The scale of investment in AI and its supporting infrastructure is breathtaking. Virtually everything the five major U.S. hyperscalers—Microsoft, Google, Amazon, Meta, and Oracle—do today revolves around AI and data centers. Collectively, they are investing $260 billion in IT in 2025 alone, double the amount in 2023. Hardware providers (Nvidia, Broadcom, Marvell, AMD, TSMC, Intel, Micron, SK Hynix), software developers (OpenAI, Anthropic, Meta, Microsoft), real estate developers, REITs, private equity firms, telecom operators, and power companies are all deploying billions more to accelerate AI adoption. Enterprises are making massive AI investments, and even households are jumping in with AI app subscriptions. Society is placing a multi-trillion-dollar bet that AI will drive the economy for years to come. The key question: what will be the return on this investment?
Economists rarely agree, but projections for AI’s impact on economic growth are particularly divergent—ranging from 0.1% to 1% higher annual productivity growth over the next decade. The closest historical parallel is the internet boom and its supporting infrastructure, such as fiber optic networks, which contributed about 1 percentage point of annual productivity growth from 1995 to 2004. These gains took time to materialize, as widespread adoption was gradual. The same will likely be true for AI—it will take years to build the necessary compute capacity, develop applications, and integrate them into businesses and daily life in ways that enhance efficiency and welfare. AI is a General-Purpose Technology (GPT) that, in time, will reshape every industry—from accounting to asset management, from real estate leasing to healthcare, from customer service to software development. We are in the early stages of adoption.
Data centers (DCs for short) power AI. Information, after all, is just zeros and ones that must be stored somewhere. Over the past two decades, data centers have evolved from simple storage facilities to hubs for enterprise IT outsourcing, cloud computing, predictive AI, and, since 2017, generative AI. Non-AI demand continues to grow at 15% per year, driven by increased cloud migration and digital content creation. However, the real inflection point has been the surge in hyperscaler demand for GenAI compute, propelling data center growth to 33% annually between 2023 and 2030. By 2030, McKinsey reckons that AI workloads will account for 70% of DC demand.
Many DCs are colocation facilities, which house multiple tenants, including enterprises, governments, and smaller tech firms. This segment has had a volatile history. The first colocation bust followed the dot-com crash of 2001-02. From 2013 to 2020, colocation-focused REITs suffered negative revenue growth each year due to declining rents from oversupply. Many companies shifted their workloads to hyperscale cloud platforms, squeezing colocation demand. Today, REITs like Equinix and Digital Realty still have significant colocation businesses. While colocation carries higher tenant credit risk, shorter leases, and sometimes non-pass-through power costs, it also commands higher rents than hyperscale leases. A well-performing colocation facility can be significantly more profitable than a hyperscale one. The DC sector itself has already experienced two boom-and-bust cycles in the past 25 years—could it happen again? And if so, when?
My current view: yes, but not in the next several years.
Why a Speculative Bubble Seems Unlikely—for Now
I see four factors that make another overbuilding crisis less likely in the near term.
- Severe Supply Constraints
- Power is the biggest bottleneck. Data centers consume 10 to 50 times more power per square foot than standard commercial properties. Global data center power demand stands at 55 GW, expected to reach 100 GW by 2029. Utilities face massive upfront investment costs, potential credit rating downgrades, and a multi-year lag between signing a power agreement and seeing full capacity utilization.
- Regulatory pushback is growing. Some jurisdictions are limiting new data center development or mandating payments for unused capacity. In Northern Virginia, where data centers already consume 26% of total power, Loudoun County changed zoning laws in February 2025, making new data centers a “special exception” use.
- Water is another constraint. In Querétaro, Mexico, 600 MW of new data centers require 15 billion gallons of cooling water, exacerbating a severe local water crisis.
- Supply chain bottlenecks are rampant. High-voltage transformers take three years to deliver, server racks two years, and skilled labor (especially electricians) is in short supply. Rising construction costs deter speculative development.
- AI inference workloads require proximity to population centers, where land and power constraints are the tightest.
- Pre-Leasing Mitigates Supply Risk
- 90% of hyperscale data centers are pre-leased with 10+ year contracts before construction begins.
- Hyperscaler tenants are among the most creditworthy corporates globally, with strong balance sheets and cash reserves.
- Pre-leasing ensures that supply growth follows demand, reducing the risk of speculative overbuilding.
- Massive Capital Requirements Limit Speculative Entry
- A single hyperscale facility costs $12 million per MW (or $1,500 per sq. ft.).
- Modern hyperscale data centers range from 150–300 MW, with AI-centric mega facilities exceeding 1 GW, requiring multi-billion-dollar investments.
- Financing involves global bank syndicates, large CMBS/ABS securitizations and equity checks in the hundreds of millions, making it difficult for smaller, riskier developers to enter the space.
- Hyperscaler Profitability Supports Demand
- Cloud services remain highly profitable for Amazon, Microsoft, and Google.
- Advances in AI efficiency (e.g., DeepSeek’s breakthroughs) have shifted more compute spending toward inference, accelerating application development and increasing near-term revenue potential.
What Could Trigger a Crash?
First, a breakthrough in cheap power could upend the industry. Small modular reactors (at least 5–10 years away), fusion energy (further out), or a surge in abundant, low-cost renewables could significantly alter the economics of data center operations. If AI proves to be an undeniable force for societal good, public resistance to expanding power infrastructure may weaken, easing regulatory barriers. Meanwhile, advancements in more efficient cooling systems are already reducing the water intensity of data centers, further mitigating one of their biggest operational constraints.
Second, technological shifts could erode demand for centralized compute, reducing data centers’ long-term value. Moore’s Law continues to improve computing efficiency, and AI chips are becoming exponentially more powerful—although they consume more energy, their performance gains outpace their rising power needs. Nvidia’s latest GPUs, for example, draw significantly more power than earlier models but deliver far greater computational capacity.
Quantum computing poses a potentially paradigm-shifting threat. Microsoft’s introduction of Majorana 1 has reignited the debate over how soon practical quantum systems will arrive. While Microsoft suggests utility-scale quantum computing is just a few years away, other experts argue a 10–20-year horizon is more realistic. Even if quantum computing remains 15 years out, that raises critical questions about data center residual values at the end of a 15-year lease signed in 2025.
Another disruptive force could be edge computing. The development of Small Language Models (SLMs) and techniques like distillation and mixture of experts—pioneered by DeepSeek—suggests that more AI workloads could shift from centralized data centers to powerful handheld devices. If inference can be efficiently distributed across smartphones and other edge devices, it could meaningfully reduce the need for hyperscale compute.
How confident can investors feel about a 5.5%–6% exit cap rate in 2035, given the pace of technological change? While landlords may have recovered their initial investment through cash flows, debt—often with minimal amortization—will still need to be repaid from sale proceeds in year 10. That equation becomes much riskier if advances in quantum computing, edge AI, or more efficient chips could diminish the residual value of traditional data centers.
Third, the balance of power could shift back to tenants. Today, hyperscalers are desperate for data center capacity, giving landlords near-total pricing power. With virtually no vacancy, developers can push rents, lock in long-term leases, and shift risks onto tenants. But this dynamic may not last. If demand for leased data center space slows or hyperscalers experience financial stress, they will cut their leased footprint first—through non-renewals or early lease terminations—before touching their owned facilities.
Unlike other real estate sectors, data center leases often include Service Level Agreements (SLAs), which guarantee uptime and operational standards. If landlords fail to meet these terms, tenants may have grounds to exit their leases early. On the other hand, hyperscalers invest heavily in outfitting their leased facilities, making them "stickier" than traditional office or industrial tenants. Landlords should carefully assess the early termination risk in their portfolios.
Fourth, tenant concentration risk is significant. Hyperscale data centers typically house just one tenant, and it’s almost always one of the same four or five firms. If even one of these hyperscalers falters in the next 10–15 years, it could send shockwaves through the entire sector. Estimating the long-term credit risk of big tech tenants is no easy task. The Global Crossing and Broadcom bankruptcies in 2002 serve as cautionary tales—companies that once seemed indispensable can still collapse.
None of these risks are imminent, nor do they appear particularly likely from today’s vantage point. 2025 is the year of the data center boom. 2028 or 2030? That’s far less certain.
References
Moody’s Ratings, Information Technology Global, “AI innovations could slow infrastructure spending but accelerate broader adoption,” February 3, 2025
Moody’s Ratings, Data Centers – Global 2025 Outlook – “Developer leverage, regulatory risk to rise as growth surges,” January 14, 2025
Moody’s Ratings, Electric Utilities and Power Companies U.S. – “Data center growth poses opportunities and credit risks for power providers,” April 8, 2024
Green Street, Data Center Insights, “Northern Virginia – The Data Center Capital of the World,” February 20, 2025
Green Street, Global Data Center Outlook, January 15, 2025
JLL, Global Data Center Outlook, January 2025
CBRE, North America Data Center Trends H2 2024, February 26, 2025
Goldman Sachs, “AI to drive 165% increase in data centerpower demand by 2030,” February 4, 2025
JP Morgan U.s. Thematic Equity Research, “DeepSeek Implications,” February 5, 2025
HSBC Global Research, “On AI, Deepseek, tariffs & Jevons paradox,” February 11, 2025
Economist, “Tech tycoons have got the economics of AI wrong”, January 30, 2025
Economist, “Training AI models might not need enormous data centres,” January 8, 2025
Davidson DaVinci, “Compute,” February 3, 2025