Investment Bubbles are Virtually Certain for Important New Technologies Such as AI

Concern about whether artificial intelligence is now, or is becoming, an “investment bubble” is widespread. Some students of technology history might argue such overinvestment is virtually inevitable. 

Economists have observed that each major technological revolution, including railroads, electricity, automobiles, the internet, and AI, has followed a pattern of initial optimism, speculative excess, and subsequent consolidation, after which the infrastructure built during the boom enables sustainable long-term growth.

Technological Era	Approx. Timeline	Peak Investment Behavior	Bust Phase	Long-Term Outcome
Canal & Railway Mania	1830s–1850s	Rapid railway and canal construction driven by speculative capital from Britain and U.S. investors	Oversupply and bankruptcies in mid‑19th century Optimal learning and new technology bubbles - ScienceDirect	Created national transport infrastructure enabling commerce and industrial expansion
Electrification & Utilities	1880s–1910s	Massive capital outlays into competing electric utilities, equipment, and distribution grids	Consolidations and failures during utility overcapacity crises Federal Reserve	Universal household and industrial electrification
Automobile & Radio Era	1910s–1930s	Dozens of automakers and radio startups seek to dominate markets; speculative IPOs	Great Depression collapse wiped out smaller firms The Bubble That Knows It's a Bubble	Durable automotive and broadcast industries emerge
Dot‑Com Internet Wave	1995–2001	Extreme venture capital inflows and equity valuations for unprofitable internet firms	2000–2002 crash destroyed trillions in market cap Optimal learning and new technology bubbles - ScienceDirect	Fiber optics, e‑commerce, and data centers laid the base for Web 2.0
Clean Tech & Solar Boom	2006–2011	Subsidy‑driven boom in renewable startups and manufacturing overcapacity	Collapse of many firms after subsidy reductions Optimal learning and new technology bubbles - ScienceDirect	Cost per watt of solar fell dramatically, enabling today’s viability
AI & Generative Models	2020s–present	Trillions in compute build‑out, IPOs, speculative valuations exceeding dot‑com scale	Potential retrenchment ahead if returns lag costs The Bubble That Knows It's a Bubble	Neural infrastructure and models likely become core of enterprise computing

So, yes, overinvestment, speculation and consolidation are virtually inevitable. But we might also suggest that sustainable growth also is virtually inevitable. 

Industry	Primary AI Use Cases	2025 Adoption / Spending Trends	Key Outcomes
IT & Telecommunications	Network optimization, predictive maintenance, customer support chatbots, AI-driven service provisioning	38% adoption rate; projected to add $4.7 trillion in value by 2035 AI Adoption Statistics in 2025	Reduced downtime, improved customer satisfaction, adaptive network efficiency
Finance (Banking, Insurance, Investment)	Fraud detection, credit scoring, algorithmic trading, customer risk profiling, robo-advisors	Over$20 billion global AI spending;68% of hedge funds use AI in trading AI Adoption Statistics in 2025	Faster fraud detection, increased trading precision, cost reduction in customer service
Healthcare & Life Sciences	Diagnostics, patient triage, drug discovery, personalized treatment, text summarization for records	36.8% CAGR in adoption;42% of hospitals using AI chatbots AI Adoption Statistics in 2025	Improved diagnostic accuracy, faster patient response, reduced care costs
Manufacturing & Industrial	Predictive maintenance, process automation, supply chain optimization, quality control	77% of producers using AI, up 7% YoY AI Adoption Statistics in 2025	23% downtime reduction, improved supply chain resilience
Retail & Consumer Goods	Demand forecasting, personalization, inventory optimization, chatbot-driven sales	20% of tech budget allocated to AI, +5% YoY NetGuru	15% conversion lift, 18% overstock reduction
Software & Cloud Services (Big Tech)	Generative AI systems, agentic AI, infrastructure automation	$320 billion capex in 2025 by major cloud companies (Microsoft, Amazon, Google, Meta) Ropes Gray	Faster LLM integration, platform differentiation, scalability gains
Robotics & Automation	Embodied AI for general-purpose robots, autonomous agents, fleet control software	342% YoY increase in deal value Ropes Gray	Accelerated labor automation, robotics ecosystem expansion
Energy & Utilities	Grid optimization, predictive maintenance for equipment, carbon monitoring	Fast-growing adoption Forbes (2025)	Efficiency gains, lower operational emissions, reduced outages

As Usual, Video Drives the Intensity of Model Processing, as it does Networking

Lumen Technologies says announced data center projects represent an additional one billion square feet of incremental U.S. data center capacity added by 2030. 

source: Lumen Technologies 

And we might already predict that artificial intelligence operations, including generation of video, are one reason for the expansion. 

Consider the use of a large language model to create an AI video that is 10 seconds long, at 240 frames per second. That requires the generation of 2400 frames (images). 

Compared to the processing required to answer a text query, that 240 fps video takes 4286 times the processing.

Task	Frame Rate (FPS)	Frames (for 10s)	Estimated TFLOPs	Ratio vs. Text
Text Query	N/A	N/A	28	1x
Video Generation	30	300	15,000	~536x
Video Generation	60	600	30,000	~1,071x
Video Generation	120	1,200	60,000	~2,143x
Video Generation	240	2,400	120,000	~4,286x

AI Training Requires a New and Faster Transport Network, Lumen Argues

Artificial intelligence training now can require exabytescale data transfers, requiring a new transport architecture and much higher capacities, a Lumen Technologies white paper notes. 

“If you go in to rent or start an AI workload and you have a 10-gig link, and a training is a petabyte, you are paying for 222 hours of idle time just to move your data into that cloud,” says David Ward, Lumen Technologies chief technology and product officer.

Consider the implications of capacity when  running an AI training session. If data delivery (transport) costs about $0.02 per gigabyte, then using a 10-gigabyte-per-second connection will cost $431,000 when loading training data into a model on a remote basis.

Over a 400-gig connection costs are about 40 times less, Ward argues.

So, at petabyte and exabyte scale, the transport network, not the graphics processing units, set AI model training time and cost, Lumen argues. 

For example, to egress an exabyte of data over a 10-Gbps connection would take 1,389 hours, Lumen argues.  When conducted over a 400-Mbps or 800-Gbps connection, the time decreases to 694 and 347 hours, respectively. 

The implication is that bandwidth now matters in a new way for optimizing the use of graphics processing units used for training operations. 

 source: Lumen Technologies