Your AI runs on oil

Everyone in AI talks about the compute bottleneck. Not enough GPUs, not enough VRAM, not enough cluster time. But there's another bottleneck upstream that rarely makes it into the conversation, and it's far harder to solve with engineering. It runs on fossil fuels, travels through one of the most contested waterways on the planet, and right now, it's under direct threat. The Strait of Hormuz, a narrow passage between Iran and the Arabian Peninsula, carries roughly 20 million barrels of oil per day. That's about 20% of global petroleum consumption and 27% of all seaborne oil trade. It also handles around 20% of the world's liquefied natural gas shipments. When Iran announced the strait would be "completely closed" in response to U.S. threats against its power infrastructure, it didn't just rattle oil traders. It put the physical supply chain behind every major AI system in the world at risk.

The energy footprint of AI is enormous

The AI industry's appetite for electricity is growing faster than almost any other sector. U.S. data centers consumed 183 terawatt-hours of electricity in 2024, more than 4% of the country's total, roughly equivalent to the entire annual electricity demand of Pakistan. The IEA projects global data center electricity consumption will double to around 945 TWh by 2030. These aren't abstract projections. OpenAI's Stargate data center will consume approximately 1.2 gigawatts of power, half the total peak load of El Paso Electric, the regional utility serving the project. The average hyperscale data center demands around 500 megawatts, enough to power 375,000 homes for a year. Goldman Sachs forecasts global power demand from data centers will increase 165% by the end of the decade compared with 2023. Deloitte estimates that AI data center power demand in the U.S. alone could grow more than thirtyfold by 2035, reaching 123 GW. Training a single frontier model like GPT-4 required around 30 megawatts of sustained power. The next generation of models will need multiples of that. Every watt comes from somewhere, and right now, a significant share of that somewhere is fossil fuels.

Where the power actually comes from

The conversation around AI energy often jumps straight to nuclear and renewables. Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart the 835 MW Three Mile Island Unit 1 reactor, expected to come back online in 2028. Amazon Web Services signed a deal with Talen Energy for a 960 MW data center campus in Pennsylvania. Meta struck an agreement with Constellation to keep the Clinton Clean Energy Center in Illinois running past 2027. These deals are real, but they're also years away from delivering power at scale. In the meantime, the grid runs on what's available. Natural gas accounts for a growing share of U.S. electricity generation, and data center operators are increasingly signing gas procurement deals directly. Atmos Energy announced a contract to provide 30 billion cubic feet per year to a data center with collocated gas-fired generation. S&P Global projects gas will comprise 29% of North American electricity demand by 2035. And when the grid falters, data centers fall back on diesel. Backup diesel generators are standard equipment at virtually every facility. In June 2025, a heatwave in Virginia forced data centers to run on diesel generators for demand response. A substation fire in Loudoun County, the densest data center market in the world, sent facilities onto backup diesel power for over 24 hours. Nearby residents reported the sound of "planes landing constantly" and the smell of diesel fumes from a Walmart parking lot. The nuclear and renewable future is coming, but the present runs on gas and diesel. Both of those depend on a global supply chain that is, at this very moment, under severe stress.

The strait is not a distant abstraction

The Strait of Hormuz isn't just an oil story. When Iran effectively closed the strait, it disrupted crude, refined fuels, and LNG simultaneously. Iranian missiles struck the Ras Laffan LNG terminal in Qatar, the world's largest, sending Brent crude above $107 a barrel. Refineries across the Gulf have been damaged or shut down. Saudi Aramco closed its largest refinery at Ras Tanura after a drone strike. Bahrain's Bapco Energies Sitra refinery declared force majeure. The Dallas Federal Reserve modeled the impact: removing close to 20% of global oil supplies from the market during Q2 2026 is expected to raise WTI oil prices to $98 per barrel and lower global real GDP growth by an annualized 2.9 percentage points. The head of the IEA called it a "major, major threat" to the global economy. For the AI industry, the consequences cascade quickly. Natural gas prices spike alongside oil. Grid electricity costs rise. Diesel for backup generators gets more expensive and harder to source. Every link in the energy chain that keeps a GPU cluster running gets stretched thinner.

Energy is the other chokepoint

The tech industry is already familiar with supply chain chokepoints. Chip export controls demonstrated how a single policy decision could reshape the global semiconductor landscape overnight. Energy is the same kind of vulnerability, but less visible and less discussed. Consider the parallel. A frontier AI lab can spend billions on GPU clusters, negotiate priority access to the latest chips, and optimize every layer of their training stack. But none of that matters if the electricity to run those chips becomes unreliable or prohibitively expensive. The most powerful cluster in the world is just a very expensive room full of metal if the power goes out. S&P Global has projected a copper supply deficit of up to 10 million metric tons by 2040, driven by AI infrastructure and accelerating defense spending. Copper is essential for power transmission, the wiring that connects generation to data centers. Even the physical infrastructure to deliver electricity is under strain. The AI industry talks about scaling laws for compute. It needs to start thinking about scaling laws for energy, and the geopolitical risks that come with them.

What happens to your API bill

The downstream effects aren't hypothetical. When energy costs spike, cloud providers absorb them temporarily, then pass them on. Data center operators with diversified energy sources and flexible logistics networks will weather disruptions better than those locked into single supply chains. But the broader market feels the impact. If the Strait of Hormuz remains closed through Q2 2026, energy-intensive industries, including AI, face a sustained period of elevated costs. Training runs get more expensive. Inference costs rise. The economics of offering AI services at current price points get harder to justify. That $200/month API subscription assumes a stable energy supply chain behind it. The firms best positioned to navigate this are the ones already investing in energy resilience: diversified generation sources, long-term PPAs, on-site generation, and the logistics networks to reroute fuel supplies when primary routes are disrupted. The rest are exposed to a risk they may not have fully priced in.

The uncomfortable truth

AI doesn't run on the cloud. It runs on electricity. And a meaningful share of that electricity, directly or indirectly, runs on oil and gas that flows through some of the most geopolitically volatile corridors on Earth. The industry's long-term bet on nuclear and renewables is the right one. But "long-term" means years, not months. In the short and medium term, the physical reality is that a regional conflict thousands of miles from Silicon Valley can disrupt the energy supply chain that powers every major AI system on the planet. Compute is a bottleneck. Chips are a bottleneck. But energy is the bottleneck upstream of all the others, and it's the one that no amount of algorithmic optimization can solve.

References