Goldman sees an AI bottleneck that can’t be vide-coded away. Ford’s CEO warns it’s already a ‘full-blown’ crisis

Goldman Sachs says AI infrastructure faces a major physical bottleneck, including a projected 45 GW power shortfall for U.S. data centers by 2028 and 760,000 additional power and grid workers needed by 2030. The article argues that agentic AI could be 60x to 130x more energy-intensive than today’s tools, shifting investor focus toward power generation, grid equipment, cooling, and other "pick and shovel" providers. Ford CEO Jim Farley echoed the labor shortage theme, saying the energy transition and data center buildout are already constrained by shortages of electricians, linemen, and other skilled trades.

Analysis

The market is still pricing AI as a hardware and software margin story, but the binding constraint is shifting to regulated, rate-baseable infrastructure with long lead times. That matters because the beneficiaries move from a winner-take-most software stack to a much broader set of capital-intensive assets whose returns are slower, but more durable and less exposed to model obsolescence. The first-order winners are not the obvious AI darlings; they are the firms that can convert scarcity into backlog, pricing power, and utility-like cash flows.

The underappreciated second-order effect is that labor scarcity turns every bottleneck into a schedule risk premium. If electricians, switchgear, and substations are the gating items, then hyperscaler capex can still be authorized without being monetized on time, which creates a wedge between headline spend and revenue realization. That favors companies with existing installed bases, service revenue, and balance-sheet capacity to pre-finance working capital; it hurts speculative developers and smaller contractors who need flawless execution and cheap labor to keep margins intact.

For Ford, the message is not just an EV/industrial anecdote; it is a sign that the same labor pool is being pulled simultaneously by automotive retooling, grid buildout, and data-center power demand. That increases the probability of wage inflation in skilled trades and makes margin compression a broader industrial risk over the next 12-24 months. The key contrarian point: the bottleneck theme is real, but the easiest way to own it may already be too crowded in pure-play grid names, while the more mispriced opportunity is in diversified industrials and services with exposure that is not yet reflected in multiples.

Near term, the data should show up first in order books, not earnings, because backlog can rise while revenue recognition lags by several quarters. If permitting, interconnect queues, or utility approvals accelerate, the trade can reverse quickly; if they do not, the shortage becomes a 2026-2028 story rather than a next-quarter catalyst. The market should watch for evidence of pricing power in high-voltage equipment, cooling, and field services before assuming the AI capex wave converts cleanly into cash flow.

Key Decisions for Investors

• Long GS vs. short a basket of AI compute beneficiaries over 3-6 months: own the advisor/allocator that can monetize infrastructure scarcity while reducing exposure to the crowded chip-led capex trade; target a 1.5-2.0x upside/downside if investors rotate toward bottleneck winners.
• Initiate a basket long in grid/electrification services names and keep it hedged with XLI or a semiconductor short for 6-12 months; the pair should work if backlog turns into margin expansion while the market de-risks pure AI capex multiples.
• Buy F on weakness for a 12-24 month horizon only as an indirect beneficiary of skilled-trade scarcity and energy conversion; risk/reward is attractive if investors re-rate industrial companies with adjacent energy optionality, but keep size modest because execution risk remains high.
• Prefer long-dated call spreads on utilities or power-infrastructure proxies rather than outright equity longs; this gives exposure to a multi-quarter bottleneck theme while limiting downside if rate pressure or permit delays slow project starts.
• Avoid or short speculative data-center developers with heavy reliance on near-term interconnects and contractor availability; if labor and grid constraints persist, these names face the highest probability of revenue slippage and multiple compression.