The One Number That Separates AI From the Dotcom Crash

The AI boom differs from the dotcom crash because of infrastructure utilization rates. Dotcom era fiber optic cables sat 85-95% unused, while today’s AI infrastructure operates at 80% capacity. The real constraint isn’t demand but electricity generation capacity.

Video – AI Infrastructure Utilization Insights

• Dotcom bubble: 85-95% of fiber infrastructure went unused (“dark fiber”)
• AI infrastructure: 80% utilization rate with GPU clusters running 24/7
• Current bottleneck: Electricity capacity, not demand
• AI data centers need 10 GW additional power in 2025, growing to 68 GW by 2027
• Infrastructure utilization rate is the leading indicator separating real demand from speculation

Is the AI boom a bubble?

You want to know if the AI boom is a bubble. Everyone does.

The comparison to the dotcom crash writes itself. Massive infrastructure spending. Valuations disconnected from revenue. Companies racing to build capacity before demand materializes.

But there is one number worth watching.

How does AI infrastructure utilization compare to the dotcom era?

Dotcom bubble: 90% dark fiber

When the dotcom bubble burst, 85% to 95% of fiber optic cables laid in the 1990s sat unused. The industry called it “dark fiber.”

Companies laid more than 80 million miles of cable driven by WorldCom’s wildly inflated claim about internet traffic doubling every 100 days.

The infrastructure existed. Nobody used it.

AI boom: 80% utilization rate

Today’s AI infrastructure operates at an 80% utilization rate. Data centers with GPU clusters run 24/7 for multi-node deep learning and high-performance computing workloads.

Nvidia’s H100s and H200s are sold out. GPU allocations are scheduled a year ahead of time.

The infrastructure exists. Everyone is using it.

Bottom line: Infrastructure utilization separates speculation from genuine demand. Dotcom infrastructure sat idle because demand didn’t materialize. AI infrastructure runs at near capacity because demand is real.

What is the real bottleneck for AI infrastructure?

Dotcom bottleneck: Last mile connectivity

The dotcom era had a last mile problem. You could lay all the fiber you wanted, but dial-up modems choked the connection to homes.

Infrastructure sat idle because the bottleneck lived elsewhere.

AI bottleneck: Electricity generation capacity

AI has a different bottleneck. Electricity.

In Texas alone, electricity loads of tens of gigawatts have been requested for AI data centers. Only a little over a gigawatt has received approval.

Nvidia CEO Jensen Huang reportedly pointed to power constraints as one reason “China is going to win the AI race.”

How much power do AI data centers need?

Globally, AI data centers could need 10 gigawatts of additional power capacity in 2025. More than Utah’s total power capacity.

If exponential growth continues, they will need 68 GW by 2027. This almost doubles global data center power requirements from 2022.

Training could demand up to 1 GW in a single location by 2028 and 8 GW by 2030. Equivalent to eight nuclear reactors.

Why is electricity the constraint?

72% of survey respondents consider power and grid capacity to be extremely challenging for data center infrastructure buildout.

Transmission needed to bring renewable capacity to load takes over a decade to build. Gas power plant projects without contracted equipment won’t become available until the 2030s.

Core insight: The constraint shifted from last mile connectivity in the dotcom era to electricity generation in the AI era. Power availability now determines infrastructure expansion, not technical capability.

Does efficiency reduce infrastructure demand?

Training cost comparison

DeepSeek claims to have trained its model for $6 million using 2,000 Nvidia H800 GPUs.

In contrast, GPT-4 cost $80 million to $100 million and required 16,000 H100 GPUs. Meta’s LLaMA 3 had similar requirements.

You might think cheaper training means less infrastructure demand. The opposite happens.

Why efficiency increases demand: Jevons Paradox

This is Jevons Paradox. When a resource becomes more efficient, its overall consumption tends to soar.

As the cost of AI reasoning dropped by 90%, the total demand for AI services exploded. More efficient training techniques mean more projects enter the market simultaneously.

By late 2025, NVIDIA recovered to a historic $5 trillion market cap. Efficiency didn’t reduce demand. It unleashed it.

Key mechanism: Efficiency gains don’t reduce total infrastructure demand because lower costs enable more projects. This creates net expansion rather than consolidation.

How much is being spent on AI infrastructure?

2025 capital expenditure

America’s hyperscalers have pledged to spend a record $320 billion on capital expenditures in 2025 alone. A 40% jump from last year’s record-setting $230 billion.

Five-year spending projections

Nvidia management expects to benefit from $3-4 trillion in AI infrastructure spending over the next five years.

The top four cloud service providers had a capital expense budget of $300 billion two years ago. This goes up to $600 billion.

Nvidia’s market capture

For every $50 billion companies spend on AI infrastructure, Nvidia gets $35 billion.

The scale is unprecedented. The question is whether this represents genuine demand or competitive pressure driving overbuilding.

Financial reality: AI infrastructure spending reached unprecedented scale with $320 billion committed in 2025 alone. The open question is whether competitive dynamics are driving overbuilding beyond what actual demand justifies.

What does infrastructure utilization tell us about bubble risk?

Utilization as a leading indicator

Infrastructure utilization is the leading indicator. The dotcom era built capacity that sat dark. The AI era builds capacity that runs at full throttle from day one.

This doesn’t mean there is no bubble risk. Competitive dynamics drive infrastructure investment beyond what demand signals alone would justify.

Companies build because their competitors are building. Nobody wants to be caught without capacity when the market moves.

The fundamental difference

But the fundamental difference stands. The infrastructure is being used.

The constraint isn’t demand. The constraint is electricity generation capacity.

Power as strategic constraint

Power is no longer one chapter of the energy transition. It has become a strategic constraint on nationwide economic growth.

AI-driven data center loads are arriving fast and in clusters and require strict reliability.

This new race for “power for compute” resembles prior historic industrial turning points. The 19th century buildout of railroads. Early 20th century mass electrification. The telecom network rollout.

Potential solution: Grid flexibility

A 2025 Duke University study found if data centers are flexible for 0.25% of their operating time, representing 22 hours per year, the U.S. grid accommodates 76GW of new data center load without building new power plants.

Critical distinction: Infrastructure utilization rates reveal whether spending reflects genuine demand or speculative overbuilding. The 80% utilization rate indicates real demand, making electricity capacity the primary constraint rather than market speculation.

Frequently Asked Questions

Is the AI boom a bubble like the dotcom crash?

The AI boom differs from the dotcom crash in infrastructure utilization. Dotcom era fiber optic cables sat 85-95% unused, while AI infrastructure operates at 80% capacity. The infrastructure is being used, not sitting idle. The real constraint is electricity generation capacity, not lack of demand.

What was the main problem with dotcom infrastructure?

The dotcom era suffered from massive overbuilding. Companies laid over 80 million miles of fiber optic cable, but 85-95% went unused because demand didn’t materialize as projected. The last mile problem meant dial-up modems choked connections to homes, so backbone infrastructure sat idle.

Why is electricity the bottleneck for AI?

AI data centers need massive amounts of power. Globally, they require 10 gigawatts of additional capacity in 2025, growing to 68 GW by 2027. Training a single model could demand 1 GW by 2028 and 8 GW by 2030, equivalent to eight nuclear reactors. Transmission infrastructure takes over a decade to build, creating immediate constraints.

Does more efficient AI training reduce infrastructure demand?

No. This demonstrates Jevons Paradox. When AI training becomes more efficient, overall consumption increases because lower costs enable more projects. DeepSeek trained a model for $6 million versus GPT-4’s $80-100 million cost, but this efficiency caused total demand to explode rather than contract.

How much are companies spending on AI infrastructure?

America’s hyperscalers committed $320 billion in capital expenditures for 2025 alone, a 40% jump from the previous year’s $230 billion. Nvidia expects $3-4 trillion in AI infrastructure spending over five years. Cloud providers doubled their capital expense budgets from $300 billion to $600 billion.

What is the 80% utilization rate?

The 80% utilization rate means AI data centers with GPU clusters run at 80% capacity, operating 24/7 for deep learning and high-performance computing. Nvidia’s H100s and H200s are sold out with GPU allocations scheduled a year ahead. This indicates genuine demand rather than speculative overbuilding.

Could competitive pressure be driving overbuilding?

Yes, competitive dynamics drive infrastructure investment beyond what demand signals alone would justify. Companies build because their competitors are building. Nobody wants to be caught without capacity when the market moves. The question is whether this represents genuine demand or creates dotcom-style overcapacity.

How does grid flexibility help with power constraints?

A 2025 Duke University study found if data centers are flexible for 0.25% of their operating time (22 hours per year), the U.S. grid accommodates 76GW of new data center load without building new power plants. This flexibility allows existing infrastructure to support expanded AI operations.

Key Takeaways

• Infrastructure utilization separates real demand from speculation: 80% AI utilization versus 90% dotcom dark fiber
• The constraint shifted from last mile connectivity to electricity generation capacity
• Efficiency gains increase total demand through Jevons Paradox rather than reducing infrastructure needs
• AI infrastructure spending reached $320 billion in 2025, with projections of $3-4 trillion over five years
• Power requirements could reach 68 GW by 2027, almost doubling global data center capacity from 2022
• Competitive dynamics may drive overbuilding, but high utilization rates indicate genuine demand
• Grid flexibility solutions could accommodate 76GW of additional load without new power plants