The Macroeconomics of AI: Why the Economists Can't Agree—And What That Means for Your Portfolio

There is roughly $25 trillion of disagreement sitting in the academic literature on AI. On one end, Daron Acemoglu's "The Simple Macroeconomics of AI" (2024) forecasts a 10-year boost to US total factor productivity of no more than 0.66%. On the other end, Anton Korinek and Donghyun Suh sketch scenarios with global GDP rising 100% to 300% over the same horizon. Goldman Sachs splits the difference and predicts a 7% GDP boost, worth around $7 trillion. McKinsey models suggest $17–26 trillion in added global output.

These are not minor calibration differences. They are different theories of the technology. And if you're allocating capital to AI—whether through NVIDIA, hyperscalers, application-layer plays, or by avoiding exposed sectors—which theory is right determines whether the current AI capex cycle is the early innings of a productivity revolution or the largest misallocation of capital since the telecom build-out of 1999.

I've spent the last few months working through the major papers in this debate. What follows is my read on each, my synthesis, and the investment framework I'm using to position around it. Standard disclaimer: this is analysis, not financial advice. You're responsible for your own positions.

Part 1: The Acemoglu Anchor

Acemoglu's paper is the most rigorous skeptical voice in the debate, and any serious analysis has to start there because his framework imposes the tightest discipline. The core idea is Hulten's theorem, which says that in a competitive economy with constant returns to scale, the macro productivity effect of a microeconomic improvement equals the share of output it affects multiplied by the cost savings.

TFP gain = (share of GDP impacted) × (average cost savings)

This is not a controversial equation. It's accounting. The disagreement is over the two numbers you plug in.

Acemoglu's numbers come from combining Eloundou et al.'s exposure data with Svanberg et al.'s feasibility estimates and the early productivity experiments. He arrives at:

• GDP share impacted in 10 years: roughly 4.6% (20% of tasks exposed × 23% feasibility rate)
• Average cost savings: 14.4% (27% labor cost reduction × 53% labor share)
• Resulting TFP gain: 0.66% over 10 years, or roughly 0.064% annually

He then introduces a distinction between "easy-to-learn" tasks (clear outcome metrics, simple action-outcome mapping) and "hard-to-learn" tasks (context-dependent, no clean outcome data) and argues 73% of exposed tasks fall in the easy bucket. Adjusting for the harder cases brings the TFP estimate down to 0.53%.

GDP grows somewhat more than TFP because of induced capital investment—Acemoglu's high-end estimate is 1.56% over 10 years if there's a meaningful capex response. He also makes an underrated point: if the capital-output ratio rises faster than TFP, GDP overstates welfare because the extra investment comes out of consumption.

Where Acemoglu's framework is strong

The discipline he imposes on forecasts is exactly right. Anyone projecting 7%+ GDP boosts has to either show much broader task exposure or much larger cost savings than the direct evidence supports. The easy/hard distinction captures something real—LLMs underperform in domains without clean outcome metrics. The bad-tasks point ages better every quarter as AI-generated content fills the internet.

Where it's vulnerable

Three places. First, the cost-decline assumption. His feasibility rate rests on conservative GPU price-performance trends. Inference costs have fallen much faster—by roughly an order of magnitude on some workloads between 2023 and 2025. More tasks become economically feasible to automate as compute gets cheaper, and his 4.6% share is probably understated.

Second, the easy/hard line. Reasoning-capable models attack tasks that Acemoglu would have put in the hard bucket. Multi-step reasoning, mathematical proof, complex code synthesis—these have moved faster than the paper's framing implied.

Third, investment response. Acemoglu assumed AI capex would be modest. But the providers' spending has been enormous—Microsoft, Google, Amazon, Meta, and AI labs committed hundreds of billions in 2024–2025 alone.

Part 2: The Competing Visions

Aghion, Jones & Jones (2018)

If Acemoglu is the empirical skeptic, Aghion, Jones, and Jones provide the more theoretically sophisticated framework. Their answer hinges on two effects. First, Baumol's cost disease in reverse: as AI automates more tasks, the non-automated sectors become a drag on aggregate growth. Second, the O-ring effect: if production requires many tasks done in sequence, automating 80% of tasks while 20% remain bottlenecks doesn't get you anywhere near 80% cost savings.

Where they depart from Acemoglu is on the long run: they argue that if AI can eventually substitute for human labor in producing ideas themselves (R&D, science), growth can accelerate substantially. The unanswered question is whether current frontier AI is approaching the "automating ideation" threshold.

Brynjolfsson, Rock & Syverson: The Productivity J-Curve

Their argument: general-purpose technologies don't show up in measured productivity for a long time after they're invented, because firms have to build complementary intangible capital—new processes, organizational structures, employee skills—before the productivity gains arrive. The pattern looks like a J. They estimate this lag was about 20 years for the personal computer revolution. Electricity took roughly 40 years.

This framework can rescue both Acemoglu and the optimists. Acemoglu may be right that 10-year TFP gains are small—we're in the flat part of the J. The optimists may be right that gains arrive eventually. If Brynjolfsson is right, the macro payoff might be a 2035–2045 story, not a 2025–2035 story.

Autor (2024): Rebuilding Middle-Class Jobs

Autor's argument: AI can democratize expertise. Tools like LLMs let less-skilled workers reach near-expert performance in domains previously gated by training and credentials. This could reverse decades of skill-biased technical change. The Brynjolfsson-Li-Raymond customer service study is exhibit A—lower-skilled workers gained the most. Where this is weak: Acemoglu's counter that productivity gains for low-skill workers can increase inequality if they trigger ripple effects across labor markets is hard to dismiss. Early 2024–2025 data on hiring for junior knowledge work suggests AI may be substituting for entry-level workers rather than complementing them.

Korinek & Suh (2024): AGI Scenarios

This is the polar opposite of Acemoglu. Korinek and Suh model scenarios for the transition to artificial general intelligence and produce baseline projections of 100% GDP growth over 10 years, with aggressive scenarios reaching 300%. The math is straightforward once you accept the premise: if AI can substitute for all human labor, the constraints that bind growth dissolve. Where this is weak: the premise is doing all the work. As a probability-weighted forecast, the expected value depends entirely on your AGI probability. Even if the AGI scenarios are unlikely, the option value affects asset prices today.

Goldman Sachs and McKinsey

Goldman's 7% global GDP boost and McKinsey's $17–26T range rest on aggressive assumptions about adoption speed and cost-savings realization. Neither incorporates Acemoglu's hard-task distinction or Brynjolfsson's J-curve lag. I treat these as upper-bound scenarios, not central forecasts.

Part 3: Where the Data Has Landed (2024–2026)

Part 4: My Synthesis

Acemoglu is mostly right on the headline numbers. The 10-year TFP gain will probably land in the 1–2% range, not 7%. The macro data hasn't given the optimists any meaningful support, and the structural arguments all push toward modesty.

Acemoglu is wrong on the upside tail. Cost declines are faster, reasoning models attack his hard-task category, and the AGI option value is non-trivial. The probability of a 5%+ TFP outcome by 2035 isn't zero—it's maybe 10–15%—and asset prices should reflect that tail.

The optimists are right that the long run is much bigger. Brynjolfsson's J-curve framework probably describes what's happening: the productivity payoff isn't a 2025–2030 story, it's a 2030–2045 story.

The most underweighted concern is Acemoglu's bad-tasks point. As AI-generated content, deepfakes, and engagement-optimized manipulation expand, measured GDP can rise while welfare falls. This is also a setup for regulatory backlash.

Part 5: The Investment Framework

Tier 1: Infrastructure (Picks and Shovels) — Highest Conviction

The capex cycle is happening regardless of whether end-user productivity gains pan out. Hyperscalers are committed to multi-year buildouts. Even if Acemoglu is right and the productivity gains are modest, the infrastructure spending has years of runway. Companies that benefit: NVIDIA (compute), TSMC (fabrication), ASML (lithography), Micron and SK Hynix (HBM memory), utilities and cooling infrastructure exposed to data center load growth.

Caveat: infrastructure plays are partially priced in. The asymmetric setup is better in less-loved parts of the chain—power, memory, fab equipment—where the AI link is less direct and valuations less stretched.

Tier 2: Hyperscalers — Moderate Conviction

Microsoft, Google, Amazon, and Meta are spending the capex. They have the cash flow to absorb it and other businesses to offset disappointment. The risk: their AI capex returns depend on whether end-customers can monetize AI productivity gains. Watch capex/revenue ratios and free cash flow trends. These are core holdings rather than asymmetric bets.

Tier 3: Applications — High Uncertainty

Most pure-play AI application companies will fail or be acqui-hired. Where this might work: narrow vertical applications with proprietary data, distribution, or regulated moats. Coding tools, customer service automation, drug discovery platforms, legal document automation in regulated jurisdictions. I'd treat this tier as venture-style bets within a public portfolio—smaller positions, more diversified, expect most to underperform.

Tier 4: What to Underweight

Acemoglu's exposure data points to sectors where AI substitutes for labor: pure-play BPO and basic call center providers, content production and basic marketing services, routine knowledge work concentrated firms, and pure-play search and ad businesses heavily exposed to generative substitution.

Tier 5: AI-Resilient — Worth a Sleeve

Local and physical services, regulated industries with high switching costs, branded consumer where physical presence matters, and real assets generally. These won't outperform AI infrastructure, but they're the hedge if the AI capex cycle disappoints.

Part 6: Signals to Watch

• →
  TFP data: US nonfarm productivity statistics. If they start running 2%+ annually for multiple years, the optimists are winning. If they keep running ~1.5%, Acemoglu is winning.
• →
  Hyperscaler capex/revenue ratios: Healthy at current levels, but if capex keeps growing faster than revenue for another 2-3 years with no operating margin improvement, that's a bubble warning.
• →
  Power constraints: Data center electricity demand growth is colliding with grid capacity. It's both an investment opportunity and a constraint on the bull case.
• →
  Adoption surveys: Census BTOS and equivalents. Acceleration past ~20% would suggest broader productivity gains are coming.
• →
  Entry-level knowledge work hiring: The clearest near-term test of whether AI substitutes for or complements labor.
• →
  Regulatory news: Bad-tasks dynamics will produce regulatory backlash. Watch for EU AI Act enforcement, US state-level legislation, China policy shifts.
• →
  Reasoning model deployment: Are reasoning-capable models showing up in measured productivity yet, or are they still demos? This is the cleanest test of whether the hard-task barrier is falling.

Part 7: What Would Change My Mind

Conclusion

Acemoglu is closer to right than the consensus is comfortable admitting. The macro data through 2026 isn't supporting the bullish case, and the structural reasons it might not—hard tasks, productivity J-curves, bottleneck effects—are getting more empirical support, not less.

But "closer to right" isn't the same as right. The cost-decline trajectory, reasoning model progress, and AGI option value all argue against treating his numbers as a ceiling. The realistic range for 10-year TFP gains is probably 1–3% with meaningful upside tail, not 0.5–0.7%.

For investors, this means the infrastructure trade is still working but is increasingly priced; the application-layer story is hard; and the macroeconomic payoff is more likely a 2030s story than a 2020s story.

The investors who do best from here are the ones with the right time horizon. If you need productivity gains to show up in macro data by 2028 to validate your positions, you're probably in the wrong trade. If you can hold infrastructure exposure through a capex cycle that may run another 5–10 years, the asymmetric setup is still attractive.

What I keep coming back to is Acemoglu's quiet observation: even if you accept all the optimistic productivity claims, the welfare gains can be smaller than the GDP gains if capital-output ratios rise faster than TFP, or if some of the new tasks are socially negative. The hardest question in AI investing isn't whether the technology will get better. It's whether the gains will accrue to the right places—and on what timeline. Position accordingly.

This piece reflects my own analysis of publicly available academic literature and market data. It is not investment advice. AI development and its economic implications are areas of active disagreement among serious researchers and forecasts here will be wrong in ways I can't anticipate. Do your own work.