Brief #13: AI firms are spending as if a boom is coming
Taking investment data seriously suggests a productivity boom, while data-driven automation could lead to stagnating wages
Welcome! This bi-weekly newsletter, published by the Windfall Trust, curates the most important developments in AI economics research and policy.
Need to Know
Treating AI firms’ investment as rational implies a ~170% productivity boom in the AI sector. It also increases GDP by 2030 by 5 to 58 percentage points in the standard investment model. The 5pp moderate scenario is the most decision-relevant benchmark; the higher estimates depend on speculative additional productivity booms. The moderate scenario is a mechanical implication of the extraordinary investment plans of AI firms, but requires strong assumptions on, e.g., market power, boom probabilities, and adjustment costs.
Wages might stagnate if data keeps improving AI systems. This could also lead to runaway growth and suggest that working on bottleneck tasks may doom workers, since they face the highest risk of automation.
In the news
The UK has set up the first AI Economics Institute (AIEI), chaired by Nobel Prize laureate Simon Johnson. It has two functions: building the evidence base, and developing models and scenarios for economy-wide impacts. The data-access plans deserve attention: the existing Future of Work Unit and its LinkedIn data partnership both transfer in; more than thirty firms (among them BT, Rolls-Royce, and Accenture) have agreed to share workplace AI data, and a working group with Anthropic, OpenAI, and other major labs is due by the end of June. The Institute will hire civil servants alongside external economists and analysts and commission external research rather than do everything in-house.
Economists split three ways on AI and jobs. A Wall Street Journal survey of prominent economists produced a near-even division on whether AI is already reducing employment. Jason Furman judged the aggregate effect “small to zero” and called the evidence of reduced hiring weak; David Autor expects effects within five to ten years and warned that institutions are unprepared; Michael Strain pointed to the Industrial Revolution, when real wages stagnated for decades. Meanwhile, employers attributed about 40% of May layoffs to AI, against 4.5% across all of 2025, even in a month when payrolls grew.
The Wall Street Journal also covered a recent workshop by Windfall Trust (the publisher of this brief) in Washington, DC, where over 40 leading thinkers in economics, technology, and public policy convened at the Peterson Institute for International Economics to discuss the following scenario: It’s 2030. The growth rate has nearly doubled. Nonetheless, underemployment has also jumped from 8% to 14%, also affecting college-educated workers. Policymakers predicted that unrest would rise, fertility would fall, and reskilling wouldn’t be enough. The social contract is visibly fraying, as UBI, taxing AI firms, and sovereign wealth funds are openly discussed. This is also true for South Korea, where Prime Minister Kim proposes linking UBI to AI wealth taxation ahead of Windfall Trust’s workshop with the Bank of Korea on July 8.
In detail
What Investment Data implies about the AI Transition
Jessica Wachter (Wharton) and Jonathan Wachter calculate the implications of large capital expenditures (capex) in the AI sector for growth. Assuming that technology firms invest in AI data centers until another data center yields zero profit, the authors expect a 170% productivity boom in the AI sector. When they attribute this to three scenarios consistent with industry forecasts, the productive AI sector expands to 8%-39% of the economy. This would imply that GDP by 2030 is between 5 and 58 percentage points higher.
The paper takes the technology firms’ revealed preferences seriously. These are a good measure for AI firms’ beliefs, but require a range of assumptions, e.g., market power, boom probabilities, and adjustment costs to be useful for productivity estimates. In contrast, others have estimated AI’s productivity contribution bottom-up from tasks (Acemoglu’s ~0.7pp over a decade) or asserted it top-down (Amodei’s trillions of dollars in capex leading to decades of scientific progress compressed into a year). Meanwhile, the authors focus on interpreting large investment numbers. The five largest hyperscalers spent roughly $380 billion on capex in 2025. The paper estimates this rises to $755 billion in 2026 and $1.09 trillion in 2027 — about 19% of U.S. private fixed investment.
They then embed an AI sector into a frictionless two-sector open-economy model. Booms raise productivity while leaving capital unchanged. This leads to jumps in the marginal product of capital, prompting AI firms to invest heavily in data centers. Since the open economy ensures that the required returns are independent of investment, the investment ratios depend only on the size of the boom.
Overall growth in GDP and AI sector’s share of GDP in three scenarios, ranging from 0 to 2 productivity booms between 2029 and 2030
This model implies that the average growth rate rises from 2% to 7%. Assuming that consumers are risk-averse and that consumers’ willingness to defer consumption when rates rise is low, the risk-free rate increases by 0.5 pp and the equity premium by 3 pp.
What this means: This isn’t a conventional forecast, but rather an extrapolation from hyperscalers’ revealed beliefs, with two additional productivity booms added on top. The central assumption is that AI firms have accurate expectations about their investment prospects. This would yield a conservative lower bound on productivity under many assumptions, especially if they are compute-constrained. But what firms must believe for spending to be rational is not necessarily the truth. If firms overinvest - whether due to overbuilding or strategic preemption in a winner-takes-all race- then the implied 170% productivity boost is overestimated.
The most decision-relevant case is the moderate scenario without an additional productivity boom in 2029 or 2030. Even there, the AI sector rises from roughly 3% to 8% of GDP, adding about 5 percentage points to GDP by 2030 and already exceeding task-based estimates. The transformative and singularity scenarios should be read as tail scenarios, not central forecasts: they depend on two years with 50% annual probabilities of further productivity booms.
This model features heavy precautionary saving against boom uncertainty rather than escaping the permanent underclass, as well as a moderate increase in interest rates. In contrast, Chow et al. (2026) suggested that AI has a large effect on interest rates but an ambiguous effect on the equity premium, e.g., due to misaligned AI that destroys equity value. Still, this paper only covers the upside of AI, not its downside.
How would AI automation based on high-quality data reverberate across the economy?
Another way to model the large-scale, long-term impacts of AI is to apply general equilibrium theory rather than interpreting AI investment figures. Maryam Farboodi, Andrew Koh, and Anchi Xia (MIT) model automation. Data is task-specific but accumulates as a byproduct of producing the task. The benefits of data also spill over into adjacent tasks.
In the long run, automation can be contagious with these spillovers. Without spillovers, if tasks complement each other more than the returns to data diminish, the economy produces more data on labor-intensive, data-poor tasks.
While this protects workers’ wages during bottlenecks in the short term, it increases incentives to automate these tasks, leading to long-run wage stagnation. Meanwhile, if tasks can be replaced quickly relative to the diminishing returns of data, capital concentrates in data-rich tasks. Consequently, automation stalls in the core, and wages grow for bottleneck tasks.
Automation spreads only slowly through the economy in the long run, but the exact dynamics are unclear. E.g., if there’s a core of easy-to-verify tasks such as coding and data analysis, then AI’s capabilities will be jagged. Other, messier tasks in the economy accumulate data more slowly, which limits automation in those areas.
In the long run, a singularity (infinite growth in finite time) is possible when the accumulation of data and compute reinforce each other. This gradually reduces the need for labor, leading to slower diffusion than the software-only singularity discussed in Brief #8. Learning by doing takes longer than automating AI R&D.
A race between spillovers and new tasks determines the future of work: if new tasks emerge quickly, humans might have a comparative advantage in them, thereby keeping wages and the labor share high. But these tasks need to be truly new: if AI systems can apply their knowledge from data on related tasks, these new tasks won’t require human labor.
This all rests on the assumption that workers train their own replacement by fulfilling a task. That reducing errors on some tasks corresponds to economic productivity is another big assumption. Additionally, data’s value shouldn’t perish, which is plausible when data feeds into automation.
What this means: This paper illustrates many conceptual points of current AI development, including fine-tuning on high-quality data and the jagged frontier. It also combines network theory with an equilibrium model. But achieving unbounded productivity growth requires simplifying Chinchilla’s scaling law by omitting the irreducible error term. Otherwise, there would be a limit to how much additional productivity data can deliver. More data might not be enough to achieve perfection.
While Jones and Tonetti’s weak-link paper suggested that task complementarity prevents a growth explosion, the authors argue the opposite holds when data accumulates as a byproduct of production: complementarity channels economic activity toward labor-intensive bottleneck tasks, which protects their wages in the short run but also generates the data that eventually automates them, leading to stagnant wages and potentially a singularity.
The existence of startups providing RL environments and high-quality data cuts both ways for the model. While this illustrates the importance of data, it also suggests that not every worker produces the data the model needs to work. Frictions, such as labor protections, limit the availability of free data even for verifiable tasks, suggesting slower change.
Compared to Wachter & Wachter, this model features slower AI expansion across the economy if AI systems focus on a core of verifiable tasks, such as coding, first. But the investment data reveal that labs bet on rapid diffusion, which is unlikely under a jagged frontier. Recursive self-improvement may explain this gap, as would a bubble.
In Other News
Labor Market & Employment
PwC’s 2026 Global AI Jobs Barometer shows that jobs that AI professionalizes are growing about twice as fast as those it democratizes, with 42% faster wage growth since 2021, and that entry-level roles in exposed fields are seven times more likely to require senior skills.
Daron Acemoglu (MIT) et al. model the relationship between automation and repression.
Auyon Siddiq and Niuniu Zhang (UCLA) use data from the freelancer platform Upwork to show that AI commoditizes labor, with wage demands becoming more important than the worker’s profile in hiring decisions.
Deric Cheng and Jacob Schaal (author of this brief) published a roadmap for the upcoming labor transition.
Carl Frey (Oxford) argues in the FT that AI may allow consumers to do more tasks themselves, reducing demand for paid work.
Lukas Althoff (Stanford) and Hugo Reichardt (CREI) model the effects of task-specific technical change and find that AI narrows wage inequality by lowering skill requirements.
Ramp and Revelio Labs find that AI adopting firms increase employment, including of juniors.
AI Capabilities & Infrastructure
The European AI policy think tanks Arq, KIRA, and the Oxford Martin School, along with others, published a scenario titled Europe 2031 on Europe’s economic and geopolitical future.
Policy Engine introduced Policy Bench to test how accurately language models calculate household taxes and benefits.
Philip Trammell (Stanford/Epoch) theorizes that recursive self-improvement may not lead to a growth explosion since “geniuses in a datacenter” aren’t parallelizable.
Andrew Caplin (NYU) argues that AI capex also has a learning-by-doing function.
Andreas Schaab (Berkeley) and Simon Scheidegger (Lausanne) introduce Equilibrium World Models, a deep-learning solver that aims to make dynamic stochastic economic models reliable beyond their simulated equilibrium paths.
Anthropic finds that the successful use of Claude Code depends more on domain expertise than coding expertise.
OpenAI reports that high adopters, such as its employees, use skills and coding agents much more and delegate work.
Entrepreneurship & Risk
Rembrand Koning (HBS) and Hyunjin Kim (INSEAD) show that AI-native Y-Combinator startups have a 25% smaller workforce, with fewer managers and entry-level workers, thereby flattening the hierarchy.
Research Opportunities
Anthropic publishes an Economic Policy Framework keyed to three unemployment scenarios and pledges $350 million for research and reskilling.
The Economist profiles the most AGI-pilled economists, including academics, nonprofit leaders, and the head of the UK’s new AI Economics Institute.
Pavel Kireyev (LSE) and Roberto Maura Rivero (Oxford, Meta) published a guide to help microeconomists contribute to AI research.
Thanks to Deric Cheng and Joel Christoph for contributing to this week’s edition of the newsletter.