The paradox writes itself. We are building artificial intelligence on old-fashioned scarcity. The new moat is not silicon, it is fuel. If access to power has become access to energy, then the next few years will test not who can code, but who can procure molecules and navigate politics. The bigger the campus, the more brittle the bet.
A massive data center pipeline is rushing toward places with gas in the ground and space on the map. Planned capacity has swollen into the hundreds of gigawatts, with Texas absorbing an outsized share. Developers are no longer waiting on utilities. They are designing their own power plants next to servers and calling that resilience. It looks decisive. It is also a shift from one bottleneck to another. The US gas network is already tight. New gas generation far from supply means new pipelines, multi-year lead times, right-of-way fights, and permitting risk. You can escape the substation queue by going off-grid, but you cannot escape physics or lead times. If the thesis is speed, the supply chain did not get the memo.
The grid was a commons that pooled risk. It spread weather, fuel, and maintenance shocks across regions and assets. When hyperscalers defect from that commons, what follows is game theory, not harmony. In a prisoner’s dilemma, rational actors defect because they trust only their own backup plans. The result is a system where everyone pays more for less robustness. Off-grid turbines and on-site substations do not add reserve margin to the network that households and small businesses rely on. They subtract scarce turbines, engineers, and transformers from the common pool. The short-term equilibrium is private uptime and public fragility. The long-term equilibrium is political.
Onsite gas is marketed as self-reliance. It is dependence in a new key. Gas-fired data campuses are now price-takers in a market that clears against LNG exports and weather. When global shocks push LNG prices higher, domestic gas follows with a lag. A shift from two-dollar gas to four-dollar gas is not exotic. It is the base case later this decade. That turns compute into a fuel derivative. Homer City in Pennsylvania, being remade into a 4.5 gigawatt gas-powered data campus, tells the story. You gain proximity to generation, yet inherit fuel deliverability, winterization, and water risks. Batteries on site can shape ramps, not seasons. They buffer minutes and hours, not weeks of cold snaps. Independence that fails under correlated stress is not independence.
The capital structure is flashing its own signal. A sliver of mega projects absorbs a huge share of dollars. A few platforms dwarf the rest, with budgets in the tens to hundreds of billions and little incremental IT advantage over existing hyperscale builds. Financial engineering fills the gap. Industrial revenue bonds where the developer pays itself are clever in the short run and fragile in the long run. Tax frameworks change when the public notices rising bills and private moats. History repeats with new costumes. The railroad era had land grants. The dot-com era had dark fiber. Overbuilds can be bargains for the next owner. But turbines that need fuel are not stranded fiber. Their carrying costs rise with commodity cycles and maintenance, and their exit optionality narrows when gas tightens.
Texas looks like a safe harbor because it has gas, land, a permissive market, and speed. It is also islanded. ERCOT does not lean on neighbors when the weather turns. Winter storm Uri showed what frozen wellheads and unprepared plants do to gas deliverability and electric reliability. Off-grid data campuses in West and North Texas will not be immune to freeze-offs or compressor outages. Cooling water is another quiet constraint in dry places. Meanwhile, turbine manufacturing is finite. Every machine diverted to a private campus is one not available to a utility trying to meet load growth for electrification. Scale without slack invites the Seneca effect: slow gains, fast losses when stress arrives.
Batteries are showing up in these designs, but mostly as fast-ramping complements to gas and renewables. They smooth volatility. They do not conjure firm energy. The queue for generation and storage, already in the thousands of gigawatts, is evidence of intent, not deliverability. Nuclear is drawing fresh interest for good reasons: high-capacity factor, low carbon, and dense energy. It also has multi-year development timelines, constrained supply chains, and a limited number of proven developers. Even if nuclear accelerates, it will not carry the entire AI load this decade. The uncomfortable truth is that today’s AI buildout is married to gas, and the relationship will be tested by commodity prices, winter weather, and turbine supply.
If bills rise and reliability wobbles, states will intervene. They always do when private plans create public costs. Expect new capacity contributions for off-grid load, emissions oversight that treats behind-the-meter gas as scope a, not a loophole, and rate designs that claw back perceived subsidies. Industrial bonds that were tax-efficient on paper can look like arbitrage under a different legislature. Interconnection rules can change. Priority queues can be re-ranked. Curtailment can be mandated. The investor story that assumes de-risked, contracted cash flows may meet a policy story that rewrites who pays and when. The option value here is political, not technical.
The distribution of outcomes is skewed. Weather and fuel shocks are not independent. Demand growth from AI, crypto, and electrification is correlated. Global LNG markets tie local gas to geopolitics. When the tail moves, it moves many variables at once. The International Energy Agency expects data center electricity use to more than double in a few years. US power demand is on track to set records. Northern Virginia’s rising residential bills offer a preview of voter attention. The expected path is tight supply, rising gas burns, and higher baseline prices. The variance comes from shocks that are more likely than the models suggest. Planning as if averages will hold is not risk management. It is wishful thinking.
Antifragility in this domain is not found in mega campuses that only work in fair weather and cheap gas. It comes from slack and diversity. Smaller, modular builds tied to multiple fuels and grids. Workloads that can flex or pause when prices spike. Real efficiency in models and hardware, not just more racks. Contracts that share risk between developers, fuel suppliers, and offtakers instead of dumping it onto the grid or taxpayers. Waste heat reuse where climate and industry make it practical. And yes, more grid investment, because the grid remains the best tool we have to pool risk across time and place. We will still build big. The inversion is to prize options over monuments. The market tends to learn this only after it pays for the lesson.
