A Real Estate Strategy for the AI Race

Inference

This is the deployment phase, where trained AI models process fresh inputs and may have to produce results in near-real time: responding to a ChatGPT query or powering AI-enabled enterprise software, for example. Inference aligns most closely with EQT’s core real estate expertise – favoring metro-adjacent locations, deep local market knowledge and disciplined site selection over pure power arbitrage.

Types of inference data centers

Not all inference requires proximity to end users. Hyperscalers can route complex or less time-critical tasks to the lowest-cost facilities, which are often large training campuses with abundant power. If a user uploads multiple documents for analysis, for example, the task may be processed wherever operating costs are cheapest. In those cases, a response time of 60 seconds, rather than 30, won’t be decisive.

Performance-sensitive inference, where functionality depends on speed and responsiveness, is different. As AI becomes embedded in workflows, consumer applications and real-time systems, latency becomes critical. It is this segment of inference demand that is most likely to gravitate toward metro-adjacent locations, where proximity to users and network density enhance performance. These sites are also well-suited to support the continued expansion of cloud workloads, which remain the primary revenue driver for hyperscalers.

The growth of performance-sensitive inference presents investors with opportunities for attractive risk-adjusted returns, provided they understand the nuances of location, power and end-user demand. This is partly due to their structural tailwinds; AI training will be done in a short, intense burst, but inference scales with user numbers and usage. That’s why McKinsey projections show inference surpassing training to become the dominant workload in AI data centers by 2030, representing more than half of all AI compute and roughly 30 to 40 percent of total data center demand.

Granted, there are uncertainties around the growth trajectory of inference workload, too, but these facilities sit on multi-use land with alternative exit options. They are smaller, distributed and diversified, offering better residual value and downside protection.

Scale unlocks power

EQT owns around 350 million square feet of industrial real estate in the U.S., and is among the largest landlords in key growth markets for data centers, including Columbus, Ohio; Fort Worth, Texas; and Nashville, Tennessee. If we were to convert just one percent of that footprint into data centers, it would make EQT one of the largest real estate funds invested in the sector.

Satisfying inference demand fits neatly with our existing real estate business. We have local teams in more than 24 markets who know the sites near end-users, highways, fiber and power. That enables us to buy or repurpose industrial land and buildings and layer inference or cloud capacity on top, creating assets that benefit from both digital demand and enduring real estate fundamentals.

That said, our strategy is not dictated by land arbitrage. Where appropriate, we develop and retain powered shell facilities, delivering the structure and securing power while allowing end users to install their own systems. This approach aligns long-term ownership with durable income streams, rather than short-term land value uplift.

Power remains critical. As with remote training campuses, access to reliable, scalable energy is still often the binding constraint. The grid is overwhelmed, and the hyperscalers are implementing innovative solutions to circumvent slow connections. Natural gas is gaining appeal, for example, as it can power on-site engines, turbines or fuel cells, providing a measure of speed and flexibility where grid access is limited. Nowadays, buying land is often akin to buying a place in the power queue. The ability to accelerate that timeline is increasingly where value is created.

We use proprietary tools to find these sites. Our team has mapped the entire U.S. electrical grid, natural-gas network and zoning framework into a single platform. Each potential site is scored on access to power, proximity to substations and gas pipelines, zoning, site size and availability, allowing us to compare opportunities quickly. The tool alerts us to opportunities to embed inference capacity within our existing portfolio, or within industrial portfolios currently on the market.

Value uplift

This is a particularly opportune moment to pursue this strategy, early in a new real estate cycle. Liquidity in the industrial market remains constrained by elevated interest rates, and larger portfolios that would previously have traded at a premium are now changing hands at a discount.

Within these portfolios, we consistently identify embedded data-center optionality. The value uplift from converting raw land into powered sites can range from two to ten times, depending on grid access timing, entitlement speed or market demand – meaning a disciplined, data-led approach can crystallize returns quickly while preserving downside protection.

In some respects, today’s AI expansion resembles a modern energy rush. Developers pore over transmission maps instead of geological surveys, searching for electricity rather than oil. Yet unlike past commodity booms, it is possible to capture much of the upside while limiting exposure to the downside by investing in proximity to people and enterprises.

For EQT’s Real Estate team, the opportunity lies in combining power expertise with real estate fundamentals at scale. By focusing on assets that serve end users, retain alternative-use value and benefit from structural growth in inference demand, we believe it is possible to participate in the AI transformation while preserving resilience across cycles.