Stefan Knight/AI Data Center Capacity: Utilization vs. Planned Growth
AI NowAI Created

AI Data Center Capacity: Utilization vs. Planned Growth

The current landscape of AI data center capacity is defined by a massive gap between investment and utilization. While hundreds of billions are being poured into planned capacity, the actual usage—especially in terms of inference—remains a fraction of the total footprint.

1. The Utilization Gap: Training vs. Inference

Currently, the vast majority of AI-specific capacity is dedicated to training rather than inference.

Training Heavy: For major labs, training a frontier model can consume up to 70–80% of their total allocated compute budget over a year. This is "burst" usage—massive consumption for a few months, followed by lower maintenance levels.

The Inference Lag: While consumer AI tools have millions of users, the actual "inference" (processing a user's prompt) currently uses far less energy and hardware than the initial training phase. Estimates suggest inference accounts for roughly 20–30% of total AI compute spend.

2. Built vs. Planned: The "Ghost" Capacity

There is a significant lag between a data center being "announced" and it actually drawing power from the grid.

Occupancy Rates: In Tier 1 markets, data center occupancy is high (often over 90%), but much of this still supports traditional cloud workloads. AI-specific workloads currently represent about 15% of total data center demand, though this is projected to double by 2027.

Idle Infrastructure: In certain regions, aggressive state-led building has led to "idle" data centers. Reports indicate that while massive capacity has been built, utilization rates can struggle to hit efficiency targets due to network latency and fragmented interoperability.

3. Energy Intensity per Rack

Usage is measured not just by physical space, but by power draw.

High-Density Racks: Standard data center racks historically pulled 5–10 kW. Modern AI racks (housing H100s or B200s) pull 40–100 kW each.

The "Stranded Power" Problem: Many data centers are "space-full but power-empty." A building might hit its total electrical limit (e.g., 50 MW) while the floor is only half-filled with AI chips.

Capacity and Consumption Comparison

Metric Current State (2024–2025) Projected (2030)
Total Global Capacity ~415 TWh ~945–1,193 TWh
AI's Share of Demand 10–15% 35–50%
Primary Use Case Model Training Model Inference
Utilization Efficiency Variable by Region Expected 90%+ Optimisation

Why the Gap Exists

Speculative Building: Hyperscalers are building ahead of the curve to ensure they have a buffer of compute for future breakthroughs.

Grid Bottlenecks: It can take 2–5 years to get a full power connection from utilities, leaving buildings only partially operational.

Efficiency Gains: Newer chips are more efficient per calculation, allowing "usage" (AI work) to increase even if total "consumption" (electricity) stays flatter.

AI Data Center Capacity: Utilization vs. Planned Growth · Stefan Knight