The narrative from the AI labs is dazzling: build AGI, unlock astonishing productivity, and watch GDP surge. It’s a compelling story, especially if you’re the one building or investing in the new thought machines. But it skips the part that makes an economy an economy: circulation.
An economy is not simply production. It is production matched to demand, and demand requires broadly distributed purchasing power. When we forget that, we rediscover an old truth the hard way: You can’t build a prosperous society that leaves most people on the sidelines.
In The Marriage of Heaven and Hell, the visionary poet and painter William Blake (writing during the first Industrial Revolution) put the circulatory logic perfectly: “The Prolific would cease to be prolific unless the Devourer as a sea received the excess of his delights.” In other words: Output has to be consumed. The system has to flow.
Image created with Gemini and Nano Banana Pro
Today, many AGI narratives assume that the “prolific” can keep producing and the broad mass of customers (“the devourer”) somehow continue to buy, even as more and more human labor is displaced and labor income and bargaining power collapses. That’s not a future of abundance. It’s a recipe for a kind of congestive heart failure for the economy: Profits and capabilities accumulate in what should be the circulatory pump, while the rest of the body is starved.
So if we want an AI economy that makes society richer, we need to ask not just “How smart will the models get?” and “How rich will AI developers, their investors, and their immediate customers get?” but “How will the value circulate in the real economy of goods and services?” Not “What can we automate?” but “What new infrastructure and institutions are needed to turn capability into widely shared prosperity?”
Two versions of the future are often discussed as if they are separate. They’re not.
The Discovery Economy: Capability Is Not GDP
I’m excited by the discovery potential of AI. It may help us solve problems that have defied us for decades: energy abundance, new materials, cures for diseases. As Nick Hanauer and Eric Beinhocker put it so well, “Prosperity is the accumulation of solutions to human problems.” That AI can grow the store of solutions to human problems is a wonderful dream, and it should be our goal to make it come true.
But discovery alone is not the same thing as economic value, and it certainly isn’t the same thing as widely shared prosperity. Between discovery and economic value lies a long, failure-prone pipeline: productization, validation, regulation, manufacturing, distribution, training, and maintenance. The valley of death is not a metaphor; it is a bureaucratic, technical, and financial landscape where many promising advances go to die. And from that valley of death, the path follows either an ascent to the broad uplands of shared prosperity, or a shortcut to a dead-end peak of wealth concentration.
If AI accelerates discovery but doesn’t accelerate diffusion, we get headlines and paper wealth, but broad-based growth takes much longer to arrive. We get a taller peak, not a wider plateau.
The distribution question begins with choke points. Who owns the discovery engines? Who controls access to compute, data, and the models themselves? Who captures the IP? Who has the channels to bring new capabilities to market? To what extent do incumbents and the moats they have built restrict innovation? Do government regulatory processes also speed up, or do they keep AI adoption at a glacial pace? Do those at the choke points use their market shaping power wisely? If those choke points are tight, the discovery economy becomes a kind of discovery feudalism: The breakthroughs happen, but the spillovers are limited, adoption is slow, and the returns concentrate.
If, on the other hand, the tools and standards of diffusion are broadly available, if interoperability is real, if licensing is designed for many routes to market, if regulatory processes can also be sped up with AI, then the discovery economy can become what we want it to be: a generalized engine of progress. There’s a huge amount of work to be done here.
Many of the questions are economic. If discovery becomes cheap, does the rest of the pipeline get cheaper, or does it get more expensive to compensate for other lost revenue? The happy dream is that a cancer vaccine becomes available at the marginal cost of production. The unhappy reality may be that the drug manufacturers conclude “We have to price this high to make up for our losses from the existing drugs that people no longer need to buy.” Even in an age of cheap discovery, it is possible that some vaccines will still cost millions of dollars per dose and only be available to people who can afford them.
The Labor Replacement Economy: Demand Is the Constraint
The other story is labor replacement. We are told that AI will substitute for a great deal of intellectual work, much as machines replaced animal labor and much of human manual labor. Businesses become more efficient. Margins rise. Output increases. Prices fall and spending power increases for those who are still employed.
But who are the customers when a large number of humans are suddenly no longer gainfully employed?
This is not a rhetorical question. It is the central macroeconomic constraint that much of Silicon Valley prefers not to model. You can’t replace wages with cheap inference and expect the consumer economy to hum along unchanged. If the wage share falls fast enough, the economy may become less stable. Social conflict rises. Politics turns punitive. Investment in long-term complements collapses. And the whole system starts behaving like a fragile rent-extraction machine rather than a durable engine of prosperity.
In a 2012 Harvard Business Review article, Michael Schrage asked a powerful strategic question: “Who do you want your customers to become?” As he put it, the answer to that question is the true foundation of great companies. “Successful companies have a ‘vision of the customer future’ that matters every bit as much as their vision of their products.”
In the early days of mass production, Henry Ford reportedly understood that if you want mass markets, you need mass purchasing power. He paid higher wages and reduced working hours, helping to invent what we now call the weekend, and with it, the leisure economy. The productivity dividend was distributed in ways that created new customers.
Ford’s innovation had consequences beyond the factory gate. Mass adoption of cars required a vast extension of infrastructure: roads, traffic rules, hotels, parking, gas stations, repair shops, and the entire social reorganization of distance. The technology mattered, but the complements made it an economy.
Steven Johnson tells a related story in his book Wonderland. The preindustrial European desire for Indian calico and chintz helped catalyze modern shopping environments and global trade networks. But there’s even more to that story. When it became cheaper to make cloth, fashion, taste, and the democratization of status display became a larger part of the economy. The point is not “consumerism is good.” The point is that economies grow because desires and capabilities change as the result of innovations, infrastructure, and institutions that allow the benefits to spread. New forms of production require new systems of distribution, experience, and exchange.
AI is at that inflection point now. We may be building the engines of extraordinary productivity, but we are not yet building the social machinery that will make that productivity broadly usable and broadly beneficial. We are just hoping that they somehow evolve.
This failure of insight and imagination is the Achilles’ heel of today’s AI giants. They imagine themselves as contestants in a race to be the next dominant platform, with the majority of the benefits going to whoever has the smartest model, the most users, and the most developers. This is not unlike the vision of Marc Andreessen’s Netscape in the early days of the web. Netscape sought to replace Microsoft Windows as the platform for users and developers, using the internet moment to become the next monopoly gatekeeper. Instead, victory went to those who embraced the web’s architecture of participation.
Now, it is true that 30 years later, we are in a world where companies such as Google, Apple, Amazon, and Meta have indeed become gatekeepers, extracting huge economic rents via their control over human attention. But it didn’t start that way. Amazon and Google in particular rose to prominence because they solved the circulation problem. Amazon’s flywheel, in which more users draw in more suppliers with more and cheaper products, which in turn brings in more users, in a virtuous circle, is a great example of an economic circulation strategy. Not only did Amazon drive enormous consumer value, they created a whole new set of suppliers.
So too, Google’s original search engine strategy was also deeply rooted in the circulation of value. As Larry Page put it in 2004, “The portal strategy tries to own all of the information….We want to get you out of Google and to the right place as fast as possible.” The company’s algorithms for both search and ad relevance were a real advance in market coordination and shared value creation. Economists like Hal Varian were brought in to design advertising models that were better not only for Google but for its customers. Google grew along with the web economy it helped to create, not at its expense. Yes, that changed over time, but let’s not forget how important Google’s support for a circulatory economy was to its initial success.
Google also provides a really good example of mechanism design to solve problems with rights holders that have economic lessons for today. When music companies sent takedown notices to YouTube for user-generated content that made unauthorized use of their IP, YouTube instead asked, “How about we help you monetize it instead?” In the process it created a new market.
The extent to which Amazon and Google seem to have forgotten these lessons is a sign of their decline, not something to be emulated. It provides an opportunity for those (including Google and Amazon, if they recommit to their roots!) who are building the next generation of technology platforms. Build a flywheel, enable a circulatory economy. AI should not be enshittified from the beginning, prioritizing value capture over broadly based value creation.
Decentralized Architectures Create Value; Centralization Captures It
An important lesson from the internet technology revolution of the 1990s and early 2000s is that decentralized architectures are more innovative and more competitive than those that are centralized. Decentralization creates value; centralization captures it. The PC decentralized the computer industry, ending IBM’s chokehold on competition during the mainframe era. The new software industry exploded. Over the next few decades, as it became dominant, Microsoft recentralized the industry by monopolizing operating systems and office applications in the way that IBM had monopolized computer hardware. The personal computer software industry began to stagnate, until open source software and the open protocols of the internet undermined Microsoft’s centralized control over the industry and ushered in a new era of innovation.
The tragedy began again, as those who had once flourished as internet innovators in turn began to prioritize control, raising moats and extracting rents rather than continuing to innovate, leading to today’s internet oligopoly. This, of course, is what allowed the current AI revolution to happen as it did. Google invented the transformer architecture, and then published it freely, but did not itself fully explore the possibilities because it was protecting an existing business model. So it was left to OpenAI to invent the future.
However, the AI revolution has a significant difference from the early internet. The U.S.’s current set up of large, closed models, enormous data centers for model training, and a highly concentrated cloud market has echoes of central planning, in which a small cadre of deep pocketed investors choose the winners at the outset rather than discovering them through a period of intense market competition and finding product-market fit (which involves finding products and services that users not only want but are willing to pay for at less than the cost of production!).
Market competition is important to ensuring that the economy is not reliant on a handful of firms reinvesting their profits into production. When this becomes the case, circulation can get cut off. Profits stop being reinvested and instead become hoarded, trapped within the sphere of financial circulation, from dividends to share buybacks to more dividends and less and less to investment in fixed or human capital.
If we are to realize the full potential of AI to reinvigorate and reinvent the economy, we need to embrace decentralized architectures. This might involve the triumph of lower-cost open weight models that commoditize and decentralize inference, and it also certainly entails protocols and technical infrastructure that can reduce the inherent concentrating tendencies of economies of scale and other technological moats that make concentration a more efficient mode of production.
Centralization is an advantage in a mature economy; it is a disadvantage when you are trying to invent the future. Premature centralization is a mistake.
A Manifesto for a Circulatory AI Economy
If AI labs wish to be architects of a prosperous future, they must work as hard on inventing the new economy’s circulatory system as they do on improving model capabilities. They need to measure success by diffusion, not just capability. They have to treat the labor transition as a core problem to be solved, not just studied. They have to be willing to win in the marketplace, not through artificial moats. That means committing to open interfaces, portability, and interoperability. General-purpose capabilities should not become a private toll road.
Companies adopting AI face their own challenges. Simply using AI to slash costs and turbocharge profits is a kind of failure. The productivity dividend should show up for employees not as a pink slip but as some combination of higher pay, reduced hours, profit-sharing, and investment in retraining. They must use the opportunity to reinvent themselves by creating new kinds of value that people will be eager to pay for, not just trying to preserve what they have.
Governments and society as a whole need to invest in the complements that will shape the new AI economy. Diffusion will be limited by the fragility of our energy grid, by bottlenecks in the supply of rare earths, but also by sclerotic approval processes for new construction or the approval of new innovations.
Governments must also develop scenarios for a future in which taxes on labor might provide a much smaller part of their income. Solutions are not obvious, and transitions will be hard, but if we face a future where capital appreciation is abundant and labor income is scarce, perhaps it’s time to consider reducing taxes on labor and increasing those on capital gains.
Over the next few months, we intend to convene a series of conversations and to publish a series of more detailed action plans in each of these areas. Let me know if you think you have ideas to contribute.
The Choice
We can build an AI economy that concentrates value, hollows out demand, and forces society into a reactive cycle of backlash and repair. Or we can build an AI economy that circulates, where discoveries diffuse, where productivity dividends translate into purchasing power and time, and where the complements are built fast enough that society becomes broadly more capable.
AI labs like to say they are building intelligence. They are making good progress. But if they want to build prosperity, they also need to discover the flywheel for the AI economy.
The prolific needs the devourer. Not as a villain, not as an obstacle, but as the sea that receives the excess, and returns it, transformed, as the next wave of demand, innovation, and shared flourishing.