Photorealistic image of a modern hyperscale AI data center with cooling systems and industrial vents, representing the emerging AI infrastructure bubble.

The AI Infrastructure Bubble: 4 Surprising Reasons the $90 Billion Data Center Boom Could End in a Bust

November 23, 2025 - By admin

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.

The $2 Trillion Question Nobody Wants to Answer

The AI data center market is experiencing what AlixPartners researchers describe as “dramatic transformation”—a battleground where speed eclipses efficiency and long-held investment certainties are being upended by technological disruption. According to their recent survey of over 400 senior data center executives, 70% expect M&A activity to become more attractive within the next year, yet two-thirds of REIT and construction players anticipate potential distress ahead.

The disconnect reveals a fundamental problem: investors are rushing into assets they may not be able to exit profitably, or at all.

1. The Great Mismatch: Short-Term Money Chasing a Long-Term Game

The first surprising reason for concern is a fundamental conflict between investor expectations and the nature of the asset itself. 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.

According to industry analysis, the majority of private equity investors expect to exit their data center investments within a 5 to 10-year timeframe. While that figure approaches 21% for massive hyperscale facilities at the heart of the AI boom, it still means nearly 80% of investors plan to exit before these assets reach their natural maturity.

This temporal mismatch creates what Tom Mannion of BDO Global calls an “unanswered question” about sustainable returns. If an exit isn’t viable within the fund’s lifecycle, general partners and their limited partners could face what AlixPartners terms a “prolonged dearth of distributions”—or suffer substantial losses if forced to hold indefinitely.

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.

The CoreWeave Cautionary Tale: The pressures of this mismatch played out in real-time with CoreWeave’s March 2025 IPO. The AI cloud provider, heavily backed by Blackstone and Magnetar with over $12.9 billion in debt accumulated in just two years, initially targeted a $35 billion valuation but was forced to slash expectations to $23 billion. Shares priced at $40—below the expected $47-$55 range—and closed flat on the first day, signaling investor skepticism about the company’s debt burden and path to profitability.

CoreWeave’s financial profile reveals the challenge: despite explosive revenue growth from $16 million in 2022 to $1.9 billion in 2024, the company reported negative Core Earnings of $36 million in 2024 and burned through cash at a rate that could only sustain operations for approximately 9 months as of December 2024, according to New Constructs analysis. With $24.5 billion in total debt (including off-balance-sheet operating leases), CoreWeave faces $7.5 billion in interest payments through the end of 2026.

Perhaps most troubling: 62% of CoreWeave’s revenue comes from a single customer—Microsoft—creating massive concentration risk that terrifies public market investors looking for sustainable, diversified cash flows.

2. The “Digital Ghost Town” Risk: How Today’s Cutting-Edge Tech Becomes Tomorrow’s Stranded Asset

The second major risk factor is technological obsolescence—and it’s more immediate than most investors realize. The AI data center industry faces a complex set of risks that could render today’s state-of-the-art facilities obsolete within years, not decades.

The Telecom Bubble Parallel

History provides a sobering precedent. During the late 1990s, telecommunications companies invested more than $500 billion in fiber optic infrastructure, financed largely with debt, based on projections that internet traffic would grow 1,000% annually. The reality? Traffic grew at “only” 100% per year—still impressive, but nowhere near sufficient to justify the massive buildout.

By 2001, an estimated 95% of installed fiber optic cable remained “dark”—unused and unlit. The overcapacity collapsed pricing by more than 90%, turning what had been premium infrastructure into commoditized capacity worth pennies on the dollar. Between 2000 and 2002, global telecom stocks lost more than $2 trillion in market value. WorldCom, Global Crossing, and numerous other giants filed for bankruptcy, leaving investors with massive losses.

The same dynamics now threaten AI data centers. Several technological shifts could trigger overcapacity:

1. Efficiency Breakthroughs: AI model training is already becoming dramatically more efficient. What required 100 GPUs last year might need only 10 next year. If this trajectory continues, the massive computing power these data centers are built to provide could become redundant far faster than their 15-20 year design life suggests.

2. The Training-to-Inference Shift: The industry is transitioning from the compute-intensive “model training” phase to a less demanding “inference” phase, where trained AI models simply execute their existing knowledge. Inference requires dramatically less power and different infrastructure. Data centers optimized for training workloads could find themselves overbuilt and underutilized.

3. Customer Concentration Collapse: CoreWeave’s dependence on Microsoft for 62% of revenue exemplifies the risk. If Microsoft decides to shift workloads to its own Azure infrastructure, reduce AI spending due to slower-than-expected monetization, or simply negotiate more favorable terms with competitors, CoreWeave’s highly specialized facilities become stranded assets. The company’s subsequent attempts to raise $1.5 billion in high-yield debt just weeks after its IPO suggest the financial strain is already acute.

4. Architectural Disruption: New chip architectures, quantum computing advances, or neuromorphic processors could make today’s GPU-heavy infrastructure obsolete. Unlike commercial office buildings that can be repurposed, a data center built for specific AI workloads has limited alternative uses.

“It’s a big, unanswered question,” Tom Mannion of BDO Global told industry analysts. 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.

3. Too Big to Sell: When Scale Becomes a Liability

The third major obstacle is elegantly simple: these facilities are becoming too large for traditional exit strategies. The hyperscale data centers driving the AI boom represent such massive capital commitments that the pool of potential buyers has shrunk to a handful of sovereign wealth funds and the largest institutional investors.

“Few investors are large enough to buy such mammoth companies or even an individual data center,” Tom Mannion explained in interviews. “I mean, they are valued in the tens of billions of dollars.”

The Switch example illustrates this challenge perfectly. The Las Vegas-based data center operator went public in 2017 with a $4.2 billion valuation, then was taken private by DigitalBridge and IFM Investors for $11 billion in 2022. Now its owners are exploring a return to public markets in 2025 at a $40 billion valuation—a nearly 10x increase in just eight years.

But here’s the problem: At $40 billion, Switch’s potential buyer universe has shrunk dramatically. Traditional private equity firms are tapped out. The KKRs and Blackstones have already made their plays, according to Stack Infrastructure CEO Brian Cox. That leaves public markets as the only viable exit—a risky proposition given the sector’s unproven business models and high leverage ratios.

While M&A and industry consolidation might provide exit routes for smaller assets, this strategy breaks down at the highest valuations. The handful of buyers with sufficient “firepower” to acquire properties worth $30-40 billion simply won’t pay the inflated valuations that sellers need to deliver returns to their own investors.

Current Market Dynamics Create a Vicious Cycle:

According to JLL’s data center market analysis, North American data center vacancy rates have plummeted to a historic 2.3% in mid-2025, with average asking rates climbing 12.6% year-over-year to $184.06 per kilowatt. While this seems to indicate strong demand, it’s actually masking future risk: 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

The fourth major concern is that traditional financial engineering solutions—the “creative exits” that have saved PE deals in the past—simply don’t apply at this scale and complexity.

The Public Market Problem

Going public via an IPO is often touted as the golden ticket exit, but CoreWeave’s troubled debut reveals the fundamental challenges. Public market investors demand certain characteristics that AI data centers currently lack:

Predictable, diversified revenue streams: CoreWeave’s 62% concentration with Microsoft and 15% with another unnamed customer (reported by The Information to be Nvidia) means 77% of revenue comes from just two sources.
Positive cash flow and clear paths to profitability: With negative economic earnings of $1.4 billion in 2024 and a debt-to-equity ratio of 1,262.8%, CoreWeave represents exactly the kind of highly leveraged, unprofitable growth story that public markets punish.
Sustainable competitive advantages: What prevents customers from building their own infrastructure or shifting to competitors? CoreWeave’s moat is unclear, especially as hyperscalers like Microsoft, Google, and Amazon continue massive internal infrastructure investments.

The company’s need to seek $1.5 billion in additional high-yield debt financing just weeks after going public underscores the precarious financial position. High-yield bonds signal elevated default risk—exactly what growth equity investors fear.

Insider Protections, Retail Exposure: 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.

The Creative Finance Catch

Other strategies being floated in the industry include:

Loan-to-own structures: Where lenders essentially take ownership positions through debt conversions
Continuation vehicles: Extending hold periods to 15+ years by moving assets into specialized funds
Asset-backed securities: Packaging data center cash flows into tradable securities

But as Andrej Danis of AlixPartners noted in industry discussions: “For these mega data centers, it is not yet clear what will be the eventual vehicle to hold them. Everything here is as speculative as it gets.”

These financial tools work for distressed or non-core assets, not for the “triple-A assets backed by a hyperscaler” that define the current boom. The scale mismatch is fundamental: creating a $40 billion continuation vehicle requires investor appetite that may simply not exist at those valuations, especially if returns disappoint.

The Warning Signs Are Already Flashing

While CBRE data shows record-low vacancy rates and strong leasing activity today, several indicators suggest the market is reaching unsustainable levels:

1. Construction Pipeline Explosion: Over 6.3 gigawatts of data center capacity is currently under construction across North America, with an additional 31.6 gigawatts in planning phases. That’s more than double the entire existing inventory—all betting on continued exponential AI demand growth.

2. Power Constraints Creating Delays: Average wait times for grid connections now exceed 4 years, with power costs rising nearly 30% since 2020. Facilities can’t generate revenue until they have power, extending the period before investors see returns.

3. Debt Market Strain: Asset-backed security (ABS) and single-asset single-borrower (SASB) loan activity increased substantially in H1 2025, but at what point does the debt market refuse to absorb more data center paper? The telecom bubble burst when debt markets closed, leaving over-leveraged companies stranded.

4. Slowing AI Monetization: While AI capabilities advance technically, actual business monetization remains uncertain. McKinsey research suggests that many enterprise AI projects struggle to move beyond pilot phases to production deployment at scale—undermining the demand assumptions underpinning multi-billion dollar infrastructure bets.

Learning From History: Why This Time Might Not Be Different

The most dangerous phrase in investing is “this time is different.” The parallels between today’s AI infrastructure boom and the late-1990s telecom bubble are striking:

Massive debt-financed infrastructure buildout based on exponential growth projections
First-mover advantage narrative driving competitive overbuilding
Technological uncertainty about which architectures and standards will prevail
Limited exit options as assets scale beyond traditional buyer capacity
Insider enrichment while retail investors bear downside risk

The telecom crash saw companies lay 80.2 million miles of fiber optic cable—76% of all digital wiring in the U.S. up to that point—only to have 85% remain unused for years. Bandwidth prices collapsed over 90%. Thousands of jobs vanished. Pension funds evaporated.

But the 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.

What This Means for Private Equity and Infrastructure Investors

The implications are sobering:

For Limited Partners: Question your GPs aggressively about exit strategies. If the answer involves vague references to “strategic buyers” or “market timing,” demand specifics. How exactly will they monetize a $30 billion data center in year seven of a ten-year fund?

For General Partners: Consider whether current valuations price in execution risks, technological uncertainty, and the real timeline to liquidity. The AlixPartners survey showing that less than 50% of respondents have visibility into data center demand over the next 12 months should be deeply concerning for anyone making 10-year infrastructure bets.

For Corporate Strategy: If you’re a hyperscaler or enterprise considering AI infrastructure commitments, recognize your negotiating leverage. Overleveraged data center operators facing exit pressure in 3-5 years will need to keep facilities occupied—giving you pricing power when it’s time to renew contracts.

For Policy Makers: The concentration of AI infrastructure in the hands of a few debt-heavy operators creates systemic risk. If these companies face distress simultaneously, who provides the compute capacity powering critical AI systems?

Conclusion: A Foundation of Capital, or a House of Cards?

An unprecedented wave of capital is flooding into an asset class defined by unprecedented uncertainty. While the world’s tech giants build infrastructure for the AI revolution, the financial architects behind them operate without a proven endgame.

This creates a fundamental conflict between the rush to build and the eventual need to exit. The CoreWeave IPO—with its lukewarm reception, insider protections, and immediate need for additional financing—may be a preview of broader challenges ahead. Switch’s revolving door between public and private markets suggests the industry hasn’t found a sustainable liquidity model.

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.

For private equity investors in tech infrastructure, the AI data center boom presents a stark choice: acknowledge these risks now and adjust positions accordingly, or wait to discover whether they’ve built the information superhighways of tomorrow—or just very expensive monuments to today’s hype.