Markets are pricing AI power demand as a permanent physical shortage — and building data centers, nuclear reactors, and grid upgrades accordingly. The telecom industry made the same assumption about internet traffic in 1999, laid 80 million miles of fiber, and left 97% of it sitting dark. The efficiency variable they missed was software. The AI buildout has the same vulnerability: small language models, quantization, and on-device processing are already cutting energy requirements by 50–90% per task.
Read the full analysis, sources, and counter-arguments ↓The Numbers the Market Is Working From
The dominant narrative in markets today treats AI as an unquenchable power consumer. The projections driving capital allocation are not subtle.
2028 projected consumption: ~580 TWh (U.S. Dept. of Energy estimates)
Implied growth: 3.3× in five years
For reference: 580 TWh is roughly equivalent to the entire annual electricity consumption of France.
These projections are driving billions of dollars into physical hardware, specialized data center real estate (REITs), and utility companies — under the assumption that we must physically rebuild the power grid to accommodate AI's growth.
power use, 2023
by 2028
in 5 years
after telecom crash
The Year 2000 "Dark Fiber" Parallel
To assess the AI infrastructure thesis, the telecom crash of 2001 is the most relevant historical data point. The structure of the error was identical.
Fiber-optic cable laid globally: ~80 million miles
Percentage of laid fiber in active use by 2002: 2.7%
Fiber sitting unused ("Dark Fiber"): 97.3%
Market cap lost across telecom sector: ~$2 trillion
The consensus in the late 1990s was that internet traffic was doubling every three to four months. Telecom giants took this as a mathematical certainty and competed to lay physical infrastructure first. The assumption driving the buildout was explicit: the only way to increase network capacity was through physical expansion.
Telecom executives assumed demand would grow linearly and that the only supply response was physical. They were correct about demand. They were catastrophically wrong about supply — because they didn't account for the efficiency of software.
The technology that collapsed the thesis was Wavelength-Division Multiplexing (WDM). WDM allowed a single strand of existing fiber to carry dozens of simultaneous signals by splitting light into different wavelengths — different colors. Overnight, one physical cable did the work of dozens. The infrastructure built for a world of linear demand became redundant before it was even fully deployed.
§ Structural ComparisonThe Pattern, Side by Side
| The Variable | Telecom, 1999–2001 | AI Infrastructure, 2024–? |
|---|---|---|
| Consensus demand assumption | Internet traffic doubling every 3–4 months forever | AI power demand tripling by 2028, growing indefinitely |
| Physical infrastructure being built | 80 million miles of fiber-optic cable | Data centers, nuclear reactors, grid upgrades |
| Capital deployed | $500B+ in debt by telecom firms | $100B+ annual capex by hyperscalers; billions more from utilities and REITs |
| The efficiency variable being ignored | Wavelength-Division Multiplexing (WDM) | Small Language Models, Quantization, On-Device AI |
| The brute-force assumption | More demand → must lay more physical cable | More demand → must build more physical power infrastructure |
How AI Software Is Outpacing Hardware
The efficiency curve in AI is real and moving fast. Three specific developments are particularly relevant to the infrastructure thesis:
1. Small Language Models (SLMs)
Current large language models are general-purpose tools applied to specific tasks — an architectural mismatch. Research from University College London found that replacing a general-purpose LLM with a task-specific small model cuts energy consumption by up to 90% on comparable tasks. For translation specifically, SLMs use approximately 35 times less energy. The industry pivot toward specialized models is already underway.
Using a massive general AI to answer a narrow question is like driving a semi-truck to pick up a single grocery bag. The industry is building the smaller vehicle. The infrastructure was priced for the truck.
2. Quantization
AI models perform billions of arithmetic operations using floating-point numbers with many decimal places. "Quantization" reduces the precision of these calculations — using fewer decimal places — in exchange for lower compute and energy requirements. Published research shows quantization can cut model energy usage by up to 50% while maintaining 97% of output accuracy. This is a software-side change that costs nothing to deploy against existing infrastructure.
3. On-Device Processing
The infrastructure buildout assumes all AI inference (the process of running a trained model to generate output) must happen in centralized cloud data centers. On-device AI — running models directly on phones, laptops, and industrial sensors — eliminates the data transmission requirement entirely. Per-task energy consumption drops by orders of magnitude. Apple's on-device intelligence framework, Google's Gemini Nano, and Qualcomm's AI-optimized chips are all production deployments of this approach, not research projects.
Efficiency improvements don't necessarily reduce total power consumption — they can increase it via Jevons Paradox, where cheaper AI encourages more usage, expanding total demand faster than efficiency gains can offset. The telecom parallel also has limits: fiber was a fixed pipe; data centers can be repurposed or decommissioned more flexibly. And sovereign AI buildouts (India, the Gulf states, the EU) may create demand the U.S. efficiency curve doesn't capture. The thesis isn't that AI power demand won't grow — it's that the market is pricing permanent hypergrowth and not pricing the efficiency response at all.
If AI inference costs stop declining — or reverse — the efficiency argument weakens significantly. If no major frontier lab deploys quantized or on-device models at scale by 2026, the timeline for the efficiency correction extends. If power demand numbers in 2025 significantly exceed DOE projections rather than tracking below them, the brute-force thesis gets stronger. We will update this analysis as those data points arrive.
The "Dark Data Center" Risk
The telecom crash didn't happen because internet demand failed to materialize. Demand was real. The crash happened because companies built infrastructure for a world where the only supply response was physical — and missed the software efficiency response entirely. By the time WDM was commercialized, the infrastructure was already in the ground and the debt was already on the balance sheet.
The current market is pricing utility stocks, data center REITs, and hardware manufacturers as if AI power consumption will compound indefinitely and the only solution is more physical infrastructure. That pricing does not appear to account for the possibility that a software update — or a shift in architecture — could make current-generation infrastructure overcapacity overnight.
The lesson from 2000 is not that infrastructure manias always end in collapse. It is that the market consistently underprices efficiency and overprices brute force during the early phase of a technology wave. Whether the timing of the correction is 2026 or 2030 is unknowable. That the correction arrives is the historical base rate.
Primary Sources
- All data points cited inline with source references. This article uses publicly available financial data, company disclosures, and historical market records.