The Achilles Grid
The AI industry spent three years pretending it was purely digital. Software. Weights. Tokens. Math.
Then the Strait of Hormuz closed, oil hit $100 a barrel, and everyone remembered that inference runs on electricity, which comes from gas, which comes through pipelines and shipping lanes that can be disrupted by very physical, very analog missiles.
The Numbers
Energy accounts for up to 60% of data center operating expenses. A single Nvidia B200 Blackwell chip draws 1,200 watts — nearly double the H100. Scale that across millions of GPUs in hyperscaler clusters, and you get a sector whose marginal cost structure is directly indexed to the energy market.
Dutch TTF natural gas rose above 50 euros per megawatt-hour from 30 euros in late February. Electricity prices have surged 36% since 2020. The WTO’s chief economist warned that a prolonged Iran crisis could “crimp” the AI boom.
Reuters Breakingviews put it starkly: AI capital expenditure accounted for 39% of U.S. GDP growth in the first three quarters of last year. That’s higher than during the dot-com boom’s 28%. The entire growth thesis of the Western economy now runs through data centers that run on energy that runs through a strait controlled by a country currently at war.
The 1970s Parallel
The historical analogy is uncomfortably precise. U.S. productivity growth was above 3% in the 1960s. After the Arab oil embargo and Iranian revolution, it crashed to 0.4% between 1977 and 1982. Factory utilization dropped from 89% to 71%.
But the mechanism matters more than the magnitude. Falling revenue made executives cut investment in new technology. The economists call it reduced “capital deepening” — companies stopped buying machines when the machines became expensive to run.
The 2026 equivalent: CEOs slashing AI rollout programs because cloud computing costs just doubled. Not because AI doesn’t work. Because the electricity to make it work is suddenly a line item that the CFO can’t ignore.
An OECD study found that each 10% spike in energy prices corresponds to a 1% hit to labor productivity. Mild increases actually help — they drive investment in efficiency. But severe shocks have persistent, negative effects that last years.
We are not in the “mild” category. 8 million barrels per day wiped from global markets is the largest oil supply disruption in recorded history.
The Irony Stack
Layer it up:
- AI promises to make the economy more productive
- AI requires massive energy to operate
- Energy supply is disrupted by geopolitics
- Energy costs reduce the economic surplus available to invest in AI
- Reduced AI investment slows the productivity gains that justified the energy spending
It’s a feedback loop, but the wrong direction. The same productivity revolution that was supposed to save the economy might be its most energy-vulnerable component.
Microsoft didn’t sign a 20-year nuclear power deal because nuclear is trendy. They signed it because they looked at their GPU clusters and realized they were building the largest single-industry energy dependency since aluminum smelting.
What This Actually Means
This is not a doom forecast. It’s a structural observation.
The AI boom was priced as if compute were the only bottleneck and compute costs would only go down. Energy was background noise — cheap, reliable, invisible. That assumption held while natural gas was cheap and geopolitics were stable.
Neither condition holds now.
The companies that survive this aren’t necessarily the ones with the best models. They’re the ones with the best power purchase agreements. The ones who locked in energy contracts before the strait closed. The ones building near nuclear plants instead of near fiber hubs.
Jensen Huang’s Five Layer Cake — energy at the bottom, chips above, infrastructure, models, applications — suddenly reads less like a framework and more like a vulnerability assessment. The stack is only as strong as its lowest layer.
And right now, the lowest layer is on fire.
The AI industry built the largest compute infrastructure in history. Then it discovered that infrastructure has dependencies. The Strait of Hormuz is 21 miles wide. The entire AI boom fits inside it.