Beyond the Bubble: Why We Think AI Infrastructure Will Compound Long after the Hype

• There is both froth in parts of the AI ecosystem and real breakthroughs in models and applications. Past overbuilds in rail, electrification, and fiber seeded critical economic change, and we believe long-term data center demand will justify current activity.
• While often compared to late-1990s fiber, today’s data center cycle is fundamentally different, underpinned by long-term contracts with the world’s most advanced technology companies, and capability, power, and land emerging as key constraints on growth.
• Several factors will separate winners from losers. The first is underwriting that considers the profitability of individual projects after the cost of power and capital.
• The second is owning the competitive moats. The barriers to entry in data centers are significant and include power, land, interconnects, permits, and the operational excellence and strong relationships necessary to work with hyperscalers, the major cloud providers that make up the largest customer base for computing capacity.
• And third, discipline and de‑risking: structuring long-tenor offtake agreements, balancing counterparty exposure, locking in terms, planning for upgrades as technology evolves, and avoiding models that rent scarce inputs at thin spreads. 

When a single chipmaker, NVIDIA, makes up 8% of the S&P 500, it’s reasonable to wonder whether an AI bubble is inflating. Lately, that scrutiny has extended to data center investment. A few statistics help illustrate both the scale and the constraints:

• McKinsey estimates that companies will invest almost $7 trillion in global data center infrastructure capital expenditures by 2030.¹ That’s the size of the combined GDP of Japan and Germany. 
• In the United States, AI-related capital expenditures account for about 5% of GDP (Exhibit 1) and have been growing at high-single- to low-double-digit pace.
• In the first half of 2025, AI‑related capex contributed materially to U.S. GDP growth – more than consumer spending. The four largest hyperscalers (Amazon, Google, Microsoft, Meta) are expected to spend more than $350 billion on capex in 2025, a year-over-year increase in the mid-30% range. Including other tech players pushes the total toward an estimated total of about $0.5 trillion in 2025.

EXHIBIT 1: AI-Related Capex Make Up Roughly 5% of U.S. GDP and Is Growing ~10% Per Year, On Par With the Late-90s / Early 2000s Tech Boom

Investment Spend as % of U.S. GDP

Our view: Yes, there’s froth. Yes, there will be a shake‑out.

However, current absorption rates show no signs of overbuilding in the world’s most active market, the United States, according to JLL Research (Exhibit 2). And in the longer term, we think demand should justify much of today’s data center build-out. Models and applications will keep advancing, and they will pull capacity from systems capable of powering them at scale.

EXHIBIT 2: Available Data Center Space Will Be Scarce at Least through 2027

North America Colocation Vacancy

Past technology-related infrastructure hype cycles suggest that the data centers, electrical infrastructure, and fiber networks being built are unlikely to go to waste. Instead, these hard assets will likely form the backbone of a new economy and achieve compounding returns. Until then, expect that some asset prices will become inflated, and some business models in the broader AI ecosystem won’t survive.

In other words: Bubbles always hurt some investors, but the capacity they create endures.

In the meantime, investing without getting burned means de-risking upfront and making realistic assumptions about project economics based on returns after the (increasingly expensive) cost of power and capital. Secure offtake agreements can guarantee a tenant will pay for contracted capacity over a period of time, whether they use it all or not. And with power and land increasingly the limiting factors in data center development, we only take on projects with both sufficient electricity and entitlements, meaning all the various permits, permissions, and agreements required to build a functional data center. 

Lessons from Bubbles Gone By

Every revolutionary technology requires new infrastructure. Those build-outs tend to follow a pattern George Soros called reflexivity. Enthusiasm drives capital, and the availability of capital invites more demand for infrastructure. But that demand often arrives ahead of its time as animal spirits outrun fundamentals. Then the shake-out comes, cleansing weak business models and balance sheets, while the assets remain.

EXHIBIT 3: A Timeline of Past Tech Infrastructure Bubbles

EXHIBIT 4: The Anatomy of Past Technology Infrastructure Bubbles

As with fiber, even when capital overshoots, durable physical networks tend to become the substrate for the next wave of applications.

We Believe Data Centers Have Important Fundamental Differences

What will the shake-out in the AI ecosystem look like? We often hear comparisons between today and the build-out of long-haul fiber connections and metro fiber networks during the late 1990s. But understanding why data centers are different can help illustrate where today’s risks are – and where they are not.

• Intelligence for Everything: The value of some technology, like railroads and the internet, is in connectivity. But AI is a horizontal computational layer that permeates workflows, creative work, analysis, routine tasks, and decision‑making across sectors. It raises the floor of productivity in many places at once, rather than concentrating value in a handful of destinations, i.e., the consumer applications and ad networks of the internet economy. A better comparison? Electrification. Productivity gains didn’t immediately arrive with electric lamps, but rather when factories began to be redesigned around electric motors. This paved the way for a wave of speed, efficiency gains, and automation that continues today.   
• Long-Haul and Metro Fiber Invited Speculative Building:⁵ After opening the ground and laying conduit (the protective pipes that strands of fiber pass through), maintenance is minimal for decades, and the marginal cost of pulling extra strands of fiber through to increase capacity is low. It’s also possible to upgrade capacity by installing new electronic equipment, or optics, at each end of the installation, rather than rip‑and‑replace construction. This scalable infrastructure was built in response to rapid anticipated demand, but without contracted long-term offtake for that capacity.
• Building a Speculative Data Center Isn’t Easy: Building a data center is so capital-intensive that raising funds for speculative investment would be difficult, and carrying extra capacity is expensive due to material ongoing operating costs. Those operating costs are rising, as the amount of power and cooling needed for every metal shelf, or rack, that holds servers, storage, and networking gear increases. In addition, computing components need refreshing every few years; idle capacity erodes returns rather than “waiting” for demand.  
• Data Centers Are Typically Contracted: Data center construction typically only happens with a customer contract in place, complete with offtake requirements. We’re not seeing vacant “Field of Dreams” sites (“If you build it, they will come”) or construction that relies on merchant power (unhedged or non-firm supply priced at spot). We see some speculative activity in work related to power, such as building substations, and buying land for potential future sites, but land generally represents a modest share of total project capex at about 10%.
• Fast Refresh Cycles Soak Up Extra Demand: Accelerators—the computing hardware in a data center—age quickly. Each generation arrives faster with step‑function performance gains, meaning that progress is sudden, significant, and discontinuous, looking more like a vertical step on a line chart than a smooth upward curve. Excess capacity doesn’t sit idle for long, as new workloads and model classes tend to pull it in. Temporary overbuilds behave like rolling upgrades rather than stranded assets.
• Power Limits Overbuild Potential: The ultimate limiting factor for building data centers isn’t capital; it’s power. Queues to connect to the grid, transformer lead times, and the difficulty of locating sites and obtaining permits make unconstrained overbuilds impractical. Owners who control grid access and have full permits are at an advantage.

Large-Scale Tech Infrastructure Builds Are Rarely a Waste

The dot-com era does have one lesson worth noting. The companies building too much fiber weren’t wrong about the long-term need for the cables. What they got wrong is how fast it would materialize.

Consumption of telecom services grew faster than total consumption before, during, and after the boom. In 2001, when the so-called bubble burst, telecom services had risen to 2.4% of total consumption from 1.7% in 1995.⁶

The construction of fiber, much like the construction of data centers, catalyzed further innovation. Every new application led to demand for more bandwidth, and more bandwidth made more new applications possible.

We believe the same is true of data centers. We expect large-language models (LLMs) and generative pre-trained transformers (GPTs) to become more efficient, requiring less computing power and reducing costs – just as happened with electrification (Exhibit 5a and 5b). At the same time, we expect adoption to rise and create new applications that stoke more demand, just as in previous cycles (Exhibit 6).

EXHIBIT 5A
Cost per Million Input Tokens (USD)

EXHIBIT 5B
Cost of Light- Hours of Work per 1000 Lumen-Hours

EXHIBIT 6: Lower Unit Costs Drive Accelerated Adoption and Demand Across Prior Supercycles

Getting in on the Ground Floor Without Getting Crushed

Two truths can hold at once: (1) This is a generational compute shift that will require tremendous infrastructure investment; (2) We’re still in the early stages in separating the signal from the noise.  We are true believers in the potential of AI and committed to investing in AI infrastructure across data centers, power, and connectivity. But we also see a need for strict discipline. A few tenets we live by when it comes to structuring data center investments include:

Those who control the moats should reap the compounding returns. Power, land, grid connections, and permits are structural bottlenecks to building data centers. Customer relationships and operational expertise are less tangible obstacles developers and managers must clear to work with the lowest-risk counterparties, investment-grade hyperscalers. We think it will be difficult for business models that rely on renting GPUs or power to achieve sustainable differentiated returns. Those who own the tangible resources and master the intangible ones are in a better position to win.

Unit economics are more important than hype. The prices of land and power are rising, especially in key data center markets. We care about return on invested capital after power and capital costs, not theoretical total addressable markets. The important question is whether contracted returns hold once utilization, cost of capital, and operating leverage normalize in the current cycle.

Execution matters. In addition to controlling scarce inputs, we think the ability to build what you promise will be a key long-term differentiator. Those who can quickly deliver a fully executed lease or development agreement, achieve reasonable costs per delivered megawatt, keep built data centers operational and available (uptime), and renew leases at profitable rates will be at an advantage.

De-risking is critical. Positioning for AI fatigue must happen before its arrival. That means securing contracts with long-term offtake agreements – meaning that data center tenants pay whether they use the facility to full capacity or not – balancing counterparty exposure, and building with flexibility in mind, so that we can pivot as technology and demand evolve.

Exits and Valuation Dynamics

The same factors that help de-risk our investments also set the stage for more favorable exits. Stabilized, fully operating data centers attract core and core-plus infrastructure funds, sovereigns, and listed platforms seeking long-duration, investment-grade cash flows. Portfolio buyers are typically willing to pay a premium for platforms with secured power access, entitled land, and expansion capacity.

Investors usually value data centers either by income yield or by enterprise value per delivered megawatt (MW), then adjust based on contract strength, expansion potential, and power certainty. Exhibit 7 illustrates a simplified unit-economics model based on income yield.

Valuations rise for platforms that combine:

• Access to scarce power and land in super-core markets — for example, Slough in London, Singapore, or Northern Virginia, where demand density and network proximity make expansion capacity uniquely valuable
• Long-term, take-or-pay contracts that require the tenant to pay for a certain level of usage, whether or not they actually use the full capacity of the site, with operations and maintenance cost pass-throughs
• Titled land and expansion permits
• Robust interconnects (high-speed fiber and network switching)
• Proven uptime and cost discipline
• Designs supporting higher power densities

Conversely, assets that lack fungibility—those built to bespoke specifications, located in non-core or peripheral areas, or situated far from major population and network centers—carry higher residual and re-use risk. Facilities with limited alternative uses or uncertain long-term relevance should be valued differently from scalable, well-located platforms with durable demand anchors. Similarly, exposure to single-tenant concentration, short-term leases, uncertain power rights, leased land, or thin margins after energy and capital costs can further weaken sale prospects and valuation multiples.

The KKR Global Infrastructure team has been investing in data centers since 2019, building one of the most active and globally diversified portfolios in the sector. We’ve now established five major data center platforms spanning hyperscale, colocation, and edge infrastructure.

Over the past six years, we’ve committed $31.3 billion in equity capital to digital infrastructure investments, reflecting our conviction in the theme.

We take a disciplined, conviction-led approach. Every assumption is challenged, with teams debating bull and bear cases and pressure-testing logic from all sides. We set clear exit or stop-loss criteria upfront and quickly shut down what doesn’t compound—freeing capital and focus for the “yes” opportunities that matter.

We also recognize that the center of gravity is shifting. AI workloads now sit at the intersection of digital, power, renewables, and industrials. We think our “One KKR” model—sharing insights, aligning customer relationships, and coordinating capital across verticals—sets us apart in a world where data, energy, and compute are rapidly converging, and where silos can kill even the best ideas.

EXHIBIT 7: An Illustrated Guide to Data Center Unit Economics

Conclusion

History shows that technological revolutions often overshoot in the short term but compound in the long term. The same forces are at work in AI infrastructure today. Valuations may look stretched, but the hard assets being built — data centers, power, and connectivity — will anchor the next wave of digital productivity.

Investors who focus on execution, unit economics, and risk discipline will separate the signal from the noise. As compute, storage, and energy converge, controlling scarce inputs — power, land, and grid access will define who wins.

The bottom line is, in our view, the AI buildout isn’t a bubble. It’s the backbone of the next industrial revolution — and those who build it with patience, precision, and conviction will draw its map.

REFERENCES

¹ McKinsey & Co. “The data center balance: How US states can navigate the opportunities and challenges.” August 8, 2025. https://www.mckinsey.com.br/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges

² Ronald C. Tobey. Technology as Freedom: The New Deal and the Electrical Modernization of the American Home. University of California Press. 1997.

³ Long-haul fiber networks run over long distances, while metro fiber connections serve one major metropolitan area. A third type of fiber connection, fiber-to-the-home and fiber-to-the-business, connects individual buildings to metro networks or regional networks in more rural areas.

⁴ Elise A. Couper, John P. Hejkal, and Alexander Wolman. “Boom and Bust in Telecommunications.” Economic Quarterly  of the Federal Reserve Bank of Richmond. Volume 89/4, Fall 2003.

⁵ This might sound strange coming from a large fiber investor. We do indeed invest heavily in fiber today, but our model is very different than 1990s fiber. We focus on wholesale models, where multiple vendors can fill open capacity, and last-mile fiber connections to individual homes and businesses, which are less vulnerable to speculative builds. 

⁶ Elise A. Couper, John P. Hejkal, and Alexander L. Wolman. “Boom and Bust in Telecommunications.” Federal Reserve Bank of Richmond Economic Quarterly. Volume 89/4, Fall 2003.