Tech bubbles old and new

Is Oklahoma setting itself up to ruin farm land and waste tax dollars in pursuit of AI riches, only to end up with massive, unusable, empty buildings? John Mecke, writing at Development Corporate, sees the same dynamics at work in AI infrastructure finance that led to the dot-com bubble and telecom crash right after the turn of the millennium.

The numbers are staggering. In a single week in late 2024, Alphabet announced a $40 billion plan for AI infrastructure, while Anthropic committed $50 billion for new data centers. An unprecedented gold rush is underway to build the physical backbone of the artificial intelligence revolution. Private equity firms, infrastructure funds, and sovereign wealth pools are pouring hundreds of billions into what they believe will be the defining infrastructure investment of the decade. But as the investment mania accelerates, a critical question is being quietly asked in boardrooms across Wall Street and London: What is the exit strategy? For the private equity and infrastructure funds backing these colossal, multi-billion-dollar projects, the lack of a clear path to liquidity presents a risk that could undermine the entire boom--or worse, trigger a collapse reminiscent of the telecom crash of 2000-2001.

In other words, how are investors going to make money in the short run on investments that may take most of a decade to generate revenue? Mecke offers and elaborates on four concerns: 1. The Great Mismatch: Short-Term Money Chasing a Long-Term Game

Data center infrastructure represents a long-duration, capital-intensive play that typically requires 10-15 years to generate optimal returns. Yet the capital flooding into the sector comes predominantly from funds with much shorter investment horizons.... The problem intensifies when you consider the construction timelines. CBRE research shows that power delivery delays and electrical infrastructure shortages mean new data centers now require 3-4 years from groundbreaking to operation. Add another 2-3 years for the facility to reach stable cash flow, and you're looking at 6-7 years before an investor sees meaningful returns--consuming most of the intended hold period before the asset is even fully operational.

Mecke points to AI cloud provider CoreWeave's lower-than-expected IPO valuation, debt burden, and burn rate as a cautionary tale. 2. The "Digital Ghost Town" Risk: How Today's Cutting-Edge Tech Becomes Tomorrow's Stranded Asset Mecke recalls the massive fiber optic build-out of the 1990s, but internet traffic growth was far less than the projections that fueled half a trillion dollars of debt-leveraged investment. Tulsans will remember the resulting price collapse and corporate layoffs. Improvements in compute efficiency, technological breakthroughs on the horizon, and the shift from compute-intensive AI model training to less demanding inference operations all point to deceleration in demand for processing, space, and power. "The risk of building what amounts to digital ghost towns--billions of dollars in concrete, steel, and silicon gathering dust--is not theoretical. It's the natural consequence of building infrastructure for a technology that's evolving faster than the construction timelines themselves." RELATED: William Langdon writes that Oklahoma's AI-infrastructure strategy is centered on an obsolescent, copper-connected, GPU-centered, power- and water-hungry technology, while more efficient Tensor Processing Unit (TPU) technology is emerging:

That undermines the entire pitch behind Oklahoma's data-center subsidies: You don't need massive cheap water, cheap electricity, huge tax giveaways -- if your hardware is built on the latest technology. It also means that if a data center built today with GPU farms gets converted (or partly reused) tomorrow for TPU-based infrastructure, much of the "infrastructure footprint" -- high voltage lines, oversized cooling, oversized water delivery -- becomes wasted. A white elephant.

3. Too Big to Sell: When Scale Becomes a Liability For Mecke, this involves not the size of the facilities, but the valuation of the investment. When the original investors demand a return on investment, there are few potential buyers big enough to pay what the investors expect, and the end result may be the Big Data customers scooping up the infrastructure at bargain rates.

73% of projects under construction are already preleased, primarily to a small number of hyperscale customers (Microsoft, Google, Amazon, Meta). This concentration means the market isn't liquid--it's locked. When it's time to sell, there are no alternative buyers beyond the hyperscalers themselves, who have every incentive to wait for distressed pricing rather than pay peak valuations.

4. The Flawed Escape Routes: Why Traditional Exits Don't Work Mecke explores the possibility of IPOs and more creative financial strategies as possibilities to attract investors, but notes worrying protections for insiders that make this industry a bad deal for future investors:

Analysis of CoreWeave's IPO structure by Mostly Metrics reveals troubling details designed to protect insiders while exposing retail investors. Magnetar Capital's "Penny Warrant" allowed them to buy shares for $0.01 each--a price unavailable to public investors. Founders cashed out nearly $500 million pre-IPO, de-risking their positions while marketing the company to retail buyers at full price. This pattern--insiders reducing exposure while retail bears downside risk--is classic bubble behavior.

Mecke goes on to list several warning signs: the massive amount of capacity in the construction pipeline or in the planning stages; long waits for grid connections and rising electric costs; the possibility of debt markets deciding enough is enough; and the question of how quickly AI capacity can be monetized by attracting customers still figuring out how to integrate AI into their businesses. Each of the factors listed above were factors in the bursting of the telecom bubble. The fiber optic infrastructure "eventually found its purpose, enabling the streaming, cloud computing, and mobile revolution that followed. The fiber was there when demand finally caught up--just not in time to save the original investors."

The question isn't whether AI will transform computing--it almost certainly will. The question is whether the current infrastructure buildout is properly sized, timed, and financed to capture that value. History suggests that revolutionary technologies often create their greatest wealth in the second wave, after the first wave of investors has built too much, too fast, with too much debt. As AI infrastructure investments scale into the tens of billions, the most important question may not be who is funding it, but who will be left holding the keys when the music stops.

Which takes us back to William Langdon's Substack essay. Politicians chasing these developments are offering discounted access to water and tax incentives, all in hopes of a small number of long-term jobs. Langdon calls on policymakers to ensure that the developers are responsible for the risks, not simply reaping the rewards:

Require infrastructure costs to be borne by developers -- not taxpayers or ratepayers. If you want to build a supercomputer campus, pay for the grid upgrades, water infrastructure, environmental mitigation, and long-term maintenance.

- December 11, 2025 at 01:10PM
Tech bubbles old and new
Click the headline to read the full story.