A free market solution to AI benefit sharing

How monetary policy and good deflation could distribute the benefits of AGI

Priors

Journalist Shakeel Hashim observed after attending The Curve:

While the rest of the world worries about whether AI is a bubble about to pop, I spent last weekend with many of the people working at the frontier of AI discussing what happens if the technology fulfills its potential.

One group is talking about the circularity of AI investment deals. The other is talking about the circularity of recursive self-improvement.

One group is asking whether the market can survive an AI crash. The other is asking whether democracy can survive an AI boom.

Say we do get that boom. How should our economic machinery respond?

In 1982, Milton Friedman wrote, “When … crisis occurs, the actions that are taken depend on the ideas that are lying around.” There is worthwhile debate to be had about whether artificial general intelligence (AGI) is coming and when. But if, as many technologists believe, transformative AI lies on the near horizon, then our real deficit is not technical foresight but social, political, and economic imagination.

We do not yet have enough thinking about how abundance should work—how prices, wages, ownership, and meaning function in a world where machines can do almost everything humans can. My hope is that more people from the humanities and social sciences work on this seriously, not to forecast capabilities but to shape the institutions that will receive them.

This blog post takes labor-replacing AGI as a given and explores one answer: that monetary policy, specifically allowing “good deflation,” offers an elegant free market solution to distributing the gains from advanced AI.

Let us begin with an older book.

Less Than Zero

The macroeconomist David Beckworth shared a book with me recently: Less than Zero by George Selgin, published in 1997. The title reminded me of the Bret Easton Ellis novel, but instead of cocaine and ennui, this one is about deflation and central banking.

The deflationary periods I am familiar with are the Great Depression (1929–1939) and the Great Recession (2007–2009), both remembered as times of collapse and misery. Selgin points out that this is only one kind of deflation—the demand-driven kind, when spending collapses, incomes fall, and prices drop because nobody is buying. He reminds us there is another kind: supply-driven deflation, when productivity rises, the economy produces more with the same resources, and prices fall because the pie is getting bigger.

Selgin challenges the view that zero inflation (or 2 percent, these days) is the ideal target for monetary policy. Under the right circumstances, he argues, the ideal inflation rate is less than zero!

Keeping the price level constant means that whenever productivity improves, the central bank must create new money to stop prices from falling. That new money doesn’t just sit there—it flows into credit markets, pushing interest rates below their natural level. Businesses respond as if real resources had suddenly become more abundant, taking out cheap loans and launching projects that would not have been profitable otherwise. But because no one has actually saved more or produced more inputs, these projects are built on false signals. When prices and interest rates eventually adjust, many of those investments prove unsustainable, producing the painful cycle of boom and bust.

Selgin’s alternative is the productivity norm: keep overall spending steady, but allow prices to fall when productivity rises and rise when supply shrinks. Let deflation happen in good times, so prices can drift downward in step with real progress. Far from slowing the economy down, this kind of good deflation sharpens price signals, reduces the need for disruptive wage changes, and prevents central banks from arbitrarily holding interest rates down.

Selgin reminds us that the Long Depression of 1873–1896, occurring just after the U.S. Civil War, was in fact a period of strong economic expansion. Prices fell nearly 30 percent, but output and real wages rose. Deflation is harmful when it is sudden and unanticipated, as in a monetary contraction or credit crisis, not when it reflects productivity gains.

Selgin ends by suggesting that nominal income targeting could serve as a practical policy rule, allowing central banks to hold spending growth steady without freezing prices in place. Economists measure spending growth using nominal GDP, which is the dollar value of everything produced, including changes in prices. (Real GDP takes out those price changes.) If nominal GDP is kept on a steady growth path, prices will rise or fall naturally depending on how quickly real output is growing.

Beckworth on Central Banking in the Age of AGI

These days, everything you read ends up sounding like it’s about AGI.

In a recent Substack post (see final section titled “The Amazing Secular Deflation Future of the Federal Reserve”), David Beckworth explores what optimal monetary policy might look like in a future world where artificial general intelligence drives exceptionally high productivity growth, say, 6% real GDP growth per year.

Beckworth writes that conventional 2% inflation targeting would be a mistake. Forcing prices to rise in such a high-growth world would require aggressive monetary expansion, risking asset bubbles and distorted investment signals.

Instead, he proposes stabilizing nominal income growth at, say, 4% a year. As AGI pushes real output up by 6%, that would mean prices falling by about 2% to keep spending and production in balance. This is not the destructive deflation of a recession. Spending and incomes are still climbing predictably, so wages and debt payments remain sustainable. Every dollar simply buys a bit more each year, which is exactly what you’d want in a world where technology is making everything cheaper to produce.

If AGI ushers in an era of abundance, prices should be allowed to show it. The challenge for central banks won’t be stopping deflation but knowing when it signals progress rather than pain. A rule-based nominal GDP target would keep spending on course, let prices fall with productivity, and let society share fully in the gains of cheaper goods and services.

AI Benefit Sharing through Good Deflation

I told David that this is a project I knew by a different name. AI benefit sharing is a growing field, spurred by projections of explosive economic growth from AGI and concerns about job displacement. The idea is to make sure AGI’s gains do not concentrate in a handful of firms or countries but are widely shared: by pooling resources like compute and data, broadening access to advanced AI systems, or redistributing part of the windfall from AI-driven growth. Policy ideas here include the Windfall Clause, which commits AI firms in advance to donate a large share of extraordinary profits once they exceed a set threshold; universal basic income, funded through taxes or dividends on AI-driven returns to provide everyone with a baseline of economic security; and diffusion frameworks, which govern how AI models, compute, and know-how are shared internationally to balance security with equitable access.

David told me that monetary policy which allows for good deflation is a simpler, more elegant solution. Taxing AGI-driven growth adds an extra layer, a redistribution mechanism built on top of the economy, whereas simply letting prices fall naturally would spread the gains of abundance to everyone automatically.

Good deflation is an excellent free market solution to AI benefit sharing! But we need to make sure that our vision of productivity-led deflation under nominal GDP targeting is robust to sectoral differences, market concentration and monopoly power, uneven international diffusion, and what happens if income goes to zero for a portion of the population.

Sectoral Differences

Even in an AGI world, productivity gains won’t be uniform. Cognitive and digital sectors may see explosive improvement, while agriculture, construction, and other physical industries advance more slowly. This means relative prices will shift dramatically: software might become nearly free while housing costs persist. It might be incredibly cheap to sue someone but expensive to eat out.

Moravec’s paradox helps us think about this—the observation that tasks requiring high-level reasoning, like chess or financial modeling, are relatively easy to automate, while sensorimotor skills that seem effortless to humans, like loading a dishwasher or making coffee, are surprisingly difficult for machines.

But honestly? Sectoral differences might not be that big of a problem. In wealthy countries, if food and housing become slightly more expensive while education, entertainment, and other sectors collapse in cost, that is still an enormous welfare gain. Most of what we spend money on would become radically cheaper.

Market Concentration and Monopoly Power

The Selgin-Beckworth vision assumes vigorous competition: many firms adopting AI, competing away profits, passing productivity gains to consumers through lower prices. But AGI is different from past technologies. It is not a tool for making specific tasks more efficient; it is a general technology that can perform all cognitive tasks.

AGI means there are no niches. The same system that writes code will diagnose diseases, manage supply chains, and argue court cases. Why would anyone hire second-tier expertise when the best intelligence exists? This may lead to radical vertical integration. We could see an OpenAI law firm, a DeepMind hospital chain, an Anthropic consulting practice. Not just AI companies selling tools, but AI companies becoming every industry.

But it might not even be an oligopoly—it could be a monopoly. Richard Sutton’s Bitter Lesson from 2019, which has held up thus far, shows that throwing more compute at a problem beats clever algorithms every time. What ultimately matters isn’t domain expertise or specialized tricks, but scale.

Trevor Chow lays out the brutal economics of scaling: GPT-4 cost around 63 million dollars to train. The jump from GPT-4 to GPT-4.5 required about 100 times more effective compute. If we want to maintain that same rate of capability improvement, doubling effective compute every year through a mix of algorithmic progress and raw compute, then by 2028 we would need roughly 300 times the compute that GPT-4 used—a 19 billion dollar training run. But that is just the training cost! The data center to run such training needs to amortize its costs across multiple training runs over its lifetime. Converting training costs to data center capex, Chow estimates a 2028 training run would require an upfront data center capex of between 152 and 304 billion dollars.

The median top model on OpenRouter stays competitive for just three weeks before something with more compute behind it displaces it. This creates consolidation pressure that pushes from many labs to a few, then from a few to two, and potentially from two to one. Only the deepest pockets (or pocket!) survive.

In that world, we do not get price-taking competitors; we get one entity setting prices across the entire economy. They capture productivity gains as monopoly rents instead of passing them through as lower prices. Good deflation stalls out because prices do not fall with costs. Nominal GDP targeting would still stabilize spending, but instead of transmitting abundance, it just accommodates monopoly pricing power.

This does not invalidate the monetary framework, but it does mean competition policy becomes critical. We might need aggressive antitrust enforcement, interoperability requirements, or other measures to prevent lock-in. The free market solution only works if we actually have free markets.

Uneven International Diffusion

Countries will not adopt AGI at the same pace. Some will experience rapid productivity growth and deflation while others lag, creating global imbalances. Capital will flow toward high-growth regions, exchange rates will shift, and the benefits of abundance may concentrate geographically. Even today, we see hints of divergence—America embraces stablecoins to facilitate digital commerce while Europe restricts them to protect traditional banking systems. These policy choices suggest different levels of readiness to absorb the economic changes AGI will bring.

As Anton Leicht observes, grand AI strategies from the US and China do not automatically translate to benefits for other nations. Middle powers must choose between expensive US frontier models with strict controls or cheaper Chinese alternatives. Low-income countries might be excluded entirely if compute remains scarce and models stay restricted.

The productivity norm works within national borders, but international coordination remains unsolved. Maybe exchange rates adjust smoothly and spread prosperity, or maybe we get technology blocs and AI colonialism.

What if income goes to zero for a segment of the population?

The productivity norm assumes that falling prices benefit everyone, but what if AGI doesn’t just change relative wages but eliminates entire categories of work, and displaced workers cannot find new jobs fast enough? OpenAI has explicitly stated its mission is to build AGI that can “outperform humans at most economically valuable work.”

Right now, we are in the augmentation phase. People are using AI to do more and earn more—lawyers handling more cases, programmers shipping more code, analysts processing more data. I want to see this period last as long as possible. But at some point, as models improve and we develop better RL environments and agent scaffolding, the equation would tip from augmentation to automation. When that happens, if a significant portion of the population sees their income approach zero, cheaper goods become irrelevant. You cannot buy discounted products with money you do not have.

This is where the elegance of the free market solution meets its hardest constraint. The productivity norm ensures that total spending grows predictably, and competition hopefully ensures that productivity gains translate into lower prices. But it says nothing about how that spending is distributed. If AGI concentrates income among capital owners and a narrow slice of AI-complementary workers, we get a peculiar form of abundance: shelves stocked with cheap goods that many cannot afford.

Again, this does not doom the productivity norm. Good deflation could still form the foundation of benefit-sharing, but it would need complementary mechanisms. The key insight from Selgin and Beckworth remains valid: letting prices fall with productivity is better than forcing them up through monetary expansion. But in an AGI world, we need to ensure that everyone retains enough income to benefit from those falling prices. Otherwise, we risk creating a peculiar dystopia where material abundance exists but remains tantalizingly out of reach for those whose work the machines have replaced.

The time to develop institutional responses to transformative AI is now, while we can think clearly, not later when crisis demands immediate action.

Acknowledgments

I am grateful to my Mercatus mentors and colleagues for their valuable feedback on earlier drafts. Special thanks to David Beckworth for sharing both George Selgin’s book and his own thinking on monetary policy under AGI—this blog post grew directly from our conversations. Thanks also to Joanna Wiaterek, Maria Kostylew, and Rudolf Laine for valuable conversations.