Data Center Colocation for AI Workloads: A Buyer's Guide

AI workloads have fundamentally changed the colocation market. A single rack of H100 GPUs can draw 40-70 kW — compared to 5-8 kW for traditional servers. This power density creates challenges in cooling, capacity, and pricing that didn't exist three years ago. Here's what you need to know before signing a colocation contract for AI infrastructure.

Power Density: The Defining Constraint

Traditional colocation was sold by the cabinet at 5-10 kW per rack. AI workloads need 30-70+ kW per rack — and the next generation of GPUs will push this even higher. Most existing colocation facilities simply cannot support this density without significant retrofitting.

When evaluating providers, ask specific questions: What is the maximum power density per cabinet in their AI-ready zones? Is this available power, or theoretical capacity that requires build-out? What's the utility power feed capacity to the building, and how much headroom exists for expansion?

Don't accept marketing claims. Ask for the utility bill or the power distribution architecture diagram. Facilities that claim unlimited power density in existing space are either oversubscribed or haven't done the math.

Cooling Architecture: Air vs. Liquid vs. Immersion

At 30+ kW per rack, air cooling becomes physically inadequate. The three options:

Rear-door heat exchangers (RDHx): Hot water loops mounted on the back of each cabinet. Handles up to 40-50 kW per rack with minimal facility modification. This is the most common solution for current-generation GPU deployments.

Direct-to-chip liquid cooling: Cold plates mounted directly on GPU modules, connected to a facility water loop. More efficient than RDHx, supports higher densities, but requires servers designed for liquid cooling (not all GPU servers support this).

Immersion cooling: Servers submerged in dielectric fluid. Handles the highest densities and is the most energy-efficient, but requires specialized tanks and is incompatible with standard racks. Best for purpose-built AI training clusters.

Your cooling choice should match your GPU hardware. NVIDIA's reference designs increasingly assume liquid cooling — ask your hardware vendor what cooling they recommend before committing to a colocation facility.

Capacity Timeline: Start 12-18 Months Early

The AI infrastructure boom has created severe capacity constraints in major colocation markets. In Northern Virginia, Dallas, Phoenix, and Chicago — the largest US markets — wait times for new high-density deployments can exceed 18 months.

If you need AI colocation capacity, start the procurement process 12-18 months before you need it deployed. This timeline accounts for: market evaluation and site selection (2-3 months), contract negotiation (1-2 months), facility build-out or retrofit (6-12 months), and hardware procurement and installation (2-4 months).

Secondary markets (Salt Lake City, Columbus, Reno, Atlanta) may have shorter timelines but fewer interconnection options. The trade-off between capacity availability and network connectivity is real.

Contract Negotiation: What You Can Actually Negotiate

Colocation contracts are negotiable, especially for larger deployments (100+ kW). The key levers:

Term length vs. rate: Longer commitments (3-5 years) get better pricing. But AI infrastructure evolves rapidly — a 5-year commitment to a facility that can't support next-gen GPUs is expensive even at a discount.

Expansion rights: Secure options on additional capacity at predetermined pricing. In a tight market, the right to expand at today's rates is extremely valuable.

Power pricing structure: Understand whether power is billed at metered usage, committed capacity, or a blend. Metered usage is cheaper during ramp-up; committed capacity provides cost certainty at full load.

SLA terms: Standard SLAs cover uptime, temperature, and humidity. For AI workloads, also negotiate SLAs around power quality (voltage stability matters for GPU clusters) and network latency to your cloud on-ramps.