A Yale economist says AGI won’t automate most jobs—because they’re not worth the trouble

Restrepo's NBER working paper argues labor's share of GDP converges to near zero as AGI directs compute (potentially ~10^54 flops) to automate economic 'bottleneck' tasks while leaving many 'supplementary' jobs untouched; the computing power of all human brains is roughly ~10^18 flops. The paper warns income will concentrate with computing-owners unless policy intervenes (redistribution/universal basic income or treating compute as public capital), citing Larry Fink's concern about wealth concentration. Near-term dislocations are likely in algorithm-driven jumps (sharp wage divergence), illustrated by data-center trade premiums: avg. $81,800 on data-center construction (~32% higher), some electricians earning ~$260,000, and a projected need for ~300,000 new electricians plus ~200,000 replacements over the next decade.

Analysis

Concentration of programmable compute will create a capital-intensity premium that accrues to hardware, hyperscalers, and the real assets that host them. Expect above-market returns for firms that control accelerator supply chains and data-center footprints, but those returns will be episodic and hinge on a handful of bottleneck suppliers (advanced lithography, specialty substrates, power transformers) rather than broad-based technology adoption.

A non-obvious second-order effect is stress on energy and construction supply chains: accelerated data-center builds will bid up electricians, copper, transformers, and local grid capacity, creating a multi-year boom for industrial suppliers and niche contractors while compressing margins in adjacent commercial construction. That creates clustering risk—regions that host hyperscale campuses will see wage inflation, permitting delays and political backlash that can materially delay project timelines and returns.

Regulatory and geopolitical catalysts are the primary near-to-medium term drivers: export controls on lithography or GPU-grade chips, or a policy push to treat compute as a public good, can rerate winners overnight. Conversely, technological decentralization (specialized edge ASICs or breakthroughs reducing GPU dependency) would democratize compute and reprice the leaders downward over 2–5 years.

The consensus framing misses that owning raw compute does not automatically translate to value capture if software, datasets, and distribution layers remain fragmented; investors should stress-test ownership narratives against four variables—manufacturing concentration, energy bottlenecks, regulatory intervention, and software lock-in—to avoid mispriced permanence in what may be a cyclical, politically contested transformation.

Key Decisions for Investors

• Long NVDA (or a 12-month NVDA call spread) — T+6–18 months. Rationale: direct exposure to accelerator demand; risk: valuation multiple compression or sudden supply-side easing. Target asymmetric payoff by buying calls or call spreads to cap downside; 3:1 upside potential if next-gen GPU cycles remain supply-constrained.
• Long data-center REITs (DLR, EQIX) — T+12–36 months. Rationale: secular demand for colocated capacity and rising lease rates; risk: regulatory restrictions on builds or rapid move to on-prem/edge. Position size: modest (5–8% of tech allocation) with stop at 20% drawdown; expected total return 10–18% annual if buildouts proceed.
• Long semiconductor equipment duopoly exposure (ASML, LRCX) — T+12–36 months. Rationale: durable pricing power via capital equipment; risk: cyclical capex pullbacks and export controls. Use outright equity or 18–36 month LEAP calls to capture multi-year scarcity; downside preserved by durable technology moat.
• Pair trade: Long BLK (T+12 months) / Short a consumer discretionary ETF concentrated in labor-intensive services (e.g., XLY) — T+12 months. Rationale: asset managers capture concentrated wealth gains and distribution fees; consumer services face stagnating wage-linked consumption. Risk: policy redistributions (UBI, capital taxes) compress asset manager multiples; target a 1.5:1 reward-to-risk by sizing the pair net-neutral on beta.