Chapter 4: Wait... How Much Is Going Into American AI Computing??
- Sean M. Walsh
- Dec 15, 2025
- 8 min read
Updated: Jan 15
US companies are investing trillions, but there is a big missing piece.
"The whole is greater than the sum of its parts, but only if all parts are present." - Aristotle, ~350BC
Key Points:
1. American companies are investing breathtaking sums, over $1 trillion annually on AI computing infrastructure.
2. But, an imbalance has emerged, as not enough investment is going toward the electricity generation and supply.
Key Stats:
$4 Trillion - estimated US investment into AI and datacenter infrastructure from 2025-2030 (per McKinsey).
$100 Billion - size of a single Microsoft/OpenAI computing cluster announced in 2024, slated for 2028 (per Situational Awareness).
$1 Trillion - projected cost of a single AI computing cluster by 2030 (per Situational Awareness).
21% - approximate percentage of total US electricity production required for a single $1T cluster (per Situational Awareness).
$6.5 Trillion - December 2024 market capitalization of top 10 chip manufacturers (per Deloitte).
0% - approximate growth in US electricity production over the past 20 years.The Maginot Line Disaster: A Lesson In Unbalanced Investment
To understand the danger of what's happening, we need to travel back to 1930s France.
After the devastation of World War I, France committed to ensuring it would never again face the horror of invasion. The solution was the Maginot Line: a massive chain of concrete fortifications, underground bunkers, artillery installations, and obstacles stretching along the entire German border.
The Line cost approximately 3 billion French francs—equivalent to roughly $50 billion today—and took over a decade to complete. It featured underground facilities with modern living quarters for garrison troops, including air conditioning and dining areas. The fortifications could withstand direct hits from German bombers. By 1939, it was widely considered impenetrable.
On May 10, 1940, the German army began its Blitzkrieg attack.
They didn't attack the Maginot Line. Instead, German forces went around it, sweeping through Belgium in barely two days. A decoy force sat opposite the Line, giving the impression of preparing for a frontal assault, while the main thrust came around the northern flank.
In just two days, German forces were racing through France. Within six weeks, France had surrendered. The Maginot Line—with its billion-dollar fortifications largely intact—became irrelevant. Worse, it had tied up 52% of the French Army in defensive positions, unable to respond when the actual threat materialized elsewhere.
The problem wasn't that France failed to invest. The problem was that France invested massively in some links of the defensive chain while neglecting others entirely.
The Maginot Line has since become a metaphor for failed national infrastructure projects, due to inconsistent investments.
Astonishing CapEx Investment Projections For American AI
Something extraordinary is happening in American technology investment—extraordinary even by the standards of an industry built on exponential growth.
In early 2025, the four technology giants—Amazon, Microsoft, Google, and Meta—announced combined capital expenditure plans approaching $320 billion for the year. By October 2025, after successive quarters of raised guidance, that figure had climbed to over $380 billion. And these numbers represent just the publicly announced commitments from just four companies.
Amazon leads the charge with $125 billion in planned capex for 2025, up from a prior forecast of $118 billion. Microsoft's spending will reach approximately $140 billion including capital leases—triple what it spent in fiscal 2024. Google has boosted its capital expenditure guidance multiple times, now targeting $91-93 billion. Meta, despite lacking a cloud business, plans to spend between $70-72 billion building AI data centers.
The semiconductor industry is riding this wave to historic heights. According to Deloitte's 2025 global semiconductor outlook, the industry posted $627 billion in sales for 2024—a 19% increase over the prior year—and is projected to reach $697 billion in 2025. The industry is on track to reach the widely accepted aspirational goal of $1 trillion in chip sales by 2030.
The market has noticed. As of mid-December 2024, the combined market capitalization of the top ten global chip companies stood at $6.5 trillion—up 93% from $3.4 trillion just one year earlier, and 235% higher than the $1.9 trillion seen in November 2022. In just two years, chip companies added more value than the entire GDP of Germany.
Nvidia shipped 3.76 million data center GPUs in 2023, representing 98% of the market. In 2024, those shipments more than doubled to exceed 4 million units. For 2025, analysts project another 55% increase in high-end GPU shipments, with Nvidia's new Blackwell architecture accounting for over 80% of volume.
Each of these GPUs requires approximately 1,400 watts of power when fully loaded—700 watts for the GPU itself and another 700 watts for the cooling, networking, and storage infrastructure that supports it. Multiply millions of GPUs by 1,400 watts, and you begin to understand the scale of what's being built.
The trajectory Leopold Aschenbrenner outlined in his Situational Awareness report is becoming reality. By 2028, individual training clusters will cost $100 billion, and require power equivalent to a small U.S. state. By 2030, we're headed toward $1 trillion clusters requiring 100 GW of power—equivalent to more than 20% of total U.S. electricity production—for a single facility.
Commercial Real Estate Upheaval
For generations, commercial real estate investment followed predictable patterns. Office towers in city centers. Shopping malls in suburbs. Industrial parks near highways. The "visible" economy—buildings you could see, photograph, and walk into.
That world is ending.
In 2025, data center construction spending exceeded traditional office development for the first time in history. According to CBRE, data center vacancy rates have fallen to a record-low 2.8%, with pre-leasing rates on new construction hitting 90% or higher.
JLL's 2025 North America Data Center Report found that total colocation inventory reached 15.5 GW, with 7.8 GW under construction—ten times the volume of five years ago. Yet even this unprecedented construction boom cannot keep pace with demand. Vacancy rates have plunged to just 2.3%.
Newmark reports $31.5 billion in annualized spending on new data center construction—an all-time high. The development pipeline has reached nearly 50 million square feet, essentially doubling the volume from five years ago.
As Andrew Batson, JLL's Head of U.S. Data Center Research, put it: "Power has become the new real estate."
America's Maginot Moment
The United States is now making a similar mistake with AI infrastructure investment—but with more zeros.
Map of America's "Maginot Line"...
We are investing trillions of dollars into computing hardware: GPUs, servers, networking equipment, data center buildings. These investments are well-planned and executed with remarkable speed. The chip industry is booming. The construction industry is scrambling to keep up with data center demand. Private capital is flooding into many layers of the AI infrastructure stack.
But there's one critical link being neglected: the electricity to power all of it.
Consider the sector breakdown of companies investing in AI infrastructure:
Chip and Server Companies: Nvidia, AMD, Intel, SuperMicro, Dell
Big Tech Platforms: Amazon, Microsoft, Google, Meta—deploying $380+ billion in 2025
AI-Focused Companies: OpenAI, Anthropic, xAI, CoreWeave
Cloud Computing Providers: Oracle, IBM
Data Center REITs: Equinix, Digital Realty, QTS
Energy Transition Companies: NextEra, Brookfield
Bitcoin Mining Transitions: MARA, IREN, Core Scientific, Cipher
Yet according to Morgan Stanley's November 2025 analysis, the U.S. faces a 44-47 GW shortfall in data center power capacity by 2028. That gap is equivalent to 44 nuclear power plants. Or enough electricity for 33 million homes. Or about 10% of total U.S. generating capacity.
The problem isn't financial—the money exists and is being deployed. The problem is unbalanced capital deployment.
The Electricity Gap
Let me repeat the scale of the disconnect, that I detailed in a previous chapter.
U.S. electricity production has remained roughly flat since 2000, hovering around 4,000-4,250 terawatt-hours annually. The grid was built for a different era—designed to power homes, offices, and traditional manufacturing. It was never engineered for the concentrated, high-density power demands of modern AI infrastructure.
Meanwhile, AI's electricity appetite is growing exponentially. According to Deloitte, AI-specific data center power demand will grow from 4 GW in 2024 to 123 GW by 2035—a 30-fold increase in eleven years. Goldman Sachs forecasts that global power demand from data centers will rise 50% by 2027 and as much as 165% by 2030.
Consider what a single $1 trillion AI training cluster would require: 100 GW of continuous power consumption. Running for a year, that's 876 terawatt-hours—equivalent to more than 20% of total U.S. electricity production consumed by a single facility.
The Morgan Stanley analysis breaks down the gap:
Total U.S. data center power demand (2025-2028): ~65 GW
Available near-term grid access: ~12-15 GW
Data centers currently under construction: ~6 GW
Resulting shortfall: 44-47 GW
A Problem Money Cannot Solve
Here's what makes this different from other infrastructure challenges: the bottlenecks are structural, not financial.
Grid interconnection: Projects face 5-8 year backlogs just to connect to the existing grid.
Transmission infrastructure: New high-voltage lines require 5-15 years from conception to completion.
Power plant construction: Even natural gas plants require 2-3 years. Nuclear won't help until mid-2030s.
Transformer supply: Lead times of 120-210 weeks (2.3-4 years). Only 20% manufactured domestically.
Meanwhile, data centers can be built in 12-24 months.
No amount of capital can compress a 15-year permitting process into 15 months. No investment can conjure transformers that take four years to manufacture. No financial engineering can accelerate the physics of grid interconnection.
We are investing $trillions into building the walls of our AI fortress while neglecting the supply lines that would feed the soldiers inside.
The Opportunity in the Gap
For those who recognize this pattern, the opportunity is extraordinary.
Each gigawatt of data center capacity represents roughly $50-60 billion in total capital expenditure. The 44-47 GW shortfall identified by Morgan Stanley thus represents a market opportunity of $2.2-2.8 trillion—just for the gap itself, not counting the ongoing expansion beyond.
Whoever can deliver electricity to data centers faster than the traditional grid—whoever can bypass the structural bottlenecks that no amount of capital can overcome—captures a market measured in trillions of dollars.
The solution requires rethinking the problem entirely. Not incremental improvements to existing systems, but orthogonal approaches that avoid the bottlenecks altogether.
My Work on Scalable Power for AI Computing
I've spent over six years developing exactly such an approach.
The core insight is simple: if the bottleneck is the grid, don't use the grid. If the constraint is transmission infrastructure, eliminate the need for transmission. If the limitation is centralized power generation, decentralize.
I invented a modular system for producing computing power that I call Solar Computing Clusters. These are off-grid, energy-independent data centers uninhibited by the bottlenecks that plague the legacy industry. They combine renewable energy generation with high-density computing in integrated units that can be deployed rapidly, at scale, without waiting for utility connections or transmission upgrades.
I've received numerous patents on this technology. More importantly, I've built commercial-scale systems to prove the concept works.
My first commercial-scale system went into service three years ago. It has demonstrated sufficient uptime for most, if not all, modern computing workloads. The system generates its own power, manages its own storage, and operates independently of the grid constraints that have left 100 MW of Silicon Valley data centers sitting dark.
The details of this technology are the subject of upcoming chapters. You can see one of the businesses I established at www.639solar.com.
The Window
History rarely offers second chances for strategic insight.
The French military had a decade to reconsider the Maginot strategy. Intelligence reports warned of German mobile warfare capabilities. Advisors urged investment in mobile reserves and flexible defense. The warnings were ignored, the alternative investments weren't made, and when the attack came, the magnificent fortifications were irrelevant.
America's AI infrastructure is now following the same pattern—massive investment in visible infrastructure while neglecting the foundational resource that makes it all function.
The companies that recognize this imbalance—that understand electricity as the true constraint, not chips or data center buildings—will capture the defining opportunity of this technological era.
The Maginot Line still stands in France, a tourist attraction and engineering curiosity. A monument to what happens when you fortify everything except what matters.
"There are a thousand hacking at the branches of evil to one who is striking at the root.", Henry David Thoreau, from "Walden" (1854)
The choice before American AI infrastructure is the same: continue hacking at branches, or finally strike at the root.
Sources:
2025 Global Semiconductor Industry Outlook, Deloitte
The Cost Of Compute: A $7 Trillion Race To Scale Datacenters, McKinsey
Can US Infrastructure Keep Up With The AI Economy, Deloitte
2025: Year Of The Datacenter Mania, AI-Supremacy.com
Racing To The $1 Trillion Cluster, Situational Awareness
AI Power: Expanding Datacenter Capacity To Meet Growing Demand, McKinsey
The Huge Problem With The AI Revolution, 44 Nuclear Power Plants By 2028, ZeroHedge
Breaking Barriers To Datacenter Growth, BCG
2025 Datacenter Outlook, JLL
AI Is Everywhere... , Bloomberg
Meta Is Spending Billions On Nvidia AI Chips, CNBC
