequency analysis of ADP payroll data through June 2026 finds no evidence of widespread, economy-wide job displacement, but a specific and widening exception: employment of young workers (ages 22–25) in AI-exposed occupations now stands 19% below where it would be had it kept pace with less-exposed peers, a gap that has widened steadily since first documented in August 2025, concentrated specifically in occupations where AI usage substitutes for human tasks rather than complements them [2]. Goldman Sachs' base-case estimate is a 6-7% displacement rate over a ten-year adoption window, with a plausible range of 3-14% depending on adoption speed [3]; the World Economic Forum's 2030 projection is more dramatic in both directions at once — 92 million roles displaced globally against 170 million newly created, a churn rather than a clean subtraction [4]. All three sources agree on the shape even where they disagree on the number: this is starting at the edges (entry-level, substitutable, codified-knowledge work) rather than sweeping the whole economy at once, which is exactly the pattern you'd expect from a technology that is currently much better at codified knowledge tasks than at either judgement-heavy human work or physical manipulation.

Knowledge Work Versus Actuators: A Real, Currently Wide Gap

TheThe reader's instinct that there's a marked difference between pure knowledge-economy jobs and those requiring physical actuators is well-supported by the current evidence, and it matters for how fast this plays out. Generative AI's demonstrated strength is specifically in codified, text-and-code-shaped knowledge work: the Stanford paper above finds the largest employment declines concentrated in exactly those occupations, and separate labour-market data shows a 21% decrease in job postings for automatable writing and coding work compared with more manual jobs over the same period [5]. Physical-world automation is a genuinely different, older, and better-studied problem: Acemoglu and Restrepo's analysis of industrial robot adoption across US local labour markets between 1990 and 2007 found each additional robot per thousand workers reduced the local employment-to-population ratio measurably and reduced wages, with effects concentrated in exactly the routine manual and manufacturing occupations robots were installed to do [6] — a slower-moving, capital-intensive, already-underway version of the same displacement mechanism, well before generative AI existed. What's different about the current moment is the possibility of these two previously-separate automation fronts converging: general-purpose humanoid robotics is a live, heavily-funded area of development specifically because pairing a generative-AI-style planning and perception layer with a physical actuator closes the gap the earlier robot-adoption wave never crossed — a robot that could only be programmed for one repetitive task on one production line versus one that can be shown a task in natural language and adapt. That convergence hasn't arrived at scale yet, and predicting exactly when it will is not something this page can responsibly do — but treating knowledge work and physical work as separately-timed fronts of the same underlying shift, rather than two unrelated stories, is the more defensible framing.

What the UBI Evidence Actually Shows — Not a Slogan, a Genuinely Mixed Result

UniversalUniversal basic income is frequently discussed as though the empirical question were settled in one direction or the other. It isn't, and the honest picture is more useful than either the "it fixes everything" or "it wrecks incentives" versions:

• Finland (2017-2019). 2,000 unemployed adults received €560/month unconditionally for two years, the most methodologically rigorous national-scale trial to date. Employment barely changed relative to the control group — recipients did not find jobs faster or slower — while self-reported stress, health complaints, and trust in others improved measurably [7]. The clean result here is against both extreme predictions: no employment collapse, and no employment boost either.
• Stockton SEED (2019-2021). $500/month to 125 low-income adults for two years. Full-time employment among recipients rose from 28% to 40%, alongside improved mental health and reduced food insecurity [8] — the opposite direction from the "why would anyone work" objection, plausibly because a income floor let people afford the costs (transport, childcare, time to interview) of taking a better job rather than the nearest available one.
• Kenya (GiveDirectly, ongoing since 2017). Over 20,000 people receiving guaranteed transfers for 12 years — the largest and longest-running study of its kind. Beyond individual effects, researchers found positive local-economy spillovers: businesses near recipient villages grew and created jobs for non-recipients too, evidence that a large enough transfer can expand the local economy rather than simply redistribute a fixed pie [9].
• OpenResearch (2019-2024, funded by Sam Altman). $1,000/month to 1,000 people across Texas and Illinois for three years against a $50/month control group — deliberately funded by an AI company's leadership rather than a government. Recipients didn't leave the workforce, though they worked slightly fewer hours; money went disproportionately to necessities, medical care, and helping others; participants reported greater agency to start businesses or take a lower-paying job for better long-term prospects. But the early reductions in stress and food insecurity measurably faded by the study's second and third years [10] — a genuinely important caveat against reading any single pilot's early results as the permanent effect.

The pattern across all four, different countries and income levels, is consistent enough to take seriously: no catastrophic exit from work, real and repeated wellbeing gains, and — in the one case testing whether the up-front stress relief holds for three years rather than two — evidence that some of the benefit fades rather than compounds. That last point matters directly for the reader's proposal: a UBI sized and delivered as a one-off pilot is not the same intervention, tested at the same scale, as a permanent income floor sized to actually replace lost wages economy-wide, and none of the four studies above tests that harder version.

"If It Doesn't Come Through Government, Companies Will Have To"

ThisThis is the part of the reader's argument this page finds most genuinely interesting, and it already has a live, real-world test case rather than being purely speculative. Sam Altman's 2021 essay "Moore's Law for Everything" argued explicitly that AI-driven productivity growth could fund a form of universal income within roughly a decade, and specifically framed this as something that would need to happen given how the wealth AI generates would otherwise concentrate [11]. Altman didn't stop at the essay: he co-founded Worldcoin, a project that scans a person's iris to establish unique "proof of personhood" specifically so that a future income-distribution scheme could verify who's eligible without a state doing the verifying, and separately funded OpenResearch's basic-income study directly out of personal and (indirectly) AI-company wealth rather than through a public research grant [12]. This is a company (and its leadership) actively building infrastructure for exactly the scenario the reader describes: private, AI-generated wealth funding an income floor because the alternative is either state action or an economy that can't absorb what it produces.

The obvious complication is that this doesn't resolve the underlying power question, it relocates it. A state-run income floor answers to (at least nominally) democratic accountability and is funded through a tax base the state doesn't itself control the underlying technology of. A company-run one is funded and administered by the same entity that owns the AI systems doing the displacing, verified through biometric infrastructure that same company built and controls — Worldcoin's iris-scanning model has drawn exactly this criticism, from privacy regulators in multiple countries and from critics who point out that a single company controlling both your employer's replacement and your income-verification credential is a much larger concentration of power over ordinary people's lives than a tax-funded welfare state, whatever its inefficiencies [13]. The reader's framing — "companies will have to provide income if they want to continue to do business" — is plausible as an economic prediction and genuinely double-edged as a political one: it describes both a company doing something arguably responsible and a company becoming something closer to a private government, with all the accountability questions that implies.

The Energy Side: Why This Compounds Rather Than Just Adds

TheThe reader's second point is the one most discussions of AI-and-jobs leave out entirely, and it changes the shape of the argument rather than just its scale. Renewable electricity has already crossed a real cost threshold, not a projected one: onshore wind's global weighted-average levelised cost fell to roughly $33/MWh in 2025 and utility-scale solar PV held at around $44/MWh, with more than 90% of new renewable capacity commissioned in 2025 now cheaper than any new fossil-fuel plant [14]. That's already reshaping electricity markets independent of anything AI-specific. Fusion is the genuinely open, higher-upside wildcard rather than a settled fact: Commonwealth Fusion Systems' SPARC tokamak, roughly 75% complete as of 2026, is targeting net-energy operation in 2026-2027, and the company has already filed a 2026 grid-interconnection application for a planned 400-megawatt commercial plant in Virginia with power-purchase agreements reportedly signed by Google and Eni — though grid connection approval alone typically takes four to six years, and no commercial fusion plant has yet put a single watt onto any grid anywhere [15]. ITER, the flagship international research reactor, remains on a much longer public-sector timeline: first plasma around 2035, sustained fusion demonstration in the late 2030s [16].

Put together with the labour question above, the compounding effect is this: AI and robotics reduce the labour cost of production; falling renewable and (on a longer, less certain timeline) potentially fusion-cheap energy costs reduce the energy cost of production; and both together push the marginal cost of producing goods and services toward capital and energy costs alone, with labour's share of that cost falling on both fronts at once rather than one being offset by the other staying expensive. That is precisely the scenario in which Keynes' 1930 assumption — that new work always eventually absorbs displaced labour — is weakest, because the assumption relied on labour remaining the input that's hardest to substitute. A world of abundant, cheap automation-compatible energy is a world where that's no longer true on either the compute side or, if humanoid robotics closes the gap described above, the physical side either. This doesn't make the outcome bad by itself — cheaper production of necessities at scale is, in isolation, a straightforwardly good thing, which is the whole basis for viewing this as an abundance problem rather than a scarcity one. Whether it's a good outcome for the people whose labour it stops needing depends entirely on whether the resulting surplus gets redistributed as income, through a government, a company, or some structure nobody has fully designed yet — which is exactly the open question the reader is asking, and exactly the one this page cannot responsibly claim to have settled.

What This Page Is and Isn't Claiming

ThisThis isn't a forecast. The employment data above shows a real, early, narrow divergence — not economy-wide collapse — and every energy figure above describes progress, not delivered abundance; fusion in particular remains genuinely unproven at commercial scale. What the evidence does support is that the reader's underlying mechanism is real rather than speculative: automation without redistribution is a demand-side problem economists have understood since Keynes, current data shows AI-driven displacement concentrated exactly where theory predicts (entry-level, substitutable, codified knowledge work), UBI pilots consistently show meaningful wellbeing gains without triggering the workforce exodus its critics predict (while also showing some of those gains fade over a longer horizon than any pilot has yet tested at full economic scale), and there is at least one clear, real, ongoing example of a major AI company's leadership treating "we may need to fund an income floor ourselves" as a live strategic position rather than a talking point. Whether that adds up to state-run UBI, company-run income schemes, some hybrid, or a poorly-managed transition with real and unevenly distributed hardship is not a question the evidence currently in front of anyone can answer with confidence — which is exactly why it's worth tracking closely rather than assuming either the optimistic or the alarming version by default.

References

1. Keynes, J. M. (1930). "Economic Possibilities for our Grandchildren." The Nation and Athenæum, 11 & 18 October 1930.
2. Brynjolfsson, E., Chandar, B., & Chen, R. (2026). "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab, August 2026. https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf
3. Goldman Sachs (2026). "How Will AI Affect the US Labor Market?" https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-us-labor-market
4. World Economic Forum (2025). Future of Jobs Report.
5. Harvard Business Review (2026). "Research: How AI Is Changing the Labor Market." https://hbr.org/2026/03/research-how-ai-is-changing-the-labor-market
6. Acemoglu, D., & Restrepo, P. (2020). "Robots and Jobs: Evidence from US Labor Markets." Journal of Political Economy, 128(6), 2188-2244.
7. Kangas, O., Jauhiainen, S., Simanainen, M., & Ylikännö, M. (Eds.) (2020). Results of Finland's Basic Income Experiment. Finnish Ministry of Social Affairs and Health.
8. Stockton Economic Empowerment Demonstration (SEED) (2021). Final report. https://www.stocktondemonstration.org/
9. GiveDirectly (ongoing since 2017). Kenya Long-Term Basic Income study. https://www.givedirectly.org/
10. Vivalt, E., Rhodes, E., et al. (2024). Unconditional Cash and Labor Supply / OpenResearch basic income study findings. National Bureau of Economic Research working papers, July 2024.
11. Altman, S. (2021). "Moore's Law for Everything." https://moores.samaltman.com/
12. Fortune (2023). "OpenAI's Sam Altman throws support behind Worldcoin crypto project amid $50M funding." https://www.benzinga.com/markets/cryptocurrency/23/12/36269529/
13. The Register (2023). "Worldcoin fundraising." https://www.theregister.com/2023/05/16/worldcoin_fundraising
14. International Renewable Energy Agency (IRENA) / International Energy Agency (IEA), Breakthrough Agenda Report 2025 — Power. https://www.iea.org/reports/breakthrough-agenda-report-2025/power
15. Commonwealth Fusion Systems (2026); grid-interconnection filing and Fall Line Fusion Power Station announcement, PJM Interconnection, April 2026.
16. ITER Organization (2026). Project status and vacuum vessel assembly progress reports.