Amazon just backed a 7.65 GW natural gas plant in West Texas. The stated destination is AI data centres. The unstated destination is the dark corner of every renewable-energy deck that claimed batteries could fix the grid.
Seven point six five gigawatts is not an incremental PPA. It is a private power station. At an 85 percent capacity factor, it produces roughly 57 TWh per year. That is somewhere between one-third and one-half of the electricity the entire Bitcoin network consumes today. I count the cracks before the dam breaks.
This is not a one-off procurement. It is a structural admission.
The project sits in the Permian Basin, with Henry Hub gas near $2.50-$3.50 per MMBtu and a pipeline supply chain that already works. Electricity from a modern combined-cycle plant runs at roughly $0.04-$0.06 per kWh when fuel is cheap. That figure matters because ERCOT, the Texas grid, has become a volatility machine. Interconnection queues stretch two to four years. Peak prices broke $1 per kWh multiple times in 2023 and touched $5 per kWh in August. The 2021 winter storm knocked wind output from normal levels down to less than five percent of nameplate capacity. Anyone who needs always-on power learned the same lesson: the central grid is a weather-tied derivative with severe tail risk.
So Amazon is not buying electricity anymore. It is converting an operating expense into a capital expenditure. The reported structure is likely not full ownership. Amazon is more likely standing behind a third-party developer with a long-term power purchase agreement, or PPA. That is the classic half-integrated play: someone else owns the pipe, Amazon owns the electrons. It is also the smartest structure for a hyperscaler that wants supply certainty without becoming a regulated utility.
The technical logic behind the choice is even more important than the financial structure.
The Battery Illusion
The battery alternative fails before it reaches the cost sheet. A 7.65 GW load that needs four hours of coverage requires 30.6 GWh of storage. At current system EPC costs for lithium iron phosphate, roughly $70-$110 per kWh, that is $2.1-$3.4 billion of storage before power electronics, grid interconnects, or land. The bigger problem is duration. A four-hour battery cannot carry a multi-day winter storm. A data centre wants reliability above 99.99 percent. Storage on a baseload profile would cycle maybe 200-300 deep cycles per year. Battery LCOS assumptions only start to favour batteries past 1,000 cycles annually. The gas turbine, meanwhile, is designed for 7,500-8,000 running hours per year. That is not a preference. It is an engineering limit.
Solar has the same hump. West Texas solar alone can produce power at $0.03-$0.04 per kWh when the sun is out. But to serve a 7.65 GW baseload facility, you would need 15-20 GW of solar plus 30 GWh of storage and 60-100 square kilometres of land. The gas plant needs two to four square kilometres. The system LCOE of a solar-plus-storage stack rises to $0.09-$0.15 per kWh. The gas combined-cycle plant, including fuel and carbon compliance, sits at $0.05-$0.08 per kWh. That arithmetic decides the project before the first turbine order is placed.
Hydrogen is financially absent. Green hydrogen at $3-$5 per kilogram translates to generation cost around $0.18-$0.30 per kWh, three to six times gas. Gas turbines can only burn about 30 percent hydrogen in commercial service today; 100 percent combustion is a 2030-era promise. Amazon could be installing machines that are ready for future blending, but it is not buying a hydrogen plant now. The gas plant might later carry CCS, or carbon capture, and IRA credits could change that math. Still, the immediate fuel is not hydrogen. It is molecules that are already in the ground.
I learned this the hard way during the DeFi summer of 2020. I spent months watching AMM liquidity models break under Ethereum gas-price spikes. The model that works in an Excel sheet breaks when the settlement layer gets congested. Replace AMM gas with natural gas and the same pattern is here. The ledger bleeds faster than the logic holds.
The Supply Chain Crunch
A 7.65 GW combined-cycle plant with GE 7HA-class machines, each rated near 400-500 MW, would need roughly 15-19 heavy turbines. The global supply of heavy gas turbines across GE Vernova, Siemens Energy, and Mitsubishi is around 200-300 units per year. AI buildouts are already competing with LNG export projects for those machines. GE Vernova’s gas turbine order book hit records in 2024, with delivery slots stretching into 2027 and 2028. The delivery queue, not the land permit, is the real finality problem.
This is where the crypto analogy becomes too clean to ignore. The gas turbine queue is the new block gas limit. Every large buyer knows it. The ones with signed contracts win; the ones waiting on a 2028 delivery slot are simply another unconfirmed transaction on an oversubscribed network.
Gas Economics and Grid Anxiety
The cost structure is not fixed because fuel is 60 to 75 percent of the stack. If Henry Hub rises by $1 per MMBtu, combined-cycle power cost increases by roughly $0.008-$0.01 per kWh. From $3 to $5 gas, the plant’s output moves from $0.04-$0.06 to $0.07-$0.09 per kWh. That is still a hedge against ERCOT spot. In fact, the gas plant is best understood as a pair of options: long physical gas and short grid-price volatility. It is the energy equivalent of a covered call. Liquidity is just borrowed time with a premium.
The strategic consequences extend well beyond one facility. US data centre electricity demand is projected to grow from roughly 140 TWh in 2023 to 300-500 TWh by 2030. Meeting that growth could require 150-250 GW of new generating capacity. Gas is the fastest adult in the room: construction in three to four years, capex near $800-$1,200 per kW, versus nuclear at $6,000-$9,000 per kW. Even if every nuclear restart and SMR project succeeds on schedule, gas will absorb the gap. The turbines already prove it. This is not a clean-energy defeat. It is a dispatchability victory.
The Contrarian Read
Now the contradiction that every positive AI-energy narrative is trying to avoid. Amazon is one of the largest corporate renewable buyers on earth, with more than 20 GW of signed PPAs. Yet in its most important power procurement decision, it chose dispatchable hydrocarbons. The reason is simple: many corporate “100 percent clean energy” claims are annual accounting constructs. A company can match total consumption with RECs and wind PPA energy while every physical megawatt during peak load comes from a gas turbine. That is the same logic as a carbon offset. The balance sheet looks clean. The electrons do not know.
There is also regulatory arbitrage. Texas has no state carbon price, no state income tax, and a permitting environment far cheaper and faster than California or the Northeast. Add IRA’s 45Q credit for carbon capture, up to $85 per ton, and a 7.65 GW plant with 90 percent capture could produce around 24 million tonnes of CO2 per year. That would generate more than $2 billion in annual tax credits before the plant even sells power. CCS is no longer just an environmental token. It is a project-finance layer.
From an options perspective, the market consensus is still looking at the wrong underlying. Retail sees AI compute and token narratives. Institutions see electrons and delivery. Microsoft chose nuclear via Constellation. Google chose SMR offtake. Amazon chose gas and pipeline optionality. Every one of them is doing what serious traders do: stop fighting the basis, take delivery. They are not exiting the grid. They are building a private settlement layer.
For Bitcoin miners, the signal is unsettling. Marginal dispatchable power is no longer being priced by a household during a summer heatwave. It is being priced by hyperscalers signing long-dated contracts for 7.65 GW blocks. Miners are flexible, interruptible load. In a strained grid, that makes them the short side of the trade. They can shut off and sell power back, but that is not a growth thesis. That is a hedged exit strategy. The era of cheap stranded power for crypto is tightening.
Takeaway
The metrics to watch are not sentiment charts. Watch the turbine delivery queue. Watch whether Amazon locks a 20-year fixed gas supply contract. Watch whether the project adds battery storage for fast frequency response. If gas-plus-storage becomes the hyperscaler template, then the clean-energy base case needs a rewrite. If pure gas wins, then every climate-friendly AI deck should be archived as marketing material.
The next bull market in crypto may not be defined by a token at all. It may be defined by who controls the electrons. Amazon just took a 7.65 GW block off the table. Survival is the only alpha that compounds.