Is this a data center bubble, not an AI bubble?

We're building infrastructure for demand that may never materialize

While analysts debate whether we're in an "AI bubble," a more uncomfortable question emerges: Are we actually in a data center bubble? The physical infrastructure boom may be racing ahead of actual demand in ways that echo the dot-com era's fiber optic excess.

The numbers are staggering

Global data center capital expenditure surged 51% to $455 billion in 2024, with the top hyperscalers Amazon, Microsoft, Google, and Meta, investing nearly $200 billion alone. McKinsey projects U.S. data center power demand will more than triple from approximately 25 gigawatts in 2024 to over 80 gigawatts by 2030, requiring massive investments in power infrastructure.

This is investment at a scale that rivals entire economic sectors. But is the demand there to justify it?

The utilization problem

Despite this flood of capital into physical infrastructure, actual utilization rates tell a troubling story. Cloud computing studies analyzing 4,000 Kubernetes clusters reveal that companies utilize only around 13% of provisioned CPUs and 20% of memory on average. Even more revealing: according to "The State of AI Infrastructure at Scale 2024," only 7% of organizations report GPU utilization above 85% at peak load, indicating widespread underutilization of the most expensive components.

Lawrence Berkeley National Laboratory noted in 2024 that "very few companies report actual data center electricity use and virtually none report it in context of IT characteristics." The lack of transparency suggests utilization may be even worse than we know.

The adoption reality check

ChatGPT has grown dramatically to approximately 800 million weekly active users processing over 2.5 billion daily prompts as of late 2025. Yet Pew Research found that only 34% of U.S. adults have ever used ChatGPT as of early 2025, with usage skewing heavily toward younger demographics, 58% of adults under 30 have tried it versus just 10% of those 65 and older.

More importantly, the nature of usage matters: approximately 78% of ChatGPT conversations focus on "everyday tasks" like seeking information, drafting emails, and practical guidance, not compute-intensive operations. About 70% of consumer queries are non-work-related. The median use case is asking ChatGPT to explain something or edit text, not running sophisticated AI models that justify massive infrastructure buildouts.

On the enterprise side, while surveys show widespread AI "adoption," implementation remains limited. Most companies are experimenting rather than deploying at scale, with the majority of AI initiatives still in pilot phases rather than production workloads that would require sustained high utilization of expensive infrastructure.

We're building Ferrari racetracks when most traffic will be Honda Civics.

Market warning signs

Several indicators suggest overcapacity is emerging:

Collapsing GPU rental prices: The GPU rental market has seen dramatic price declines throughout 2024 and 2025. The H100 rental index dropped 23% from September 2024 to June 2025, falling from $3.06 to $2.36 per hour. Individual providers now offer H100 rentals as low as $1.49-$2.85 per hour, compared to the $4+ per hour rates that were common in 2023. Analysis by industry experts indicates that once H100 rental rates fall below $1.65 per hour, revenues no longer recoup the investment, and prices need to be above $2.85 to beat stock market returns.

SemiAnalysis, which accurately predicted this decline in December 2023, now characterizes it as "a buyers' market for GPU rentals" with "widespread availability from over 100+ AI Neoclouds and Hyperscalers." The firm notes that H100 production ramped up throughout 2024, with the decline continuing as buyers pivot to focus on next-generation Blackwell chips.

Speculative building: Data center vacancy rates hit record lows of 1.9% in primary markets at year-end 2024, but nearly 80% of the 6,350 MW under construction was already pre-leased. This isn't organic demand, it's fear-driven speculation. As Ares Management Co-President Kipp deVeer noted, "These trends tend to lead to overbuilds in certain places."

The efficiency wild card: Chinese firm DeepSeek released its V3 model in December 2024, claiming approximately 18x reduction in training costs and 36x reduction in inference costs compared to GPT-4o. While debates continue about these specific claims, the broader point stands: algorithmic improvements could dramatically reduce infrastructure requirements faster than we're building capacity.

The dot-com parallel

The comparison to the late-1990s telecom bubble is unavoidable. Telecommunications companies laid millions of miles of fiber optic cable, convinced bandwidth demand would grow exponentially. They were right about the direction but catastrophically wrong about the timing and required capacity.

The infrastructure eventually found use, but not before triggering massive bankruptcies and wiping out hundreds of billions in shareholder value. The capacity was ahead of demand by years, perhaps decades.

Today's data center boom shares unsettling similarities: massive capital deployment driven by competitive fear, assumptions about exponential demand growth, long lead times between investment and utilization, and everyone building simultaneously.

What a burst looks like

A data center bubble burst wouldn't mean the infrastructure becomes worthless, just as fiber optic cable wasn't worthless in 2001. But it could manifest as:

Stranded assets in secondary markets without adequate power or proximity to demand. Chinese industry analysis notes that over 13,000 intelligent computing centers have emerged, many built on a "construct first, find applications later" model that has led to oversupply.

Valuation corrections for data center REITs and operators, particularly newer entrants heavily leveraged to AI growth narratives with limited revenue diversification.

Consolidation among smaller GPU-as-a-service providers facing margin compression. Industry analysis indicates many providers are "barely breaking even" at current rental rates below $2.85 per hour.

Accelerated depreciation: Hyperscalers currently depreciate data center assets over 5-15 years, but NVIDIA's product roadmap tells a different story: Ampere (2020), Hopper (2022), Blackwell (2024), Rubin (expected 2026). If actual useful life of cutting-edge GPUs is closer to two years than ten years due to rapid obsolescence, companies may face write-downs as they recognize infrastructure won't remain competitive for its assumed lifespan.

The infrastructure-demand mismatch

Academic research on LLM workloads shows they often leave substantial power headroom, allowing for oversubscription of power to servers, suggesting current capacity may exceed actual requirements. The shift from training to inference workloads further complicates matters: inference requires very different infrastructure with lower power density and better economics on older, cheaper GPUs.

Industry observers in China note: "We have yet to see a 'killer app' for AI or a clear application scenario." Given oversupply of computing power and fewer immediate applications than anticipated, further price declines appear likely.

The real question

The central issue isn't whether AI is valuable or will grow, it certainly will. The question is whether the infrastructure being built today matches near-term demand or whether we're constructing capacity for applications that remain years away.

Infrastructure investment should match expected utilization with a reasonable margin for growth. Instead, we're seeing capacity expansion that assumes everything works perfectly: models scale predictably, applications proliferate rapidly, efficiency improvements are fully offset by demand growth, and power constraints get resolved on schedule.

History suggests such perfect scenarios rarely unfold.

Irrational exuberance in concrete and silicon

The data center boom of 2024-2025 exhibits classic bubble characteristics: exponential growth in investment, assumptions about sustained exponential demand, everyone building simultaneously, and a narrative so compelling that skepticism seems like Luddism.

The painful truth may be that we're not in an AI bubble so much as a data center bubble. The AI applications will eventually emerge. The question is whether they arrive in time to fill the infrastructure we're constructing today, or whether we're building monuments to investor exuberance that will sit underutilized for years.

The dot-com era taught us that being right about the destination doesn't protect you from being wrong about the timing. Fiber optic cables buried in 1999 didn't prevent the telecom crash of 2001. Today's hyperscale data centers may tell a similar story, technically impressive, potentially useful someday, but fundamentally ahead of their time and larger than immediate demand justifies.

That's not a technology failure. It's a capital allocation failure. And those tend to be expensive lessons.