Nuclear-Powered AI: Thesis Validation (Part 1)

To start validating my thesis I wrote in my initial post, I broke it down into it’s foundational elements, the first one being “Will AI actually create a significant energy bottleneck?”.

If you haven’t read my first post, the thesis is basically this:

My (4-part) Working Thesis

As AI becomes increasingly compute-intensive, giants like NVIDIA, Amazon, Google and others will step up to solve the compute bottleneck.

The solution for the compute bottleneck will create another critical constraint: energy.

Energy will experience incremental improvements around current infrastructure (e.g. transmission, distribution, storage, etc), but the next step-change innovation in energy will be nuclear.

Finally, AI (and other) data centers will serve as a beachhead market for nuclear, given the “product” is self-contained, modular, and “generator-like”.

So, I had Perplexity develop a deep-dive report, sourcing from 180+ different citations. In nutshell, data suggests the answer is an empathic yes! More specifically:

The quantitative evidence strongly suggests that AI-driven compute growth is creating unprecedented energy demands. With projections indicating data centers could consume between 4.5-12% of global electricity by 2030 (depending on the forecast), and AI workloads specifically growing at 40-45% CAGR, there are clear signs of an emerging bottleneck.

Next, my plan is to reach out to some actual humans to double click on some of these stats. In the meantime, here’s the (pretty good imo) report that Perplexity generated 👇

Perplexity-Generated Report

Before exploring whether AI will create a significant energy bottleneck, this research report synthesizes quantitative data from leading industry analysts and major technology companies to evaluate projected energy demands and growth rates for AI infrastructure.

Global Data Center Energy Demand Projections

Industry Analyst Forecasts

Industry analysts unanimously project substantial increases in data center energy consumption, with AI workloads driving the most dramatic growth:

IDC Projections

Global datacenter electricity consumption will more than double between 2023-2028 with a 19.5% CAGR, reaching 857 TWh by 20281
AI datacenter energy consumption specifically is growing at an accelerated 44.7% CAGR, projected to reach 146.2 TWh by 20271
Electricity represents the largest operational expense for datacenter operators, accounting for 46% of total spending for enterprise datacenters and 60% for service providers1

Gartner Projections

Power required for data centers to run AI-optimized servers will reach 500 TWh per year in 2027, which is 2.6 times the level in 20233
40% of existing AI data centers will be operationally constrained by power availability by 20272 3
Reliable 24/7 power can only be generated by hydroelectric, fossil fuel, or nuclear power plants, limiting options for sustainable growth3

McKinsey Projections

US data center power consumption will grow from 3-4% of total US power demand today to 11-12% by 20304
This expansion will require infrastructure investments exceeding $500 billion4
European data center power consumption will surge from 62 TWh to over 150 TWh by 20305
European IT load demand will grow from 10 gigawatts in 2023 to approximately 35 GW by 2030, representing a CAGR of about 20%5

BloombergNEF Projections

Global data center demand will triple to approximately 1,500 terawatt-hours by 2030, growing from around 1.5% to 4.5% of total global consumption8
Bloomberg Intelligence calculates that AI-driven data center electricity use could surge 4-10x by 20307
Boston Consulting Group expects US data center power demand to triple to 7.5% of total US power by 20308
Goldman Sachs projects this figure will reach 8%, requiring an additional 47GW of dispatchable supply8

Cross-Analyst Comparison

There is significant variance in forecasts, highlighting the uncertainty in this rapidly evolving field:

McKinsey (most bullish): Global data center demand might grow by 240GW between 2023-2030, with US data centers potentially consuming up to 12% of total US power by 20308
IEA (most conservative): Global power demand will increase by only 35-100% by 2030, adding between 120-390 TWh8
FERC: US data center load will grow by no more than two-thirds by 20308

Tech Giants’ Energy Consumption and Strategies

Amazon

Achieved 100% renewable energy for all operations in 2023, seven years ahead of their 2030 goal11
Scope 1 emissions increased to 14.27 million metric tons CO2e in 202311
Invested in 513 renewable energy projects totaling 28GW capacity11
Has not announced plans to move to a 24/7 matching model for renewable energy11

Microsoft

Consumed 23,564,161 megawatt-hours of renewable energy in fiscal year 202310
Emissions have increased almost 30% since announcing carbon negative goals in 2020, primarily due to AI data center infrastructure10
Plans to invest $80 billion in data center infrastructure in fiscal year 202510
Has signed power purchase agreements adding 34 gigawatts of “carbon-free” energy to the grid10

Google

GHG emissions increased 13% year-over-year in 2023, reaching 14.3 million metric tonnes CO2e14
Emissions have grown 48% since 201914
Data center consumption exceeded 24 TWh in 2023, representing approximately 7-10% of global data center electricity consumption14
Implementing various strategies including TPU efficiency improvements (3x better carbon-efficiency) while achieving 64% carbon-free energy globally14

AI Computing Growth Trends

Exponential Growth in AI Compute Requirements

Computational power for AI training has been doubling approximately every six months since 201016
Model parameters have doubled yearly, with some modern models reaching trillions of parameters16
Training data has doubled every nine to ten months since 201016
Current large models require hundreds of days of computation on specialized hardware, with training costs reaching $40 million16

AWS CEO Projection

Future language model training (2-3 generations from now) may require between one to five gigawatts of power per model, equivalent to powering a large city17

RAND Corporation Analysis

AI data centers could need ten gigawatts of additional power capacity in 2025 alone15
If current trends continue, AI data centers will require 68 GW in total by 2027 — almost equivalent to California’s 2022 total power capacity of 86 GW15
Training could demand up to 1 GW in a single location by 2028 and 8 GW by 203015

AI Server Market Growth Projections

The AI infrastructure market is experiencing explosive growth across all segments:

Precedence Research: AI server market will grow from $30.74 billion in 2024 to $352.28 billion by 2034, at a CAGR of 27.62%18
TrendForce: AI server market will reach $298 billion in 2025, representing over 70% of total server industry value19
The broader AI infrastructure market is expected to grow from $56.98 billion in 2024 to $223.85 billion in 2029 at a CAGR of 31.9%20

Power Efficiency Trends

While efficiency is improving, absolute power consumption is increasing significantly:

Per-computation efficiency improvements include 3nm/2nm nodes reducing power by 30% and 3D stacking cutting power by 50%21
However, absolute power consumption continues to increase across generations:
- NVIDIA’s H100 uses 75% more power than A100 while being 2X more efficient per operation22 23
- AMD’s MI300x consumes 50% more than MI25023
- Intel’s Gaudi 3 uses 50% more than Gaudi 2, with upcoming Falcon Shores expected to reach 1,500W per chip23

Regional Power Grid Constraints

The power challenges are already materializing in key markets:

Northern Virginia: Power for new facilities delayed until 2026, with data centers consuming approximately 20% of the region’s electrical power and growing at 25% per year6 25
Santa Clara: Substation energization delayed until 2028-2029 due to grid capacity constraints25
Ireland: Data center power consumption reached 21% of the country’s total in 2023, up from 5% in 2015, prompting a moratorium on new developments until 20288
Dominion Energy (Virginia) had to pause new connections for several months in 20228
Markets with better power availability, such as Atlanta, are benefiting from these constraints, with over 1 GW of new planned capacity25

Conclusion

The quantitative evidence strongly suggests that AI-driven compute growth is creating unprecedented energy demands. With projections indicating data centers could consume between 4.5-12% of global electricity by 2030 (depending on the forecast), and AI workloads specifically growing at 40-45% CAGR, there are clear signs of an emerging bottleneck.

Major technology companies are already experiencing emissions growth despite aggressive renewable energy investments. Regional grid constraints are creating immediate bottlenecks in prime data center markets, with many locations unable to accommodate new facilities until 2026-2029.

The mismatch between 2-3 year data center development cycles and 4-8 year grid infrastructure upgrade timelines suggests that without significant intervention, energy constraints will likely become a limiting factor for AI development within the next 5-10 years, particularly in regions with already-strained electrical grids.

Citations:

Answer from Perplexity: https://www.perplexity.ai/search/i-m-trying-to-validate-a-thesi-lXuSB41rTui8fyXgMiRBJQ?utm_source=copy_output

Abhi Dewan