The Industry Is No Longer One Asset Class
The data center market crossed $347 billion in 2026 and is tracking toward $801 billion by 2033. That aggregate number obscures what actually matters for enterprise infrastructure decision-making: the market has divided into two structurally different asset classes that cannot be evaluated with the same underwriting criteria, lease terms, or operational assumptions.
On one side: AI factories. Facilities designed from inception around sustained GPU utilization, extreme rack density, direct-to-chip liquid cooling, and non-interruptible power delivery at hundreds of megawatts per campus. On the other: general-purpose data centers. Facilities built for elastic, mixed workloads — cloud hosting, enterprise applications, storage, interconnection — with grid-connected power, hybrid cooling, and phased capacity delivery.
The distinction is not marketing taxonomy. It reflects fundamentally different capital structures, power architectures, regulatory exposure, and community relationships. Enterprises that conflate the two will overpay for the wrong infrastructure, underprovision for AI workloads, or lock into lease structures misaligned with their actual demand profile.
This is the strategic question 2026 forces: Which infrastructure class does your workload actually require, and what is the right decision framework for accessing it?
What Defines an AI Factory
The AI factory is not a data center with more GPUs. It is a purpose-built facility organized around a single governing constraint: power — its scale, continuity, and delivery architecture.
Where a general-purpose facility treats utility grid power as its primary source and diesel generators as contingency, an AI factory treats onsite generation as a co-equal or primary architecture element. Sightline Climate’s 2026 data center outlook tracks this directly: onsite and hybrid power approaches account for fewer than 10% of projects by count but nearly half of all announced capacity. The math reflects the gigascale concentration — a single grid-independent campus like New Era Energy & Digital’s 7 GW project in New Mexico shifts the aggregate dramatically.
Behind-the-meter natural gas provides dispatchability and scale in the near term. Nuclear — specifically small modular reactors and microreactors — is already influencing long-cycle site selection decisions. Google’s acquisition of Intersect Power’s 10.8 GW pipeline and Amazon’s direct investments in solar and storage signal that hyperscalers are no longer waiting for utilities to solve the interconnection problem. They are becoming their own power architects.
For the enterprise buyer, this matters in two ways. First, AI factories carry a different power cost structure: onsite generation reduces exposure to spot grid pricing volatility but introduces fuel supply and operational dependencies that don’t exist in a standard colo lease. Second, AI factory campuses routinely plan for hundreds of megawatts of firm, non-curtailable load — a profile that compresses utility planning cycles and makes these facilities visible to regulators, communities, and grid operators in ways that a 10 MW enterprise colo rack never was.
Cooling Is Not an Upgrade — It Is a Redesign
General-purpose data centers have adopted liquid cooling incrementally. Rear-door heat exchangers, localized liquid-cooled zones for high-density clusters, selective direct liquid cooling where a specific row requires it. This approach works when AI workloads are a minority of the floor. It does not work when AI is the primary workload.
AI factories design cooling into the structural architecture before a single server is installed. Direct liquid cooling to the chip is the baseline, not the premium option. Thermal architectures are engineered for continuous high-intensity GPU utilization — not peak handling, but sustained throughput. Immersion cooling is viable selectively where density and thermal headroom demand it.
The operational implications extend beyond the mechanical systems. Cooling defines floor loading specifications, piping infrastructure, redundancy architecture, commissioning sequences, and maintenance protocols. A facility that was designed for air cooling — even with localized liquid cooling overlays — cannot be retroactively converted to support AI factory workload profiles without a capital investment that effectively constitutes a rebuild.
Submer’s 2026 analysis frames it plainly: liquid cooling is no longer optional for high-density AI infrastructure. Proficiency in both direct liquid cooling and immersion is now a prerequisite for operators serving AI workloads, not a differentiator.
This is the first classification question for any enterprise evaluating infrastructure: Does the workload require sustained GPU compute at scale, or is it a mixed profile that can be supported by an air-cooled facility with liquid cooling zones? The answer determines which asset class is relevant — and which operators have infrastructure that can actually deliver the performance specification.
The Invisibility Loss: AI Factories Under Public Scrutiny
General-purpose data centers have historically operated in a regulatory blind spot. Power draw is distributed and modest by industrial standards. Cooling systems are largely self-contained. Community impact is limited. Permitting is routine.
AI factories have ended that era.
Axios reported in February 2026 that Sightline Climate had identified more than ten new moratorium proposals across U.S. states in a single month — New York, Michigan, Virginia, and Oklahoma among them. These proposals are concentrated around the energy intensity and grid impact of AI infrastructure, not general-purpose data centers. Virginia, which has 6.3 GW under construction — roughly 25% of the entire Americas pipeline according to Cushman & Wakefield’s Q4 2025 market report — is now a jurisdiction where approval complexity and community opposition materially affect development timelines.
The 777 large facilities tracked by Sightline Climate since 2024 represent 190 GW of announced capacity. Of the 16 GW slated for 2026 delivery, only 5 GW is actively under construction. The remaining 11 GW sits in announced status — and 30 to 50 percent of the full 2026 pipeline is projected to slip. In 2025, 26 percent of expected capacity was delayed; another 10 percent pushed back commercial operation dates without public notice.
This is not a supply chain story. It is a regulatory and community opposition story concentrated at the AI factory tier. General-purpose data centers are not driving moratorium proposals. Enterprises planning AI factory-class infrastructure — or evaluating co-location within facilities that operate at that scale — need to build regulatory risk into their site selection and timeline assumptions in ways that simply did not apply to data center decisions five years ago.
The Build-Buy-Lease-Sell Decision Framework
The structural split between asset classes creates four distinct enterprise decision paths, each with a different risk and capital profile.
Build (AI Factory): Appropriate for enterprises with sustained, proprietary AI workloads that cannot tolerate multi-tenant latency, security, or compliance constraints. Capital requirements are prohibitive for most — hundreds of millions to billions for greenfield AI factory campus development. Timeline risk is severe: construction takes 12 to 18 months under ideal conditions, and approval complexity at primary markets extends that materially. Brownfield retrofits are gaining traction as a speed-to-deployment alternative, but the cooling and power infrastructure requirements still demand significant capital.
Build (General-Purpose): Appropriate for enterprises with stable, predictable workloads and long-term horizon. General-purpose facilities remain the more accessible build option — air cooling, standard power architecture, conventional permitting. The strategic risk is workload evolution: a facility designed for general-purpose compute in 2026 may be inadequate for AI-augmented workloads by 2029.
Buy (Acquire Existing Capacity): M&A activity in data center assets is accelerating as operators seek capital to fund expansion. The AMD-Meta $100 billion agreement for 6 GW of AI capacity illustrates how major enterprises are moving from procurement to partnership or direct acquisition. For enterprises with the capital and strategic intent, acquisition of existing AI factory capacity — or general-purpose capacity with retrofittable infrastructure — avoids permitting risk and compresses timelines. The acquisition underwriting, however, must distinguish which asset class is being acquired. General-purpose assets acquired at AI factory valuations are a capital allocation error.
Lease (Co-location): The most accessible path for most enterprises. Americas colocation preleasing reached 81.5 percent and vacancy sits at 4.2 percent, with no meaningful easing expected before 2030 per Cushman & Wakefield. Supply constraint means lease terms favor operators. Enterprises entering colo negotiations now must distinguish between AI factory-class facilities and general-purpose facilities, and negotiate accordingly. An AI workload placed in a general-purpose colo — even one with a liquid cooling overlay — will underperform. A general-purpose workload placed in an AI factory colo will be overpriced by a significant margin.
Sell (Divest Existing Assets): For enterprises holding owned general-purpose data center assets that do not support their forward workload profile, the current market represents a liquidity window. Operator demand for existing infrastructure — particularly brownfield sites with power access — is elevated. The strategic logic of holding a 5 MW on-premise general-purpose data center when the operational workload is migrating toward managed cloud and AI-augmented processing is weakening.
Workload Classification: The StackedAI Framework
Every infrastructure decision should begin with workload classification, not vendor evaluation. The classification determines the asset class, which determines the relevant operator set, lease structure, pricing benchmark, and risk profile.
Tier 1 — AI Factory Class: Sustained GPU clusters running training, large-scale inference, or simulation workloads. Requirements: >50 kW/rack sustained, direct liquid cooling, non-interruptible power, dedicated or private colo environment. Decision path: AI factory co-location at scale, hyperscaler partnership, or owned build. Operators without direct liquid cooling infrastructure and onsite power architecture are not qualified vendors.
Tier 2 — High-Density Hybrid: Mixed workloads with AI inference components, high-throughput analytics, or HPC clusters alongside standard enterprise applications. Requirements: 20–50 kW/rack in AI zones, hybrid cooling, standard grid power with contingency. Decision path: general-purpose co-location operators with demonstrated liquid cooling capability. Evaluate cooling architecture carefully — liquid cooling as a marketing statement is not equivalent to liquid cooling as designed infrastructure.
Tier 3 — General-Purpose Enterprise: Cloud-connected workloads, enterprise applications, storage, backup, interconnection. Requirements: standard rack density, air cooling, grid power, SLA-based redundancy. Decision path: standard co-location, cloud migration, or edge facilities where latency requirements apply. This is where brownfield and tertiary market operators are gaining relevance — lower cost, less approval friction, adequate for the workload.
For each tier, the lease structure and pricing benchmark differ materially. AI factory capacity is priced by power commitment — $/kW — with long-term contracts and limited flexibility. General-purpose capacity is priced by rack or cabinet with more elastic terms. Enterprises negotiating AI factory contracts using general-purpose mental models will underestimate both cost and commitment depth.
Advisory Takeaway
The structural split between AI factories and general-purpose data centers is not a transitional phase. These are now distinct asset classes with different financial profiles, regulatory exposure, cooling architectures, power strategies, and community relationships. The industry’s trajectory — from the Mag 7’s combined $650 billion in AI infrastructure commitments to the wave of state-level moratorium proposals — confirms the divergence is deepening, not converging.
For enterprise infrastructure teams, the near-term action items are concrete:
Classify before you procure. Run every workload through a density and continuity analysis before engaging operators or brokers. Misclassification is expensive in both directions.
Read the lease structure, not just the rate card. AI factory co-location contracts carry power commitment floors, curtailment provisions, and term lengths that don’t appear in standard colo agreements. Know which contract type you are signing.
Build regulatory timeline risk into AI factory decisions. Primary markets face approval complexity. Any AI factory-class project or co-location in a high-density market should carry a 6 to 12 month schedule buffer for permitting and community review.
Evaluate brownfield and tertiary markets for general-purpose workloads. The colocation vacancy rate of 4.2 percent and 81.5 percent preleasing at primary markets create pricing pressure that tertiary markets can relieve for workloads that do not require hyperscale proximity.
Treat onsite power as a due diligence variable, not a footnote. Whether you are leasing AI factory capacity or evaluating an acquisition, the power architecture — grid-connected, hybrid, or grid-independent — determines cost structure, regulatory exposure, and operational resilience.
The enterprises that get this classification right in 2026 will have infrastructure cost structures and delivery timelines that outcompete peers who treated the two asset classes as variants of the same product. The ones that get it wrong will be renegotiating leases and reconfiguring cooling infrastructure at exactly the moment when AI workload demands are accelerating.
StackedAI provides enterprise data center advisory services across site selection, lease negotiation, operator due diligence, and infrastructure strategy. Contact us at StackedAI.net to discuss your infrastructure classification and procurement strategy.