The Hook: Buying the Source Code
In late 2024, the polite fiction of “green energy credits” finally collapsed. For a decade, Silicon Valley has claimed to be 100% renewable by purchasing Renewable Energy Certificates (RECs). This mechanism essentially involved paying a wind farm in Texas to offset a data center in Virginia. It was a paper accounting trick that worked when power was cheap and abundant, allowing corporations to claim carbon neutrality without physically altering their energy supply chains.
That era is over. As 2026 approaches, the paper trading has been replaced by physical acquisition.
In a stunning pivot, Microsoft didn’t just buy credits; the company effectively bought a resurrection. By signing a 20-year power purchase agreement (PPA) with Constellation Energy to restart Three Mile Island Unit 1—now delivering 835 megawatts of carbon-free power—the tech giant signaled that the grid is no longer reliable enough for AI. Amazon attempted to secure 960 megawatts directly from the Susquehanna nuclear plant. The plan was to plug their servers directly into the reactor’s busbar, effectively operating “behind the meter” and bypassing the public utility infrastructure entirely.
This is no longer about sustainability. It is about Baseload Reliability. AI models do not sleep, and neither can their power source. The sector has entered the “Joule Wars,” a zero-sum game where Big Tech leverages its trillion-dollar balance sheets to privatize the most stable generation assets in America. This strategy leaves the rest of the economy (and the aging public grid) to fight for the scraps of available capacity.
The Physics of the Problem: 24/7 vs. The Weather
To understand why Microsoft is resurrecting a nuclear plant, one must look at the physics of a modern H100 GPU cluster. Unlike a steel mill, which can ramp down production during peak pricing hours, or a residential neighborhood where thermostat adjustment is possible, an AI training cluster is a rigid, gargantuan load.
A single rack of Nvidia NVL72 servers consumes approximately 120kW. A modern gigawatt-scale data center houses thousands of these racks. The load profile is a flat line; it needs maximum power, every second of every day (8,760 hours per year). This demand curve is fundamentally incompatible with the generation curve of renewable energy.
The Intermittency Gap
Renewables like solar and wind have a “capacity factor,” which is the percentage of time they produce maximum power.
- Solar PV: ~25% capacity factor (Sun doesn’t shine at night).
- Onshore Wind: ~35% capacity factor (Wind is variable).
- Nuclear: >92% capacity factor.
To run a 1 GW data center on solar, engineers would need to build roughly 4 GW of solar panels and a massive battery array to store energy for the night. This “firming cost” triples the price of electricity. Nuclear, however, generates steady power 24/7, perfectly matching the AI load profile without the need for massive storage overhead.
Big Tech has done the math. It is cheaper to pay a premium for nuclear than to build the battery infrastructure required to make wind reliable enough for a five-nines (99.999%) uptime requirement. As the calendar turns to 2026, the cost of batteries has not fallen fast enough to close this gap.
The Thermodynamics of Compute
Beyond simple electrical load, the “Joule Wars” are driven by the second law of thermodynamics. Every watt of electricity that enters a GPU is eventually converted into heat. A 1 GW data center is not just an electrical strain; it is a thermal bomb.
Handling this heat rejection requires industrial-scale cooling infrastructure. Traditional air cooling is no longer sufficient for the density of modern AI racks. The industry is shifting to liquid cooling, which adds another layer of complexity and energy demand. Pumping coolant through thousands of racks requires its own substantial power budget, further increasing the “parasitic load” of the facility.
This thermal constraint is why co-location with nuclear plants is so attractive. Nuclear sites already possess massive water intake and discharge permits, originally designed to cool the reactor core. Repurposing this infrastructure to cool server racks represents a massive capital efficiency. Instead of building new cooling towers and fighting for new water rights, tech companies can simply tap into the existing thermal rejection systems of the nuclear plant.
The PJM Shock: A Warning Signal
The financial implications of this demand shock are already visible in the capacity markets. In the PJM Interconnection (the grid operator serving 65 million people in the Mid-Atlantic and Midwest) capacity prices for the 2025/2026 delivery year skyrocketed. Prices jumped from roughly $28.92/MW-day to a staggering $269.92/MW-day. This represents an 800% increase in the cost just to ensure power is available.
This massive spike was driven largely by two factors: the retirement of aging fossil fuel plants and the sudden, unpredicted rise in load from data centers. When supply contracts and demand expands, prices explode. For residential customers, this translates directly into higher monthly bills. The AI revolution is not just costing compute credits; it is beginning to cost households in Pennsylvania and Ohio hard currency.
Contextual History: The Utility Death Spiral
For 100 years, the “Regulatory Compact” governed electricity. Utilities were granted local monopolies in exchange for an obligation to serve everyone at reasonable rates. Industrial giants like Alcoa or Dow Chemical paid lower rates because they were stable, predictable customers.
The Amazon-Talen Energy deal threatened to shatter this compact.
Amazon proposed to co-locate their data center physically at the Susquehanna nuclear plant. By connecting “behind the meter,” the company would effectively avoid paying transmission and distribution fees (the costs that maintain the wires and poles for everyone else). This would have created a precedent where the wealthiest entities could opt out of the communal infrastructure costs.
The FERC Rejection
In November 2024, the Federal Energy Regulatory Commission (FERC) rejected the amended interconnection service agreement for this project. The ruling was a rare rebuke of Big Tech. The logic was simple: if the wealthiest company in the world buys the power generator and stops paying for the grid, the remaining cost burden falls on grandmothers and small businesses.
This ruling effectively declared that the grid is a shared national asset. Participants cannot opt-out of the maintenance payments simply because they bought the power plant next door. It set the stage for the conflict seen in 2025: Tech companies are now forced to participate in the grid, but they are using their capital to outbid everyone else for available capacity.
The Forward-Looking Analysis: 2026 and Beyond
As the industry moves deeper into 2025 and looks toward 2026, the “Joule Wars” will shift from legal battles to physical constraints.
1. The Heavy Industry Squeeze
Market observers are already seeing “load shedding” requests in PJM. If a data center bids $150/MWh for steady power, a steel manufacturer operating on thin margins at $60/MWh cannot compete. The United States risks de-industrializing the “Rust Belt” not because of labor costs, but because the electricity was sold to chatbots. The economic multiplier of a steel mill (jobs, supply chain) is vastly different from a highly automated data center, creating a potential political flashpoint.
2. The SMR Pivot and Timeline Gaps
With existing nuclear plants essentially sold out, the focus is shifting to Small Modular Reactors (SMRs). Google’s 500MW deal with Kairos Power serves as the blueprint here. Unlike the bespoke, decade-long construction projects of the past, SMRs are factory-built. However, the first commercial electrons from these won’t flow until 2030. The gap between 2025 and 2030 is the “Danger Zone” for grid stability. The only immediate solution is keeping aging coal and gas plants online longer than planned.
3. The Natural Gas Bridge
Denied immediate access to nuclear and waiting on SMRs, the unspoken reality is that Big Tech is quietly funding a massive expansion of natural gas peaker plants. Despite net-zero pledges, the immediate need for 24/7 power is driving a fossil fuel renaissance. The carbon emissions of the AI sector will likely rise, not fall, over the next three years as gas fills the gap left by insufficient nuclear capacity.
4. The Rise of “Sovereign Grids”
Expect to see Big Tech explore “Sovereign Grids” (completely islanded power systems that do not touch the public infrastructure at all). If FERC blocks co-location, companies might build new SMRs and data centers in remote locations, completely off-grid. This would create a two-tier energy system: a high-reliability, privatized grid for AI, and a crumbling, expensive public grid for everyone else.
The Bottom Line
The Joule Wars are not just about electricity; they are about the hierarchy of the American economy. By purchasing the rights to the nation’s most stable power, Big Tech has signaled that computing compute is now the most valuable output of the United States economy (more valuable than steel, chemicals, or manufacturing).
The grid was designed for lightbulbs and washing machines, not for gigawatt-scale thinking machines. As the electrons flow toward the highest bidder, the nation is about to find out just how elastic its infrastructure really is. The result may be a shiny new AI future, powered by the most expensive electricity in history.
🦋 Discussion on Bluesky
Discuss on Bluesky