An OpenAI cofounder ‘vibe coded’ an analysis of the U.S. labor market’s exposure to AI, and the highest-paying jobs have the worst scores

Key datapoint: Andrej Karpathy’s BLS-based analysis assigned an overall weighted AI/automation exposure score of 4.9 (scale 0–10), with occupations earning >$100k averaging 6.7 versus <$35k averaging 3.4. High-skilled roles (software developers, data scientists, financial analysts, writers, designers) scored ~9, while manual and service roles (construction laborers, janitors, home health aides, bartenders) scored 1–2, implying disproportionate exposure for white‑collar, higher‑paid jobs. The chart was subsequently removed by Karpathy and is described as informal, so interpret results as directional risk signals for workforce composition and hiring rather than definitive forecasts.

Analysis

The uneven exposure of occupations to AI will create concentrated winners at the infrastructure and enterprise orchestration layers and concentrated losers in fee-for-time businesses and low-margin staffing. Expect capital spending on GPUs, interconnect, and model hosting to scale faster than software seat-license growth: a 12–24 month procurement cycle for large enterprises implies front-loaded capex for cloud providers and GPU vendors, and a more gradual reallocation of labor budgets into software and monitoring functions.

Second-order effects matter: large reductions in junior analyst/headcount will compress hiring pipelines, raising replacement costs for mid-senior technical roles over a 2–5 year horizon and benefitting firms that operate reskilling platforms or provide “human-in-the-loop” validation services. Conversely, lower demand for entry-level work reduces churn and temp staffing revenue, creating a structural headwind for payroll/staffing firms even if overall economic activity is steady.

Tail risks and catalysts are asymmetric. Fast adoption (6–18 months) driven by verticalized, high-ROI automations can re-rate infrastructure names materially, while a regulatory clampdown, high-error event, or a failure to reduce total cost of ownership vs humans could stall adoption and compress multiples across the stack. The consensus underestimates the persistence of domain expertise premiums — firms bundling models with specialist oversight can keep pricing power, so blunt short positions on professional services are risky without a paired long in pure-play automation software or infrastructure.

Key Decisions for Investors

• Long NVDA via 12–18 month call spread (buy NVDA 12mo ITM call / sell higher-strike call) — directional exposure to GPU-driven model demand with capped cost; reward asymmetric if enterprise AI capex accelerates, risk is multiple compression if adoption stalls.
• Overweight MSFT or GOOGL (6–12 month horizon) — play capture of model-hosting and enterprise AI stack; hedge with short exposure to a staffing name (MAN) to isolate AI infrastructure upside vs labor-replacement downside.
• Pair trade: Long ServiceNow (NOW) or Alteryx (AYX) 9–15 month calls, Short Manpower (MAN) stock — automation SaaS should show higher revenue per seat and margin expansion while staffing faces secular headcount declines; expect 2:1 upside/downside if enterprise procurement favors software over temps.
• Tactical hedge: Small long volatility position on TSLA (2–3 month straddle) sized as ~1–2% portfolio risk — protects against sentiment swings if AI narratives spill into auto/autonomy headlines or macro-driven de-risking causes correlated tech drawdowns.