Why AI is running out of Power

Summary

This episode reframes AI infrastructure as an electrical-grid problem, not a semiconductor problem. SemiAnalysis argues that GPU clusters are effectively machines that turn electricity into tokens, so AI demand ultimately becomes power demand. Because datacenter timelines move in months while grid timelines move in years, companies are using behind-the-meter generation, BYOG, and site selection arbitrage to accelerate deployment. The discussion focuses on the practical mechanics of onsite power—turbines, reciprocating engines, fuel cells, redundancy, inertia, and frequency stability—as well as the labor and supply-chain constraints behind them. The broader conclusion is that AI leaders are becoming de facto energy buyers and infrastructure developers, and that the next phase of the AI race may be decided by who can secure credible power fastest.

Notes

• AI infrastructure is framed as a power-grid bottleneck, not a GPU or HBM bottleneck.
• A GPU cluster is described as an industrial machine that converts electricity into tokens.
• AI power demand was said to rise from about 3 GW in 2023 to more than 28 GW by 2026.
• Expected new datacenter gross power demand was projected at 84 GW by 2030.
• A gigawatt-scale AI campus is economically enormous because an AI cloud can generate $10B-$12B per GW per year.
• A 200 MW cluster online six months earlier is worth roughly $1B.
• The industry is using BYOG as a bridge strategy: Bring Your Own Generation.
• Behind-the-meter generation can include gas turbines, reciprocating engines, fuel cells, batteries, or hybrid systems.
• The datacenter needs power in 2027 or 2028 while grid interconnection may not arrive until 2030.
• Gigawatt-scale AI facilities are compared to industrial factories manufacturing intelligence.
• Roughly 1 TW of load requests were said to be in the US queue, exceeding a 700-800 GW national grid peak.
• AEP Ohio had 35 GW of load requests in October 2024, and 68% lacked land control.
• The interconnection process was described as taking around five years for most generation types.
• SemiAnalysis says available grid headroom is approaching zero and turns negative by 2027.
• xAI stood up a 100,000-GPU cluster in four months, showing how speed matters in power procurement.
• xAI’s energy strategy was onsite generation, not waiting on the grid.
• xAI used 16 MW Solar Turbines units, highlighting truckable modular generation.
• xAI rented turbines from Solaris Energy Infrastructure and leased VoltaGrid truck-mounted gas engines.
• xAI picked a border site between Tennessee and Mississippi to exploit permitting differences.
• xAI is said to have more than 500 MW of turbines deployed near its datacenters.
• OpenAI and Oracle reportedly placed a 2.3 GW onsite gas generation order in Texas.
• The onsite gas market was described as triple-digit annual growth.
• Twelve suppliers reportedly each have more than 400 MW of US datacenter orders.
• Doosan Enerbility booked a 1.9 GW order to serve xAI.
• Wärtsilä signed 800 MW of US datacenter contracts.
• Boom Supersonic announced a 1.2 GW turbine contract with Crusoe.
• Aeroderivative turbines are described as jet engines bolted to the ground.
• Aeroderivatives were said to run 30-60 MW, ramp in 5-10 minutes, cost about $1,700-$2,000/kW all-in, and take 18-36 months lead time.
• Industrial gas turbines were described as 5-50 MW units with roughly 20-minute ramp times and $1,500-$1,800/kW cost.
• Reciprocating engines were compared to oversized car engines, with 11 MW units over 14 meters long.
• Reciprocating engines ramp in about 10 minutes and handle heat, dust, and dirty fuel better than turbines.
• Solid-oxide fuel cells generate power electrochemically with no combustion.
• Fuel cells cost $3,000-$4,000/kW and their stacks last 5-6 years.
• The market rewards the vendor with an open order book and a credible delivery date.
• Meta and Williams’ Ohio site, Socrates South, uses a mixed fleet of turbines and engines from multiple vendors.
• The transcript says nobody designs a power plant that way on purpose; it reflects whatever can be deployed on time.
• The US grid delivers about 99.93% uptime through system-wide redundancy.
• Behind-the-meter sites must overbuild to reproduce grid reliability for a single customer.
• Vendors typically require at least N+1 redundancy, with N+1+1 better for maintenance coverage.
• A 200 MW site using 11 MW reciprocating engines could deploy 26 units for 286 MW nameplate.
• A 200 MW site using 30 MW aeroderivatives could use 9 units for 270 MW nameplate, with extra units needed in hot climates.
• Crusoe’s Abilene site for Oracle and OpenAI uses five GE Vernova LM2500XPRESS aeros and five Solar Titan 350s.
• Vantage’s Shackelford County campus is paired with 2.3 GW of VoltaGrid systems, implying a 64% overbuild.
• Behind-the-meter is not cheaper; it is earlier.
• AI training loads can surge and dip on a sub-second basis, creating power-quality challenges.
• Sites add synchronous condensers, flywheels, and batteries to stabilize the load.
• Traditional datacenters were built around three-, four-, or five-nines uptime, but AI changes those assumptions.
• Training can tolerate lower availability via checkpointing and recovery.
• At Abilene and Memphis, training clusters were built without diesel generator backup.
• Running behind the meter means staffing and operating a power plant 24/7.
• A 2.3 GW high-speed engine site could imply roughly 500 engines and more than 2,000 service events per year.
• The gas turbine industry ran at production lows of under 10 GW per year from 2017 through 2022.
• GE Vernova says it will increase output to 24 GW per year, and Siemens Energy plans to exceed 30 GW by decade-end.
• Turbine blades and vanes depend on single-crystal nickel alloys and a narrow set of Western suppliers.
• Behind-the-meter is framed as timing arbitrage, not economic optimization.
• ERCOT is an energy-only market, unlike capacity-market structures such as PJM.
• ERCOT’s April 2025 forecast projected up to 77.9 GW of datacenter load by 2030.
• ERCOT later discounted generic and officer-attested requests and pushed in-service dates back 180 days.
• PJM capacity pricing is described as a simulation based on the Variable Resource Requirement curve, not ordinary market bidding.
• During Winter Storm Fern, ERCOT peaked around $300/MWh while PJM lost about 21 GW and averaged $700/MWh system-wide.
• The AI supply chain now includes gas turbines, fuel cells, transformers, switchgear, substations, power electronics, and permitting specialists.
• Yttrium and other materials make AI power buildout sensitive to rare-earth policy.
• ProEnergy’s PE6000 is described as repurposing retired Boeing 747 engine cores into aeroderivative turbines.
• Datacenter announcements should be judged on credible power, not campus size.
• The gating checklist is grid path, onsite generation, gas pipeline, turbines, transformers, air permit, interconnection, and local load absorption.
• Air permitting for onsite generation can take a year or more.
• Texas had already delayed at least one gigawatt-scale Stargate facility.
• Behind-the-meter power creates environmental and political tension because many solutions use gas.
• The AI industry wants fast, reliable, clean, and cheap power, but these goals conflict.
• Gas is framed as a bridge technology while grids, transmission, and cleaner firm resources catch up.
• The frontier of AI is moving through power markets as well as algorithms and chips.
• The bottleneck is moving from silicon to energy.
• Behind-the-meter could power well over half of new US datacenters from 2028 onward.
• The equipment market for datacenter behind-the-meter solutions could cross 50 GW per year by 2029.
• Large AI players are becoming energy buyers, infrastructure developers, and almost power companies by necessity.
• The AI race is becoming a power race.