Every week, I consume an enormous amount of research looking for ideas that challenge how I think about markets and especially AI and its impact. I use X to help guide me toward viral research that other people find useful, which also gives me a sense of broader sentiment. This week, one of those pieces was Stijn Van Nieuwerburgh’s Brookings paper, Financing the AI Buildout. The paper applies a traditional financial framework to the economics underlying the extraordinary amount of capital currently flowing into artificial-intelligence infrastructure.
Van Nieuwerburgh estimates that the United States will invest approximately $10.3 trillion in AI infrastructure between 2025 and 2032, equivalent to an average of 3.63% of GDP annually. Under central assumptions of a 10% unlevered return and a 50% operating cash-flow margin, that infrastructure would ultimately require roughly $3.7 trillion of annual revenue by 2032 to earn an adequate return on the capital invested. That represents approximately 9% of projected U.S. GDP.
The 9% figure has understandably attracted the bear’s attention because it appears enormous relative to the current economy and when compared to other large innovation buildouts in the history of the country. But the more important question is what that calculation can tell us about the economic system that will exist in 2032. Capital-recovery analysis is extremely useful for determining the revenue required to support an investment. Its explanatory power becomes more limited when the productive capacity, cost structure, demand base, and composition of the economy itself are changing during the forecast period.
That distinction matters enormously for AI. If we have learned anything over the past year as the pace of progress has accelerated, it is that assumptions about capability, cost, productivity, and demand can change materially within relatively short periods. Investors focused on this ROIC issue therefore face two related analytical problems: estimating the economics of the infrastructure being financed today and estimating the scale and structure of the economy that will eventually use that infrastructure.
Those questions require different frameworks.
1. You Cannot Hold the Economy Still
There is a familiar analytical sequence being applied to AI. Start with current revenues, estimate the capital required to build the infrastructure, impose a required return, calculate the revenue necessary to justify the investment, and compare the result with a forecast of future GDP. The math is internally consistent, but its conclusions depend heavily on the assumptions made about how the economy evolves during the forecast period.
Game 4 of the Knicks’ NBA Finals provides a simple analogy for the limitations of linear extrapolation when the future contains a wide distribution of possible outcomes. The first quarter ended with the Knicks trailing 41–22. Multiplying those numbers by four would have produced a projected final score of 164–88. The calculation is mathematically correct, yet its usefulness as a forecast depends on assuming that the remaining three quarters unfold in roughly the same way as the first.
Real games evolve. Coaches adjust strategy, players tire, injuries occur, foul trouble affects rotations, defenses adapt, and the score itself changes incentives. New information generated during the game changes the path of the game. Time therefore contains information rather than functioning simply as a multiplier. In the end, as we now know, the Knicks won 107–106.
AI presents the same problem at considerably greater scale. Models become more capable, inference becomes cheaper, agents perform increasingly complex tasks, coding improves, robotics advances, and scientific discovery accelerates. These developments interact with one another, making the technology both more productive and more widely usable. And this occurs before considering further advances in autonomous vehicles, robotics, humanoids, or more general forms of machine intelligence.
If we knew the capability, cost, and productivity of AI in 2032, estimating the appropriate amount of infrastructure to build today would be a far easier exercise. The uncertainty surrounding those variables is therefore central to the investment problem.
The timing of the Brookings research itself illustrates the analytical challenge created by exponential technological change. The original Columbia working paper is dated March 19, 2026, and was presented at the Brookings Papers on Economic Activity conference on September 25. Six months represents a relatively short interval in most areas of economics. In AI, however, six months can encompass multiple model generations, substantial reductions in inference costs, new agent architectures, and the commercialization of capabilities that previously existed only as demonstrations.
This observation says nothing about the quality of the Brookings work. It highlights the unusual measurement problem confronting anyone studying AI: the underlying technology can change materially during the period required to analyze it.
As capability, cost, and productivity evolve, the assumptions surrounding the revenue opportunity evolve with them. Capital-recovery analysis remains essential for understanding whether investors ultimately earn an adequate return on the capital deployed. The larger uncertainty concerns the amount of economic value that will exist for participants to capture.
2. From Intelligence Liquidity to Productive Liquidity
This brings me back to an idea I explored previously in a paper I wrote last year titled QE for the Mind: How Artificial Intelligence Is Flooding the Economy with Intelligence Liquidity. I argued that AI can be understood as a new form of liquidity, with an economic effect that resembles quantitative easing in one important respect: both increase the availability of something previously constrained.
Traditional quantitative easing lowered the cost of money and increased the amount of financial liquidity available to the economy. AI lowers the cost of intelligence. Reasoning, coding, analysis, design, planning, and decision-making become cheaper and more abundant, allowing individuals and businesses to access capabilities that were previously scarce or expensive.
I described this as intelligence liquidity. The concept becomes more powerful when extended one step further because cheaper intelligence can increase the utilization of resources that already exist. It can release labor, capital, physical assets, and human time from low-productivity activities and redeploy them toward higher-value uses.
I think of this second-order effect as productive liquidity.
The distinction is particularly important when assessing the economic capacity available to support AI investment. The future revenue pool depends partly on the amount consumers and businesses spend directly on AI. It also depends on the productive resources AI releases elsewhere in the economy and on the new forms of economic activity those resources subsequently enable.
3. The Economy Is Full of Trapped Capacity
Consider a simplified economy producing $100 of output. Suppose $10 of that output is devoted annually to solving a particular problem. A technological improvement reduces the resources required to solve the same problem to $2. The remaining $8 represents labor, capital, and time that can now be redeployed.
Real economies contain enormous quantities of this underutilized capacity. A truck returning empty represents physical capital earning nothing on the return journey. A factory operating well below capacity reflects installed capital with unused productive potential. A doctor spending hours on paperwork represents highly valuable human capital allocated to administrative work. Excess inventory sitting unnecessarily in a warehouse ties up financial and physical resources while producing little economic value.
Technology frequently creates value by increasing the productive intensity of resources already inside the system. AI has the potential to apply that mechanism simultaneously across manufacturing, transportation, healthcare, finance, software, logistics, scientific research, professional services, and a wide range of administrative functions.
The economic impact therefore extends well beyond the revenue generated directly by AI providers. Lowering the amount of labor, time, inventory, working capital, or physical capacity required to accomplish an existing task effectively creates additional productive capacity elsewhere in the economy.
This is where static comparisons between AI spending and future GDP can become difficult to interpret. The same technology absorbing large amounts of investment may also be changing the efficiency with which the rest of the capital stock is used.
4. Imagine AI Helps Cure a Disease
Consider an extreme example. Assume society spends $500 billion annually treating and managing a particular disease. Now imagine AI accelerates scientific discovery sufficiently to produce a cure or dramatically better treatment, reducing the resources required to manage that disease to $100 billion per year.
The direct revenue earned by the AI companies involved would capture only a portion of the economic effect. Roughly $400 billion of resources would become available for other purposes. Doctors could treat other patients, researchers could pursue other diseases, hospital capacity could be redirected, insurers and governments could allocate capital elsewhere, and families could regain both income and time. Improved health could also increase labor-force participation and productive output.
National accounting creates an important complication. A large reduction in healthcare expenditures could initially reduce measured nominal GDP because GDP records spending rather than the full welfare value created by eliminating the need for that spending. Some of the released resources would eventually reappear as consumption, investment, profits, or other measured activity, while some of the benefit could appear as savings, additional leisure, or improved health.
A diagnostic procedure that costs $1,000 today and $50 tomorrow provides a smaller contribution to nominal spending despite representing an enormous productivity gain. If aggregate nominal GDP declined as a result of similar efficiencies, a fixed $3.7 trillion AI-revenue requirement would actually appear larger relative to GDP.
This accounting effect reinforces the need to separate several concepts. Nominal GDP, economic welfare, productive capacity, and revenue available to AI infrastructure are related, but they measure different things. AI could create extraordinary economic value while producing a much smaller amount of directly monetized AI revenue.
For investors, that distinction is fundamental.
5. Five Channels Through Which AI Expands Productive Capacity
AI can affect productive capacity through at least five major mechanisms.
Substitution occurs when companies reduce spending on certain forms of labor, outsourcing, software, administration, or other activities and redirect a portion of those budgets toward AI. This represents the most familiar form of AI monetization because an existing expense line becomes available to fund a new technology.
Optimization increases the output generated by existing assets. Factories can operate at higher utilization, inventories can decline, trucks can spend fewer miles empty, power can be allocated more efficiently, and working capital can spend less time idle. The economic benefit appears through higher asset productivity rather than simply through increased technology spending.
Liberation occurs when AI dramatically reduces the resources required to solve existing problems. Labor, time, physical capacity, and capital previously committed to those activities become available for alternative uses. The disease example represents an extreme version of this mechanism.
Creation encompasses products, services, discoveries, and business models that previously were technically impossible or economically uneconomic. Historically, this has been one of the most important consequences of general-purpose technologies because entirely new demand categories emerge that could not have been incorporated into an ex ante market-size calculation.
The fifth channel has the most direct implications for AI infrastructure: machine demand. Software has historically served human users whose consumption is bounded by biological limitations. Humans sleep, work a finite number of hours, and possess limited attention. Autonomous agents operate under a very different constraint set.
Agents can search, reason, code, communicate, monitor, negotiate, and transact continuously. The relevant user base for compute could therefore expand from human beings operating software to machines consuming intelligence in order to perform economic activity themselves. That shift has the potential to alter the demand curve for compute in ways that are difficult to infer from current human usage patterns.
There is another dimension to machine demand that may be even more important for productive capacity: time. The average human workweek is roughly 34 hours, or about 1,770 hours per year. A machine capable of operating continuously has 8,760 hours available each year, nearly five times as many. For economic activities that can be transferred from human labor to autonomous agents, the available time input into production can therefore expand dramatically even before assuming any improvement in productivity per hour.
This does not imply that GDP mechanically increases fivefold. Agents cannot perform every economic activity, demand is finite, and production remains constrained by capital, energy, physical infrastructure, and other scarce inputs. But it changes the potential production function. An agent performing economically useful work continuously can generate almost five times the annual working hours of a human operating on a conventional schedule. If the productivity of those machine hours also rises as models improve, the effect becomes multiplicative: more productive hours operating at higher productivity per hour.
This is one reason I think machine demand deserves to be treated as something larger than another category of software consumption. Agents and increasingly robots and humanoids create a new class of economic actor whose productive clock runs continuously. The transition from human time to machine time could materially expand the amount of economically useful activity the existing capital stock can support.
The revenue outcome will ultimately depend on the interaction between quantity and price. In its simplest form:
Compute revenue = price per unit × units consumed.
AI is simultaneously increasing the potential quantity of intelligence consumed and reducing the cost of producing each unit of intelligence. This creates a classic Jevons-style problem. Falling inference prices can stimulate dramatically higher consumption, while improvements in efficiency can reduce the revenue generated by each individual task.
The infrastructure bull case therefore requires growth in machine activity to exceed the decline in the unit price of performing that activity. Utilization could increase substantially at the same time that revenue per task falls. The central economic question becomes whether the elasticity of demand for intelligence is sufficiently high that consumption expands faster than prices decline.
That question will become increasingly important as agents move from experimentation toward persistent economic activity.
6. Time Is Both an Opportunity and a Risk
Rapid technological progress expands the potential size of the AI economy while creating a separate challenge for the infrastructure being financed today. The faster the technology improves, the greater the possibility that current hardware becomes economically obsolete before investors fully recover its cost.
New generations of GPUs can reduce the competitiveness of older hardware. Model improvements can lower the compute required to perform a given task. Algorithmic advances can increase the amount of useful work generated from the same physical infrastructure. Declining inference costs can also compress the prices that infrastructure owners and operators are able to charge.
The economic life of an AI data center therefore varies significantly by component. Land, access to power, and interconnection rights may retain strategic value for decades. Buildings and cooling systems can also have relatively long useful lives. Networking equipment generally operates on shorter replacement cycles, while GPUs can experience rapid functional depreciation as newer architectures become significantly more efficient.
This creates a duration mismatch inside the financing structure. Long-lived debt, leases, or project-finance obligations may depend on cash flows generated by assets whose technological competitiveness declines considerably faster than the physical infrastructure surrounding them.
Time consequently affects both sides of the AI investment equation. Technological progress can increase utilization by expanding the number and sophistication of economically viable AI tasks. The same progress can reduce unit prices, improve efficiency, and shorten the economic life of existing hardware.
For capital providers, the distribution of outcomes is therefore unusually wide. The economics of the buildout depend on the interaction among utilization, pricing, technological depreciation, financing structure, and the speed at which new applications generate demand.
7. Economic Revolution and Investment Returns Are Separate Questions
Transformational technologies routinely create enormous economic value while distributing financial returns unevenly across the capital deployed to build them. Railroads reshaped commerce and geography while producing repeated waves of bankruptcy. Electrification generated extraordinary productivity gains across the economy while delivering very different outcomes across the companies that financed and supplied the system. Telecommunications and fiber became essential foundations of the internet even as large amounts of invested capital were impaired.
AI could follow a similar pattern.
Suppose AI allows a corporation to save $1 billion annually and that corporation pays $100 million to the AI providers responsible for those savings. The corporation retains $900 million of economic surplus. That outcome would represent substantial economic value creation and potentially substantial value for the company adopting the technology. Only a fraction of the total benefit would become revenue available to service the capital invested in AI infrastructure.
This distinction places the Brookings financing question at the center of the investment debate. GPU manufacturers, memory suppliers, power producers, data-center owners, cloud platforms, model developers, application companies, and end users all participate in the same technological ecosystem. The distribution of economic surplus across those layers will determine returns.
The transmission mechanism can be thought of as a waterfall:
Economic value created or released
↓
Economic activity and spending
↓
AI revenue
↓
Infrastructure revenue
↓
Cash flow available to capital
Each stage narrows the amount of value available to the next. A technology can create an enormous amount of economic value at the top of the waterfall while delivering a far smaller pool of cash flow to a particular layer of infrastructure capital at the bottom.
This becomes increasingly important as AI infrastructure financing migrates toward leases, joint ventures, project debt, private credit, securitizations, and special-purpose vehicles. These structures increasingly separate the physical asset, the technology occupying that asset, the corporate tenant generating the revenue, and the investor ultimately assuming the risk.
For investors, the relevant question is therefore highly specific: what proportion of the economic value generated by AI ultimately reaches the particular layer of the capital structure I own?
8. Two Questions and Four Variables
This framework brings us back to the $3.7 trillion revenue requirement. The hurdle is economically meaningful because capital ultimately requires cash flow. Its relationship to a 2032 GDP forecast provides a useful sense of scale, although the economy generating that GDP will increasingly reflect AI-driven changes in productivity, resource allocation, prices, consumption patterns, and machine activity.
Investors therefore need to separate two questions.
The economic question asks how much productive capacity AI can create, improve, or release. This determines the potential size of the economic surplus generated by the technology and the scale of the opportunity available across the system.
The investment question asks how much of that economic value can be monetized by the specific asset or security an investor owns. This determines the return on invested capital.
I believe four variables will dominate that second question:
Utilization. Price. Value capture. Asset life.
Utilization determines how intensively the installed infrastructure is used. Price determines the revenue generated by each unit of compute or intelligence delivered. Value capture determines the share of the economic surplus retained by each layer of the ecosystem. Asset life determines how long the infrastructure remains sufficiently productive and competitive to recover its original cost.
These four variables provide a more useful framework for analyzing the AI capital cycle than the size of infrastructure spending alone. An asset with very high utilization can still produce disappointing returns if unit prices collapse. Strong pricing can provide limited protection when technological obsolescence shortens asset life dramatically. Enormous economic value creation can coexist with poor infrastructure returns when the majority of the surplus accrues to applications, enterprises, consumers, or other layers of the stack.
The investment outcome depends on the interaction among all four.
There is also a broader economic implication. Much of technological history can be understood as a sequence of innovations that released scarce resources from activities that previously consumed them. The steam engine increased the leverage of human and animal labor. Electricity transformed the productive intensity of physical capital. Computers reduced the cost of processing and transmitting information.
AI reduces the cost of intelligence itself.
That represents the next stage of what I described as QE for the Mind. Quantitative easing created financial liquidity by increasing the availability of capital. AI creates intelligence liquidity by increasing the availability of reasoning, analysis, creation, and decision-making. As that intelligence becomes embedded throughout the economy, it can generate a further effect: productive liquidity.
Productive liquidity represents labor, time, capital, and physical capacity released from lower-value activities and made available for higher-value uses. Its scale will influence the productive capacity of the economy that ultimately emerges from the AI transition.
Van Nieuwerburgh’s financing question is therefore essential: How much revenue must AI generate to earn an adequate return on the capital being invested today?
Investors should place a second question beside it: How much productive capacity will AI release, and where will the resulting economic surplus ultimately accrue?
The first question determines whether today’s capital structures can support themselves. The second helps determine the size and composition of tomorrow’s economy. Understanding both will be essential for allocating capital through the AI investment cycle.
