---
slug: "evicting-the-server"
title: "Evicting the *Server*"
subtitle: "The Basal Grid: Why Hyperscalers Became Nuclear Utility Monopolies."
volId: "vol-003"
volMonoId: "003-001"
monographNumber: 11
publishedDate: "2026-05-08"
author: "Caleb Brown"
readingTime: "7 Min"
editorialName: "Vanishing Acts"
isFeatured: true
excerpt: "Microsoft, Amazon, and Google are no longer pure software providers; they are de facto power companies financing reactor restarts and monopolizing baseload electricity to sustain Blackwell inference clusters."
featuredImage: ""
galleryCaption: "When compute outstrips the grid, the cloud ceases to be software; it becomes an electrical utility monopoly."
tags:
  - "Nuclear Energy"
  - "Datacenter Power"
  - "Cloud Infrastructure"
---

# Evicting the *Server*

In September 2024, Microsoft committed billions over twenty years to resurrect an eighty-foot pressurized water reactor on an island in the Susquehanna River.

The facility was Unit 1 of Three Mile Island. Decommissioned in 2019 against cheap shale gas, the reactor sat cold, its turbines locked and fuel assemblies submerged in cooling ponds. Under an agreement with Constellation Energy, the plant will be reëngineered, recertified by the Nuclear Regulatory Commission, and renamed the Crane Clean Energy Center. 

Its entire 835-megawatt electrical output will flow to a single customer. 

Not one watt will heat a home, illuminate a classroom, or power a municipal transit system. Every electron generated by its uranium core will route through dedicated step-down transformers to run dense matrix multiplications across AI accelerator clusters.

The romantic myth of "the cloud" died on the Susquehanna.

For two decades, Silicon Valley marketed computing as weightless: *elastic scale*, *serverless functions*, *ambient intelligence*. Software was framed as an intangible ledger operating above the dirty realities of industrial hardware. Compute was treated as an infinite utility decoupled from thermodynamics.

That illusion is shattered. 

Today, computing is bounded not by instruction sets, but by high-voltage copper busbars, substation transformer lead times, and steam turbines. Hyperscalers are no longer software businesses. They have mutated into private energy conglomerates, deploying balance sheets to commandeer sovereign baseload power before the public grid recognizes what has been extracted.

The server has been evicted from the center of systems architecture. The power plant took its room.

## The Thermodynamic Wall: 130 Kilowatts per Rack

To understand why software monopolies are buying nuclear reactors, one must inspect a compute rack down to its fluid dynamics.

| Architecture Metric | Legacy Enterprise Rack (2014) | Nvidia Blackwell NVL72 (2026) |
| :--- | :--- | :--- |
| **Power Density** | 5 – 10 kW / rack | 120 – 132 kW / rack *(15x continuous density)* |
| **Cooling Architecture** | Perforated raised-floor air distribution | Direct-to-chip deionized liquid loops (>2 L/sec) |
| **Workload Profile** | Stateless web servers & serial transactions | Synchronous all-reduce tensor parallelism |
| **Failure Characteristic** | Graceful failover; Kubernetes container reschedule | Single-node brownout drops cluster gradient barrier |


For thirty years, datacenters operated on human-scale physics. A 42U rack consumed 5 to 10 kilowatts, cooled by HVAC blowers pushing chilled air through perforated raised-floor tiles. If a server overheated, distributed software absorbed the fault: a health check tripped, traffic rerouted, and Kubernetes instantiated an identical container on an adjacent node.

The Nvidia Blackwell GB200 NVL72 liquidated that model.

A single GB200 rack integrates 72 Blackwell GPUs and 36 Grace CPUs across an NVLink copper backplane, drawing between 120 and 132 kilowatts of continuous power. 

That is not an IT chassis. That is an industrial blast furnace.

At 132 kW per rack, air cooling is thermodynamically impossible. Air lacks the volumetric heat capacity to extract thermal energy at that density without hurricane velocities that would shear surface-mounted diodes from circuit boards. Facilities must route direct-to-chip liquid loops—pumping deionized water and propylene glycol through quick-disconnect manifolds across silicon cold plates—at flow rates exceeding two liters per second.

The macro economics are staggering. A frontier cluster housing 100,000 GPUs demands 150 to 200 megawatts of continuous, uninterruptible power—the continuous draw of 150,000 suburban American homes concentrated on a twenty-acre pad.

This load is uniquely unforgiving. Web applications cycle diurnally. Frontier AI training demands a flat, unyielding horizontal line. 

Tens of thousands of GPUs tied across 800-gigabit InfiniBand networks continuously exchange tensor weights via collective all-reduce operations. If a regional substation suffers a 200-millisecond voltage sag, or if a cooling loop pressure drop forces a thermal throttle on three racks, the synchronous gradient barrier fractures. 

The entire 200-megawatt cluster halts. Months of computation risk corruption. Checkpoint recovery consumes millions of dollars in wasted compute cycles. 

Hyperscalers do not need intermittent green offsets. They need absolute, unbreakable baseload electricity operating at a 99.999% capacity factor, twenty-four hours a day.

## The Intermittent Accounting Trick

For a decade, Big Tech cloaked power consumption behind a financial fiction: Renewable Energy Certificates (RECs) and net-zero purchase agreements.

A cloud vendor contracted for a 200-megawatt solar array in Nevada or a wind farm in West Texas. When generation surged, the installation injected kilowatt-hours into the regional grid. The vendor claimed environmental credits, declaring its footprint "100% matched by renewable energy."

It was a spreadsheet arbitrage that ignored electrical physics.

Solar arrays generate electricity at an average capacity factor of 20 to 25 percent. Wind turbines average between 30 and 35 percent. When twilight settles over Northern Virginia, or when summer high-pressure systems stall wind in the Ohio Valley, output drops to zero. 

A GPU cluster executing backpropagation cannot pause all-reduce synchronizations for weather fronts to clear eight hundred miles away. Bridging that chasm requires multi-day battery storage systems (BESS)—whose capital costs and cell degradation curves destroy hyperscaler margins—or fossil-fueled peakers. 

During the sixty-five percent of the year when the sun was down and the air was still, those server racks were drinking electrons generated by pulverized coal and combined-cycle gas.

With millions of high-draw accelerators entering production, this accounting fiction collapsed against grid transmission limits. Hyperscalers had to secure genuine, uninterrupted, zero-carbon thermal energy. 

There was only one physical candidate: commercial nuclear fission.

## Behind the Meter: Siphoning the Generator Bus

Once hyperscalers acknowledged that weather-dependent generation could not sustain Blackwell clusters, they initiated a hostile acquisition of America's nuclear infrastructure:

* In March 2024, Amazon Web Services (AWS) paid $650 million to acquire the Cumulus Data campus adjacent to Talen Energy’s 2.5-gigawatt Susquehanna Steam Electric Station in Pennsylvania, contracting for up to 960 megawatts of direct nuclear power.
* In September 2024, Microsoft signed its twenty-year agreement with Constellation Energy to monopolize the entire 835-megawatt capacity of Three Mile Island Unit 1.
* In October 2024, Google signed an advance framework with Kairos Power to deploy 500 megawatts across seven molten-salt Small Modular Reactors (SMRs) by 2035.

Crucially, the Amazon-Talen agreement bypassed the public grid entirely through *behind-the-meter colocation*.

| Distribution Model | Electrical Routing Path | Infrastructure & Ratepayer Impact |
| :--- | :--- | :--- |
| **Regulated Utility Model** | Nuclear Plant $\rightarrow$ High-Voltage Switchyard $\rightarrow$ Public Grid (PJM) | Capital costs, spinning reserves, and maintenance distributed across all public ratepayers. |
| **The Hyperscaler Bypass** | Nuclear Plant $\rightarrow$ Direct Co-located Busbar $\rightarrow$ AWS Cumulus Datacenter | 960 MW siphoned behind-the-meter; public grid loses clean baseload while public ratepayers absorb grid stabilization costs. |


In standard merchant generation, a nuclear plant steps voltage up to 500 kilovolts and feeds the regional transmission network. Grid operators dispatch power across high-voltage corridors, and every participant pays transmission tariffs.

In the behind-the-meter model, Amazon connected datacenters directly to the power plant’s generator bus within the station security perimeter. AWS siphons nearly a gigawatt of baseload power before it ever touches the public transmission grid.

The business rationale is ruthless: bypass regulatory interconnection delays and evade hundreds of millions in transmission fees. 

They did not build new clean energy. They walled off existing clean generation, coöpting a public asset for private corporate matrix multiplication.

## The Seven-Year Queue and the FERC Rejection

This maneuver was forced by grid paralysis.

Consider PJM Interconnection, the regional transmission organization coördinating electricity across thirteen states for 65 million people. PJM encompasses Northern Virginia's "Data Center Alley"—the dense corridor in Loudoun and Prince William counties through which roughly seventy percent of global internet traffic routes.

PJM's interconnection queue is an administrative graveyard. 

Developers seeking to interconnect new generation or massive industrial loads face delays of five to seven years for regulatory studies. Thousands of projects languish, paralyzed by staffing shortages and supply bottlenecks; lead times for high-voltage substation step-up transformers exceed four years.

A technology monopoly operating on an eighteen-month silicon cadence cannot wait seven years for a utility hookup. Behind-the-meter colocation was an end-run to skip the line.

The maneuver failed.

In November 2024, the Federal Energy Regulatory Commission (FERC) delivered a landmark ruling. In a 2-1 decision (Docket ER24-2172), FERC rejected the amended Interconnection Service Agreement allowing Amazon to expand its direct draw from the Susquehanna plant from 300 to 480 megawatts, targeting 960 megawatts.

The challenge was brought by competing utilities: American Electric Power (AEP) and Exelon. 

Their argument was mechanically undeniable: siphoning 960 megawatts of baseload off the public grid undermines regional reliability. Missing nuclear generation must be replaced by dispatching dirtier, costlier fossil reserves. Pulling power behind the meter allows Amazon to evade transmission charges, shifting tens of millions in annual grid maintenance costs onto regular utility customers.

FERC agreed. Regulators blocked Big Tech from privatizing clean baseload while leaving society to absorb the costs.

## The Socialized Circuit: Who Finances the Copper?

When tech conglomerates consume power on the scale of nation-states, the bill spills across the balance sheets of every citizen connected to the grid.

Dominion Energy’s integrated resource plans project power demand in Northern Virginia will surge by more than one hundred percent over fifteen years, driven almost entirely by datacenters. To prevent brownouts in suburbs outside Washington, D.C., Dominion and PJM are forcing through billions in new 500-kilovolt transmission corridors, substations, and delayed retirements of fossil plants.

Under cost-of-service utility regulation, capital expenditures are not paid from corporate margins. Every mile of copper wire and every substation added to the rate base earns a guaranteed return of nine to ten percent, socialized across the ratepayer base.

In Virginia, Pennsylvania, and Ohio, residential electric customers are absorbing the financial drag of the AI boom. Monthly utility bills are escalating as regulators approve rate hikes to reinforce grids built explicitly for hyperscale compute parks. Working families and municipal transit systems are paying a silent, involuntary surcharge on their electric meters to underwrite high-voltage corridors feeding LLM clusters.

The cloud has executed a profound transfer of resources. Hyperscalers have captured the private economic upside of artificial intelligence, while socializing capital costs, environmental strain, and reliability risks onto the public ratepayer. They have coöpted sovereign generation, walled off the cleanest reactors in the country, and left surrounding municipalities to manage the volatile remains of an overburdened transmission grid.

The eviction of the server is complete. 

The illusion that software exists separate from the industrial economy has dissolved into the hum of transformers and the steam of cooling towers. The cloud is no longer a software abstraction. It is an electrical utility monopoly, fortified behind private perimeter wire, sustained by the public grid, and burning through sovereign power reserves one matrix multiplication at a time.