Inference vs. Training Data Centers: Where AI Demand Actually Lands

The Thesis

Two asset classes are emerging — and only one of them is scarce where users live

The market still says "data center" as if it were one product. It isn't anymore. AI training and AI inference are diverging into two distinct asset classes with different economics, different siting logic, and different tenants — and understanding the split is the single most useful lens for anyone allocating capital or capacity in this cycle.

Training campuses are gigawatt-scale, remote, and single-tenant. They chase the cheapest available power at almost any location, because a training run doesn't care about latency. Demand comes from a handful of frontier labs, growing at roughly 13% annually.

Inference sites are the opposite: metro-adjacent, latency-sensitive facilities in the 20–100 MW range, where demand scales with every user, agent, and query. Inference has to live near population — which means it competes for power in exactly the places power is hardest to get.

The Numbers

Inference overtakes training as the dominant workload

~35%

Inference demand CAGR

2026–2032, versus roughly 13% for training and traditional hyperscale workloads.

90+ GW

Global inference demand by 2030

More than half of all AI compute, per McKinsey's modeling of workload demand.

2028

The crossover year

Modeled point where inference passes training — rising from ~40% of AI demand in 2026 to ~67% by 2032.

Industry forecasts put roughly 70% of all data center demand on the inference side by 2030 as adoption and agentic workloads scale. Total AI workload demand grows from ~44 GW in 2025 to ~156 GW in 2030 in McKinsey's continued-momentum scenario — with inference driving the majority of incremental capacity.

Implications

What the divergence means for tenants and investors

For capacity buyers: the constraint isn't total megawatts — it's megawatts near users, deliverable on real timelines. Power-rich secondary markets in the Southeast and Midwest are where time-to-power still exists.
For investors: training exposure is concentrated in a handful of counterparties; inference exposure diversifies with adoption itself. Mid-size standardized campuses match the demand curve that's actually compounding.
For both: the 20–100 MW block — modular, liquid-cooling-ready, credit-leased — is emerging as the unit of the inference era. That is precisely the product Phoenix builds.

Phoenix's platform — ten standardized 52 MW campuses across power-rich Southeast and Midwest markets — is purpose-built for this curve: power secured first, campuses sized to inference demand, leased long-term to credit-quality tenants.

Sources: McKinsey & Company, AI data center workload demand analyses (2025); Goldman Sachs Research; Epoch AI; JLL North America Year-End 2025; Knight Frank Data Centres Global Forecast 2026; CBRE North America Data Center Trends H1 2025. Figures are third-party estimates; forecasts are inherently uncertain.

Inference vs. training: where AI demand actually lands