‘Humans could go the way of horses’: Goldman calculated how bad the AI ‘job apocalypse’ will be—and its analysts were pleasantly surprised

Goldman Sachs analysts Joseph Briggs and Sarah Dong estimate AI could automate 25% of work hours, with a baseline 15% AI-driven labor productivity uplift implying 6–7% of jobs displaced over the adoption period and a peak gross unemployment increase of ~0.6 percentage points (about 1 million workers). They note historical precedent for technology creating new occupations—only 40% of current workers are in jobs that existed 85 years ago—and point to >6 million workers in computer-related roles and a further 8–9 million in gig/e-commerce/content roles, suggesting significant structural labor reallocation rather than permanent labor obsolescence.

Analysis

Market structure: AI adoption concentrates economic rents into compute, cloud, and model-owners (NVIDIA, MSFT, GOOGL, AMZN) while compressing margins for routine-labor-exposed firms (staffing, BPO, low-end retail). Goldman’s baseline (15% productivity uplift, 25% hours automatable, 6–7% job displacement) implies a multi-year surge in demand for GPUs, datacenter capacity and enterprise AI software, tightening supply of high-end chips and cloud slots and lifting pricing power for infra providers over 12–36 months.

Risk assessment: Key tail risks are regulatory restrictions (export controls/AI usage taxes), a political backlash leading to protectionist labor policy, and an energy-cost spike that raises marginal compute cost; any could cut projected productivity gains by >50% and re-rate multiples. Near-term (days–weeks) volatility will hinge on model launches and earnings; short-term (3–12 months) on enterprise pilots and layoffs; long-term (3–5 years) on reskilling pace and data-access economics.

Trade implications: Go overweight AI infrastructure (NVDA 2–3% portfolio exposure via 3–9 month calls or equity) and cloud software (MSFT/GOOGL 1–2% each) funded by tactical shorts in staffing (MAN) or under-invested legacy semis (INTC). Use pair trades (long NVDA / short MAN) to neutralize market beta; sell covered calls on large-cap cloud names to monetize high IV ahead of model-release windows. Target entries on pullbacks >8–12% or ahead of confirmed enterprise rollouts; trim into 30–50% rallies.

Contrarian angles: Consensus understates reskilling friction, data-ownership Duopoly and export controls—AI may be more oligopolistic and less broadly job-creating than tech booms of the past, so semiconductors’ multiples could be stretched. Historical parallels (mechanization) show eventual job growth but protracted social/political shocks; hedge for policy shock (1–2 year horizon) and be ready to rotate into cyclicals if regulators slow adoption.

Key Decisions for Investors

• Establish a 2–3% portfolio long in NVIDIA (NVDA) via 3–9 month at-the-money calls or direct equity; entry on pullback >8% from current levels or immediately with a 20% stop, target 30–50% profit take within 3–9 months.
• Allocate 1–2% each to Microsoft (MSFT) and Alphabet (GOOGL) equities for durable cloud/AI exposure; sell 1–2 month covered calls (25–35% OTM) to generate income ahead of major model/earnings catalysts, reassess after 2 quarters.
• Establish a 1% short/equivalent notional put position in ManpowerGroup (MAN) or Robert Half (RHI) using 3-month 15% OTM puts (or equity short) to capture labor-displacement risk; cover if unemployment uptick exceeds 0.6 percentage points or after 6 months if no deterioration.
• Execute a dollar-neutral pair trade: long NVDA (or MSFT) vs short MAN (equal notional) sized to 1–2% net exposure to hedge macro beta; rebalance after major policy announcements or if NVDA outperforms by >40% relative to MAN.