Who Owns the Gains: If Machine Intelligence Raises Output, Who Receives the Returns?
Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-11-22). "Who Owns the Gains: If Machine Intelligence Raises Output, Who Receives the Returns?". SomaSoft Research. https://somasoft.ai/papers/who-owns-the-gains. Licensed under SAGL-1.0.
SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections
Who Owns the Gains
If machine intelligence raises output, who receives the returns?
Draft, 22 November 2026. Not peer reviewed. Part of a series on the transition to an economy with highly capable AI, after "The Power to Say No" and "Beyond the Price."
Reading conventions. [I] marks our own inference or argument: not sourced, offered to be disagreed with. [measured] marks a figure read off a primary document or peer-reviewed study, with the publisher and date given in the references. Where we could not verify something we say so and leave the gap open. Every number in this paper was checked against a source on 22 November 2026; the claim-by-claim list is in the accompanying facts file.
1. The question this paper asks
The earlier papers in this series asked whether a worker can refuse a bad job, and which goods should stay outside markets. Both kept returning to the same question and leaving it open: who owns the machines?
[I] If machine intelligence raises output, that output is paid out in two ways: as wages to the people who work, and as returns to whoever owns the capital, meaning the chips, data centres, models and the firms that hold them. Cash floors and bargaining change the first stream. Only ownership changes the second. If the second stream grows faster than the first, a society can become much richer while most of its members gain little.
This paper looks at the evidence on that second stream. It asks where returns have been going, who owns the means of machine intelligence today, and which mechanisms for spreading ownership have been tested rather than only proposed.
2. Where the returns have been going
The labour share is at a record low in the United States. [measured] The Bureau of Labor Statistics reported that labour's share of output in the nonfarm business sector was 52.8% in the second quarter of 2026, the lowest level since the series began in 1947. For most of the post-war period it was in the range of 60 to 66 percent.
The decline is not only American. [measured] The International Labour Organization estimated in September 2024 that if the global labour income share had stayed at its 2004 level, workers would have earned US$2.4 trillion more in 2024 alone. The share fell 0.6 percentage points between 2019 and 2022, and nearly 40% of the reduction since 2004 happened in those pandemic years. The ILO warned that generative AI could add further downward pressure.
The explanations point at technology and concentration. [measured] Karabarbounis and Neiman (2014) found a decline in the labour share across most countries and industries since the early 1980s, and attributed roughly half of it to the falling price of investment goods, which led firms to substitute capital for labour. Autor, Dorn, Katz, Patterson and Van Reenen (2020) found that the US decline is driven largely by "superstar firms": high-profit, low-labour-share firms gaining market share.
[I] Machine intelligence fits both explanations: it is capital-intensive, and it is being built by a small number of very large firms. The default expectation, unless ownership changes, is that a growing part of the gain goes to capital.
3. Who owns the means of machine intelligence
Compute is concentrated. [measured] Epoch AI estimated in April 2026 that five companies (Amazon, Google, Meta, Microsoft and Oracle) owned 70.7% of global AI compute in the fourth quarter of 2025, up from 62.5% in the first quarter of 2024.
Model building is concentrated in industry. [measured] Stanford's 2025 AI Index reported that nearly 90% of notable AI models released in 2024 came from industry, up from 60% in 2023.
The investment is enormous. [measured] Bloomberg reported in February 2026 that Alphabet, Amazon, Meta and Microsoft planned roughly $650 billion of capital spending in 2026, mostly on AI computing.
One new form of ownership deserves attention. [measured] In October 2025 OpenAI recapitalised: its for-profit arm became OpenAI Group PBC, a public benefit corporation, and the nonprofit OpenAI Foundation received a 26% equity stake valued at about $130 billion, with the right to appoint the company's board. [I] This is a large charitable claim on the returns of a frontier AI firm. Whether it becomes a mechanism for sharing gains broadly depends on how the foundation spends, who holds it to account, and whether its control survives future funding rounds. None of that can be evaluated yet.
4. Public wealth funds: one tested, one admired, several new
Alaska: the evaluated case
Alaska has paid every eligible resident a yearly dividend from the Alaska Permanent Fund since 1982. It is the closest thing the world has to a permanent, universal dividend from a shared asset.
[measured] Jones and Marinescu (2022, American Economic Journal: Economic Policy), using Current Population Survey data and a synthetic control method, found that the dividend had no effect on employment. It did increase part-time work, by 1.8 percentage points (about 17%). They interpret this as a general-equilibrium effect, consistent with a more positive employment response in non-tradable sectors such as local services.
[measured] The amount is set by politics, not by the fund's earnings. The 2022 payment was $3,284, including a one-time $662 energy relief payment. The Legislature set the 2025 dividend at $1,000, paid to more than 600,000 eligible Alaskans. The 2026 payment is $1,200, including a $200 energy relief payment, after the House passed $1,500 and the Senate pushed for $1,000.
[I] Alaska shows that a universal dividend from a shared asset can run for four decades without reducing employment. It also shows the weakness: when the state budget is tight, the dividend is what gets cut.
Norway: ownership without a dividend
[measured] Norway's Government Pension Fund Global received its first transfer in 1996. It was worth NOK 21,268 billion at the end of 2025, after a 15.1% return that year, and it owns on average about 1.5% of all listed companies worldwide. Under the fiscal rule, government spending may use only the fund's expected real return, currently estimated at 3%, not its capital.
[I] Norway is the opposite design to Alaska. Citizens receive no cheque; the returns finance the ordinary budget. It is widely admired for discipline, but it is not an evaluated distribution mechanism in the way Alaska's dividend is, and we found no study measuring what its returns do for household ownership.
2025–2026: new public stakes, untested as distribution
- Child investment accounts. [measured] Under the One Big Beautiful Bill Act, the US Treasury deposits $1,000 into a "Trump Account" for each eligible child born from 2025 through 2028, with contributions allowed from July 4, 2026. The IRS reported on 31 March 2026 that more than 4 million children had been enrolled, more than 1 million of them with elections for the $1,000 deposit.
- A government stake in a chipmaker. [measured] In August 2025 the US government took a 9.9% stake in Intel, buying 433.3 million shares at $20.47 for $8.9 billion, funded from $5.7 billion of previously awarded CHIPS Act grants and $3.2 billion from the Secure Enclave programme. Intel described the stake as passive, with no board seat.
[I] Both are public ownership. Neither is yet a mechanism that returns gains to people: the accounts are small per child and locked until adulthood, and the Intel stake's returns go to the Treasury. They are early, and nobody has evaluated them.
5. Employee ownership: modest, consistent evidence
[measured] The National Center for Employee Ownership counts 6,609 employee stock ownership plans (ESOPs) in the United States covering 15.1 million participants (2023 data).
Performance. [measured] O'Boyle, Patel and Gonzalez-Mulé (2016) pooled 102 samples covering 56,984 firms and found a small but positive and statistically significant relation between employee ownership and firm performance (r = 0.04).
Job stability. [measured] Kurtulus and Kruse (2017), studying US firms from 1999 to 2011, found that companies with employee stock ownership had about half as many layoffs as other companies in the last two recessions, and were somewhat more likely to survive. Burdín and Dean (2009), using Uruguayan social security records from 1996 to 2005, found that worker cooperatives adjusted wages rather than employment when shocks hit. Pérotin's 2016 review for Co-operatives UK reached the same conclusion across countries: worker cooperatives survive at least as long as other businesses and keep employment more stable.
Mondragon. [measured] The Mondragon cooperative group reported 71,400 people and €11.322 billion in sales for 2025. Each cooperative puts at least 13% of its profit into a shared divisional fund, and workers can be redeployed between cooperatives. That mechanism was tested in 2013, when the group's appliance maker, Fagor Electrodomésticos, sought creditor protection with about €1.1 billion of debt.
[I] The evidence says employee ownership does not hurt performance, may help it slightly, and buys real job security. It does not show that employee ownership spreads wealth at national scale: 15.1 million US participants is a minority of workers, and many ESOP stakes are small. It also faces a selection problem. Firms that choose employee ownership may differ in ways the studies cannot fully remove.
6. The Windfall Clause: a proposal
[measured] O'Keefe, Cihon, Garfinkel, Flynn, Leung and Dafoe (2020) proposed the Windfall Clause: an advance commitment by AI firms to donate a large share of any extremely large profits from transformative AI. Their illustrative schedule scales with a firm's profits as a share of gross world product (GWP), like an income tax:
| Profits as share of GWP | Share of marginal profits pledged |
|---|---|
| 0–0.1% | 0% |
| 0.1–1% | 1% |
| 1–10% | 20% |
| 10–100% | 50% |
In their worked example, a firm earning $5 trillion when GWP is $268 trillion (2010 dollars) would owe $488.12 billion.
[I] The idea is careful, and it addresses exactly the scenario this paper worries about. But it is a proposal. We found no public record of any firm signing one. It has never been triggered, and its enforceability, including whether a future board could undo it, has never been tested in court. It belongs in the "proposed" column.
7. Data cooperatives and data trusts: the thinnest record
Data cooperatives and data trusts would let people pool their data and bargain collectively over its use, including its use to train AI.
[measured] - The Open Data Institute ran three data trust pilots with the UK government, ending in March 2019, on illegal wildlife trade, food waste and public services in Greenwich. It found strong interest, but concluded that UK trust law was "inappropriate" for data trusts and that further legal work was needed. - In Toronto, Sidewalk Labs proposed an independent Urban Data Trust for its Quayside project. It dropped the idea in November 2019 in favour of oversight by Waterfront Toronto, and cancelled the project on 7 May 2020. - MIDATA, a Swiss health-data cooperative, lets members store their data and vote on research access, and its bylaws bar paying profits to members as dividends. We found no published evaluation of its outcomes or membership numbers. - A 2025 multi-case study of six data cooperatives (PACIS 2025) reported that many struggle to sustain themselves, and that research on them remains largely conceptual.
[I] As a way of sharing AI gains, the data cooperative is so far an idea with pilots, not a mechanism with a record. Individual data is worth little on its own. Collective bargaining over it may be worth more, but we found no case where a data cooperative has moved meaningful returns to its members.
8. Public compute: widening who can build
Public compute programmes do not directly share returns. They lower the barrier to building, so that more of the capability, and more of its gains, can sit outside the largest firms.
[measured] - United States. The National AI Research Resource Task Force recommended $2.6 billion over six years in January 2023, citing the concentration of computing power in large technology companies. A pilot launched in January 2024 and supported more than 800 research projects. In September 2026 NSF made a $35 million, five-year award to the San Diego Supercomputer Center, with the Texas Advanced Computing Center, to run a NAIRR Operations Center and move the pilot toward permanent infrastructure. - United Kingdom. The AI Research Resource includes Isambard-AI at the University of Bristol, with 5,448 Nvidia GH200 superchips. The government has committed an extra £1 billion to expand public compute at least 20 times by 2030. - European Union. On 30 July 2026 the Commission opened a call for up to seven AI gigafactories, with up to €10 billion of public money and at least €20 billion of private investment sought. The call closed on 12 November 2026. - Canada. Budget 2024 committed C$2 billion to a Sovereign AI Compute Strategy, including up to $300 million for an AI Compute Access Fund for smaller firms.
[I] These are real and growing, but small next to private spending. Their effect on who owns AI's returns has not been measured. They are best judged by whether projects using them produce firms, tools or public goods that would not otherwise exist.
9. The evidence, side by side
| Mechanism | Status | What has been measured |
|---|---|---|
| Universal dividend (Alaska) | Operating since 1982 | No employment loss; part-time work up 1.8 points |
| Sovereign fund (Norway) | Operating since 1996 | Fund size and returns; not distribution effects |
| Child accounts, public equity stakes | New, 2025–2026 | Enrolment counts only |
| Employee ownership (ESOPs, cooperatives) | Operating at scale | Small performance gain; fewer layoffs in recessions |
| Windfall Clause | Proposal (2020) | Nothing; no known signatory |
| Data cooperatives and trusts | Pilots | Feasibility and legal obstacles; no member returns |
| Public compute | Pilots and new programmes | Projects supported; not ownership effects |
10. Recommendations, staged
[I] All of the following are our proposals. Each stage is chosen so that it can be checked before the next is attempted.
Firms
- Now: publish how productivity gains from AI deployment are being split between pay, hours, training and profit, using the gain-share rule from "The Power to Say No."
- Within one to three years: offer broad-based employee ownership (an ESOP, an employee ownership trust or a cooperative conversion) when founders sell, so that ownership transfers to staff rather than to an outside buyer.
- For AI developers: sign a binding windfall commitment now, while it is cheap, and put enforcement in the hands of a party with no stake in the firm.
Municipalities
- Now: fund conversion advice for retiring business owners who might sell to employees.
- Within one to three years: pilot a small community wealth fund from a public asset the city already holds, such as land leases or utility revenue, paid as a per-resident dividend and evaluated against a comparison area.
- For data: run a data trust pilot with a sealed evaluation and publish the result, including if it fails.
Policymakers
- Now: commission an independent evaluation of child investment accounts and public equity stakes before they expand, with a stated measure of who benefits.
- Within one to three years: make public compute programmes permanent, and report what share of their capacity goes to small firms, universities and public-interest projects.
- Longer term: consider a public wealth fund seeded with a share of revenue or equity from AI firms, with the fund's capital protected by a spending rule like Norway's and part of its return paid out as a dividend as in Alaska, so that both discipline and a direct dividend are kept.
11. What this project can say about itself
SomaSoft is a small research project. It has no users and no revenue, so it has no gains to share and no ownership structure that has been tested by money. A benefit fund is proposed in our governance drafts; it does not exist. We report this because a paper about who owns the gains should not imply that its authors have solved the problem.
12. Limits
- This is a review of secondary evidence. We have not re-analysed any study. The figures should be checked against the cited sources.
- The evidence is mostly from the United States and Europe. Ownership mechanisms may behave differently elsewhere.
- The strongest studies concern older technologies. No mechanism here has been evaluated against gains from AI specifically.
- Selection. Employee-owned firms, successful funds and surviving cooperatives are partly self-selected. Failures such as Fagor show that ownership alone does not make a firm safe.
- Recent developments may change quickly. The 2025–2026 items above describe programmes that were weeks or months old when we checked them.
- Our recommendations are untested. We have applied none of them.
13. Conclusion
[I] Machine intelligence will raise output in some sectors, perhaps by a great deal. Whether most people share in that depends less on the technology than on who owns it. The record suggests three things. Universal dividends from shared assets can work without harming employment, but they are politically fragile. Employee ownership is a modest, reliable way to put returns and security in workers' hands. And the mechanisms designed specifically for AI, the Windfall Clause and data cooperatives, are still almost entirely untested. The honest next step is to try them at small scale, measure them, and publish what happens.
References
- Autor, D., Dorn, D., Katz, L. F., Patterson, C. & Van Reenen, J. (2020). The fall of the labor share and the rise of superstar firms. Quarterly Journal of Economics 135(2), 645–709. doi:10.1093/qje/qjaa004
- Bloomberg (6 February 2026). How much is Big Tech spending on AI computing? A staggering $650 billion in 2026.
- Bureau of Labor Statistics (15 September 2026). Labor share at its lowest level, 52.8 percent, in second quarter 2026. The Economics Daily.
- Burdín, G. & Dean, A. (2009). New evidence on wages and employment in worker cooperatives compared with capitalist firms. Journal of Comparative Economics 37(4), 517–533.
- Epoch AI (14 April 2026). Hyperscalers control most compute. epoch.ai
- European Commission (30 July 2026). EU launches AI gigafactories call.
- Innovation, Science and Economic Development Canada (2024–2025). Canadian Sovereign AI Compute Strategy; AI Compute Access Fund.
- Internal Revenue Service (31 March 2026). IR-2026-42, on Trump Account enrolment.
- International Labour Organization (4 September 2024). World Employment and Social Outlook: September 2024 Update.
- Intel Corporation (22 August 2025). Announcement of US government investment.
- Jones, D. & Marinescu, I. (2022). The labor market impacts of universal and permanent cash transfers: evidence from the Alaska Permanent Fund. American Economic Journal: Economic Policy 14(2), 315–40.
- Karabarbounis, L. & Neiman, B. (2014). The global decline of the labor share. Quarterly Journal of Economics 129(1), 61–103.
- Kurtulus, F. A. & Kruse, D. L. (2017). How Did Employee Ownership Firms Weather the Last Two Recessions? W.E. Upjohn Institute.
- MONDRAGON Corporation (2026). Annual Report 2025.
- National AI Research Resource Task Force (January 2023). Strengthening and Democratizing the U.S. Artificial Intelligence Innovation Ecosystem (final report).
- National Center for Employee Ownership (2026). ESOPs by the numbers.
- National Science Foundation (1 September 2026). NSF establishes operations center for the National Artificial Intelligence Research Resource.
- Norges Bank (2026). Annual Report 2025; Norges Bank Investment Management, About the fund.
- O'Boyle, E. H., Patel, P. C. & Gonzalez-Mulé, E. (2016). Employee ownership and firm performance: a meta-analysis. Human Resource Management Journal 26(4), 425–448.
- O'Keefe, C., Cihon, P., Garfinkel, B., Flynn, C., Leung, J. & Dafoe, A. (2020). The Windfall Clause: distributing the benefits of AI for the common good. AIES 2020; arXiv:1912.11595.
- Open Data Institute (April 2019). Report on three data trust pilots, as reported by Pinsent Masons, Out-Law, April 2019.
- OpenAI (28 October 2025). Our structure.
- Pérotin, V. (2016). What Do We Really Know About Worker Co-operatives? Co-operatives UK.
- Stanford HAI (2025). The 2025 AI Index Report.
- State of Alaska, Department of Revenue (19 September 2025; 21 September 2026). Permanent Fund Dividend amount announcements.
- UK Department for Science, Innovation and Technology (2026). AI Research Resource: advanced supercomputers for the UK.
- SomaSoft Research (2026). The Power to Say No; Beyond the Price; both at somasoft.ai/papers.
AURI's concept graph was queried on 2026-11-22 ("employee ownership"). The concept was not in the graph, so nothing in this paper comes from AURI. All figures were checked against the web sources listed above on 2026-11-22.