Every major technology wave of the past 25 years has driven a data center demand surge. The internet boom. Cloud migration. Mobile. Streaming. Each drove a cycle of development, a period of oversupply, and then equilibrium. AI is different — not in degree, but in kind.

This lesson explains why, what physical changes AI is forcing on the built environment, and what the investment implications are through 2035.

AI is different from prior data center waves

To understand why AI is a structural inflection point rather than another cyclical wave, it helps to compare it to what came before.

The internet era (1999–2005) drove the first real data center construction boom. Companies needed servers to host websites, run email, and process e-commerce. Power density was low — 1 to 3 kW per rack — and the primary constraint was floor space. The dot-com crash created a severe oversupply, and many facilities sat half-empty for years.

Cloud migration (2008–2015) was more durable. Amazon Web Services, Microsoft Azure, and Google Cloud drove hyperscale demand. These facilities were efficient, highly engineered, and optimized for cost. The big difference from dot-com: the hyperscalers had durable revenue models. Cloud demand never stopped — it just grew more slowly than initial projections. Power density crept up to 5 to 8 kW per rack on average.

Mobile and streaming (2012–2020) added enormous demand for video delivery infrastructure, CDN edge nodes, and content storage. Netflix, YouTube, and the app economy drove billions in new CapEx. But the fundamental compute architecture remained similar — CPUs serving files and running web applications.

AI is different along five dimensions:

Power density is 5 to 10 times higher than traditional cloud workloads. A GPU rack running AI training pulls 30 to 100 kW. A CPU server rack runs 2 to 8 kW. This is not a marginal increase — it is an order-of-magnitude shift that invalidates most existing infrastructure.
Every enterprise is building AI capabilities simultaneously. In prior waves, adoption was sequential — enterprises moved to cloud over a decade. AI adoption is happening in parallel across every industry at once. Banks, hospitals, manufacturers, retailers, logistics companies, and government agencies are all procuring AI compute in the same window.
Training new AI models requires enormous one-time compute bursts. Training GPT-4 required approximately 25,000 A100 GPUs running continuously for months. Every major AI lab is training the next generation of models right now. These training runs consume megawatts of continuous power in facilities that must operate without interruption.
Inference at scale requires always-on GPU clusters. Training is one-time per model. Inference — serving that model's outputs to users — is continuous. Every ChatGPT query, every Copilot suggestion, every AI-driven recommendation engine is consuming GPU compute around the clock. Inference demand compounds with adoption and does not plateau.
No signs of demand moderation. GPU shortages have persisted since 2022. NVIDIA's lead times on H100 and H200 GPUs stretched to 12 months or more at peak. Hyperscaler CapEx — the most reliable leading indicator of data center construction — reached record levels in 2024 and 2025 and is forecast to continue growing. This is not a bubble pattern; it is constrained supply against genuine demand.

The net result: AI is not a demand spike that the industry will build through. It is a permanent step-change in compute intensity that requires a fundamentally different built environment.

Power density — the AI problem for real estate

The power density gap between traditional and AI workloads is the defining challenge for data center real estate right now.

A traditional data center was designed around a planning assumption of 2 to 5 kW per rack. Some older colocation facilities were built for as little as 150 to 200 watts per rack — they were glorified climate-controlled warehouses for 1U servers.

AI training clusters require 10 to 30 kW per rack as a baseline. Cutting-edge AI labs running NVIDIA DGX H100 or H200 systems are pulling 40 to 100 kW per rack. Some experimental GPU dense configurations are pushing higher.

To put this in concrete terms: a typical 100-rack colocation cage in a traditional data center draws 200 to 500 kW total. The same 100-rack footprint configured for AI training would draw 1,000 to 3,000 kW — two to six times the power capacity of the entire traditional facility, concentrated in a fraction of the floor space.

This creates three immediate real estate dynamics:

Retrofit demand. The building shell may be fine. The problem is the power infrastructure — the utility feed, the transformer vault, the UPS systems, the generator capacity, and the distribution within the facility. Retrofitting a 5 MW traditional facility to deliver 15 MW of AI-capable power requires replacing electrical infrastructure from the utility interconnection to the rack PDU. These projects cost tens of millions of dollars and take 18 to 36 months.

Premium pricing for AI-ready facilities. Facilities that can deliver high-density power today command substantial premiums. Colocation pricing for AI-dense deployments runs 3 to 5 times the per-kW rate of traditional colocation. Tenants pay the premium because the alternative — waiting for new construction — adds a year or more to their deployment timeline.

Massive greenfield development pipeline. The hyperscalers have decided that retrofitting is not fast enough. The scale of AI infrastructure buildout requires purpose-built campuses designed from day one for extreme power density. This is driving the largest greenfield data center development pipeline in history.

Liquid cooling — the physical infrastructure shift

Air cooling has been the standard for data center thermal management since the first raised-floor machine room. Hot aisle / cold aisle containment, precision air conditioning units, and computer room air handlers (CRAHs) have served the industry for 40 years.

Air cooling hits its physical limits at approximately 20 to 30 kW per rack. Above that threshold, the volume of air required to carry heat away from the rack becomes impractical — you would need air velocities that create excessive noise and pressure drop, and the ductwork required would consume more space than the servers themselves.

AI workloads break the air cooling model. The industry has shifted to liquid cooling, which carries heat 1,000 times more efficiently than air per unit volume. There are three primary liquid cooling architectures now deployed in AI data centers:

Direct Liquid Cooling (DLC) routes coolant through a heat exchanger mounted directly in the rack, or through cold plates attached to the GPU and CPU. The chips are not submerged — coolant runs through metal plates in direct contact with the processor package. DLC can handle 30 to 100+ kW per rack and is now the standard approach for NVIDIA DGX-based systems.

Immersion cooling submerges servers entirely in a dielectric fluid — a liquid engineered to be electrically inert. Single-phase immersion uses a fluid that remains liquid throughout. Two-phase immersion uses a fluid that boils off the chip surface and condenses on a heat exchanger above, capturing the phase-change energy. Immersion cooling can theoretically handle any power density the hardware can generate, but it requires purpose-built tanks and is incompatible with standard rack form factors.

Rear-door heat exchangers attach to the back of standard server racks and capture heat from the exhaust air stream by passing it through a coolant-filled radiator. This is the easiest retrofit path — it works with standard rack infrastructure — but has limits in how much heat it can remove and requires a facility-level chilled water loop.

Why liquid cooling matters for CRE investors: liquid cooling is not a minor equipment upgrade. It requires:

Facility-level plumbing infrastructure — supply and return manifolds, secondary cooling loops, leak detection systems
Different floor loading — immersion tanks and filled pipe runs are significantly heavier than raised floor air systems
Mechanical room redesign — chillers, cooling towers, or dry coolers replacing or supplementing CRAC/CRAH units
Different maintenance protocols — facilities staff must be trained and certified for liquid systems

This changes the retrofit calculus for older facilities. A building that cannot support the structural loads of liquid cooling infrastructure, or that lacks the mechanical room footprint for supplemental cooling plant, may be effectively stranded regardless of its electrical capacity. Liquid cooling compatibility is now a specification requirement in hyperscale pre-lease RFPs.

The CapEx pipeline

The numbers behind AI data center buildout are difficult to contextualize because they are genuinely unprecedented.

In 2025, the major hyperscalers announced the following data center CapEx programs:

Microsoft: $80 billion
Amazon (AWS): $105 billion
Google: $75 billion
Meta: $65 billion

Combined, the four largest hyperscalers committed approximately $325 billion in a single calendar year. For comparison, the entire U.S. office real estate sector — all office buildings in America — has an estimated total value of roughly $3 trillion. The hyperscalers committed over 10% of that value in a single year, in a single asset class.

Where does this capital go? The rough allocation for a purpose-built hyperscale data center campus:

Land and site work: 3 to 5% of total project cost
Shell construction: 8 to 12%
Power infrastructure (utility feed through UPS to distribution): 30 to 40%
Mechanical / cooling systems: 20 to 25%
IT equipment (servers, networking): variable — often 50%+ of total campus investment

For CRE investors, the relevant observation is this: the supply of powered, cooling-ready data center space cannot be built fast enough to absorb $325 billion per year of hyperscaler demand, let alone the enterprise and colo demand on top of it. The utility interconnection queue — the list of projects waiting for grid power — in most major U.S. markets now extends 4 to 6 years. You cannot build a data center if you cannot get power, and right now you cannot get power quickly anywhere in America.

This supply constraint is the fundamental reason data center investment fundamentals remain strong.

AI campus — the new asset class

The hyperscaler buildout at this scale has created an asset class that did not meaningfully exist five years ago: the AI campus.

An AI campus is a purpose-built facility or cluster of facilities designed to deliver 500 MW to 2 GW of IT load for AI training and inference at extreme power density. These are not conventional data centers scaled up. They are fundamentally different facilities.

Characteristics of AI campuses:

Extreme power density — 30 to 100 kW per rack throughout, not in pockets
Colocation of power generation — AI campuses are increasingly being sited adjacent to power sources: wind farms, solar arrays, natural gas peakers, or nuclear plants
Purpose-built mechanical infrastructure — liquid cooling designed in from the start, not retrofitted
Single hyperscale tenant — AI campuses are almost always built-to-suit for a single tenant rather than multi-tenant colo
20 to 25-year leases — the lease terms match the expected lifecycle of the mechanical infrastructure
Critical facilities redundancy — N+1 or 2N redundancy on all critical systems, but focused on uptime rather than traditional Tier classification

Notable examples:

Microsoft's deal with Constellation Energy to restart the Three Mile Island nuclear plant — renamed Crane Clean Energy Center — is the clearest illustration of where this is heading. Microsoft committed to a 20-year power purchase agreement for 835 MW of nuclear output specifically to power data centers. The logic: AI training cannot tolerate carbon-intermittent power because the compute cannot be interrupted, and nuclear provides always-on, zero-carbon baseload. This was not a real estate transaction in the conventional sense — it was a real estate transaction that bundled power generation.

xAI's Memphis AI campus, built by Elon Musk's AI company in 2024 and 2025, demonstrated how quickly these facilities can be deployed when a hyperscale tenant has sufficient urgency and capital. The facility went from site selection to operational GPU cluster in under a year — an extraordinary construction pace that required prefabricated modular power infrastructure.

For CRE investors, AI campuses represent a new category at the top of the risk-adjusted return spectrum: long-term, credit-tenant leases, trophy asset characteristics, and scarcity value driven by the power interconnection queue.

Investment implications for the next decade

Looking at data center CRE through 2035, several structural trends are clear.

Cap rate compression will continue. Core data center cap rates have already compressed below office and retail to levels competitive with multifamily and industrial. As AI demand persists and supply remains constrained by power availability, the institutional capital flowing into this sector will continue to push prices up. The risk premium for data center over traditional CRE is compressing as the asset class becomes better understood and more liquid.

Development yields remain strong for well-positioned projects. The spread between development cost and stabilized value — the development profit — remains wide in data center because construction cost has not caught up to the value appreciation. A developer who can secure utility power interconnection, deliver a spec powered shell in a supply-constrained market, and attract a hyperscale tenant can still generate development yields in the 15 to 25% range on cost in the right markets.

Bifurcation between AI-ready and legacy facilities will widen. The data center market is splitting. AI-ready facilities — built or retrofitted for high density, liquid cooling capable, with multi-megawatt utility feeds — will command significant premiums. Legacy facilities — built for 2 to 5 kW per rack, air-cooled, with limited power capacity — face obsolescence risk. Some will be repurposed for warm storage or edge use cases. Others will be stranded. Investors holding legacy data center assets need an honest obsolescence analysis before 2030.

Florida-specific consideration — utility interconnection queues. Florida's data center market is emerging, driven by population growth, tax climate, and proximity to Latin American enterprises. But Florida utilities — Duke Energy Florida, Florida Power & Light, Tampa Electric — are managing congested interconnection queues like the rest of the country. Developers and investors who want to build or acquire data center assets in Florida need to be in the utility queue now. The queue process takes 2 to 4 years in most Florida markets. Projects that do not begin the interconnection process in 2025 or 2026 will not have power until 2028 or 2029 at the earliest.

How to position in data center CRE today

Different investor profiles have different optimal entry points.

Passive investors ($1M–$10M equity)

For investors who want data center exposure without the operational complexity, the most accessible path is publicly traded REITs:

Digital Realty (DLR) — the largest data center REIT by square footage, with a global portfolio and strong hyperscale relationships
Equinix (EQIX) — the largest data center REIT by market cap, focused on interconnection-dense colocation
Iron Mountain (IRM) — originally a document storage company, now one of the fastest-growing data center operators with a focus on powered shell development

For private market exposure at this scale, private credit to data center developers (senior construction loans or mezzanine debt) can provide yield with less equity volatility than REIT shares.

Active investors and developers ($20M–$100M equity)

The most compelling opportunity for mid-market developers is powered shell development in secondary markets with available utility capacity. The hyperscalers cannot wait for primary markets to clear their interconnection queues. They are actively looking in markets like Central Florida, the Carolinas, Ohio, Indiana, and the Mountain West — anywhere that can deliver power at scale in a 24 to 36-month window.

A powered shell strategy: acquire a well-located industrial site with proximity to a substation, work through the utility interconnection process, build a shell building with the mechanical and electrical infrastructure in place, and pre-lease to a hyperscale or large enterprise tenant. The premium over a standard industrial development is significant, and the tenant quality is superior.

Institutional investors (above $100M equity)

At this scale, the primary opportunities are hyperscale build-to-suit development and portfolio acquisition from REIT spinoffs or private equity data center platforms. REITs are increasingly separating their data center assets from other business lines, creating acquisition opportunities. Private equity platforms that assembled data center portfolios in 2018 to 2022 are now reaching their hold period and will be exiting through sale processes.

What to take away

AI is a structural inflection point for data center real estate, not a cyclical wave — demand is permanent and compounding
Power density in AI workloads is 5 to 10 times higher than traditional cloud, requiring fundamentally different facilities
Air cooling is physically inadequate for AI workloads above 30 kW per rack — liquid cooling (DLC, immersion, rear-door) is the new standard
Hyperscaler CapEx commitments hit $325 billion in 2025 — the largest single-year capital deployment in CRE history
Utility interconnection queues now extend 4 to 6 years in most major U.S. markets, creating a durable supply constraint
AI campuses — purpose-built facilities near power generation with 20+ year single-tenant leases — represent a new asset class at the top of the institutional market
The data center market is bifurcating: AI-ready facilities command premiums, legacy air-cooled facilities face obsolescence
Florida investors need to enter the utility interconnection queue now — projects starting in 2026 will not have power before 2028 to 2029
Passive investors access the sector through Digital Realty, Equinix, and Iron Mountain; active developers have opportunity in powered shell development in secondary markets with available utility capacity

This lesson completes the AI and demand section of the data center course. The next lesson covers how to underwrite a data center acquisition — the specific financial metrics, lease structure analysis, and due diligence checklist for CRE investors entering this asset class for the first time.

The AI Demand Surge — What It Means for Data Center Real Estate Through 2035