TL;DR
Most IT and facilities leaders treat their colocation or data center contract as a solved problem once the racks are installed and the cross-connects are live. But the contract was almost certainly priced against yesterday’s power assumptions — a world of 5-10 kW racks — while AI and GPU workloads are pushing real deployments to 20-100+ kW per rack. If your agreement doesn’t spell out expansion rights, overage pricing, and cooling upgrade paths, the moment your roadmap needs more power than the rack was provisioned for, you’re not negotiating from strength — you’re negotiating from a forced relocation, with your dedicated fiber and cross-connects stranded at the old address.
The Status Quo Trap: “We have colocation space, we’re set for growth”
Ask most IT directors when they last re-read the power provisions of their colocation contract, and the honest answer is: never. The rack was sized for the workload at signing, the cross-connects were provisioned, the dedicated fiber was lit, and the assumption since has been that “we have space” is the same thing as “we have room to grow.” That’s the trap. A colocation contract isn’t a growth plan — it’s a snapshot of power and cooling assumptions frozen at the moment of signature, and those assumptions are aging faster than almost any other line item in the IT budget.
The ground has shifted underneath that snapshot. AFCOM’s 2025 State of the Data Center report puts average rack density at approximately 16 kW, up from roughly 7 kW in 2021. That’s already more than double in four years — and the curve is steepening, not flattening. Average rack density increased from approximately 16 kW in 2025 to 27 kW in 2026, while only one in five operators report being prepared to support the 50–70 kW racks now common in AI deployments, with forecasts putting average densities near 40 kW within a few years and the newest AI systems reaching up to 246 kW per rack. A contract signed even two years ago, sized for a “standard” enterprise workload, is now competing for power and cooling headroom against tenants running GPU clusters an order of magnitude denser.
The market has responded by splitting in two. The colocation market has split into two distinct tracks: standard colocation handling traditional enterprise compute at 3-10 kW per rack, and high-density AI colocation purpose-built for GPU clusters at 20-100 kW or more per rack, requiring liquid cooling and dedicated high-capacity power distribution. If your contract and your facility sit in the first track and your roadmap is quietly drifting toward the second, you don’t find out until you ask for the upgrade — and by then, the leverage has already shifted to the provider.
The Tele Data Guru Framework: The Power Density Runway Matrix
Before you renew, expand, or sign a new colocation or dedicated fiber agreement, score the facility — not just the rate card — against the four variables that determine whether it can carry your compute roadmap without a forced migration:
| Facility Type | Typical Rack Density Supported | Liquid Cooling Readiness | Expansion Headroom | Cross-Connect / Fiber Portability Risk |
|---|---|---|---|---|
| Standard Enterprise Colo (legacy air-cooled) | 3–10 kW per rack | None — air-cooled aisles only | Low — power distribution not built for a step-change | High if you outgrow it — full re-termination required elsewhere |
| High-Density Retrofit Colo | 20–50 kW per rack | Partial — in-row or rear-door units added post-build | Moderate — contingent on the facility’s grid allocation | Moderate — some cross-connects portable within the same campus |
| Purpose-Built AI/GPU Colo | 20–100+ kW per rack | Native — direct-to-chip or immersion designed in | High — but pre-lease timelines can run 12+ months | Low if you land here first — no future forced move |
| Hyperscale Wholesale / Edge Pod | Custom-provisioned to workload | Native, workload-specific | Highest — but minimum commitment sizes exclude most mid-market buyers | Low, but term length and MW commitments limit flexibility |
Most mid-market buyers default to whichever facility had space available when they needed it — not the one that scores highest against where their compute roadmap is actually headed. Run the matrix against your 24-month plan, not your current rack count, before you sign anything.
The Power Ceiling Exposure Formula
Power Ceiling Exposure = [(Target Rack Density − Contracted Power Allocation per Rack) × Racks Requiring Upgrade × Provider Overage Rate ($/kW/month) × 12] + One-Time Relocation Cost (Fiber Re-Termination + Cross-Connect Re-Provisioning + Cutover Downtime)
Example: a company has 20 racks contracted at 6 kW each in a standard enterprise facility. Its AI roadmap requires 8 of those racks to run GPU inference at 40 kW each. Using an illustrative overage rate at the low end of the current primary-market range — colocation pricing in 2026 ranges from roughly $180 to $400 per kilowatt per month in primary US markets, and $130 to $250 in secondary markets — the math looks like this: (40 kW − 6 kW) × 8 racks × $180/kW/month × 12 = $587,520 per year in overage billing alone, assuming the facility can even physically deliver that power. If it can’t, the relocation cost — re-terminating dedicated fiber, re-provisioning cross-connects, and absorbing cutover downtime — gets added on top, and it’s the number that turns a “power upgrade request” into a full facility migration project.
Commercial Realities & Vendor Pitfalls
- “Per-rack” pricing hides the real constraint. Colocation vacancy hit 1% in 2026 and pricing is now per-kW, not per-rack. If your contract or quote is still denominated in rack units rather than committed kW, you don’t actually know what you’re buying — and neither does the sales rep who quoted it.
- Space isn’t the bottleneck anymore — power is. Vacancy in primary North American colocation markets declined to a record low of 1.4% at year-end 2025, and the primary constraint on new supply is no longer construction timelines but power grid access. A facility can show open floor space and still have no power left to sell you.
- The clauses that actually determine your total cost aren’t the headline rate. Committed power, power overages, reserved capacity, escalation clauses, cooling charges, and expansion rights can materially affect the total cost of the agreement. Two facilities quoting the same per-kW rate can produce wildly different five-year outcomes once these terms are read closely.
- High-density capacity carries a real premium, but not always where you’d expect. GPU colocation costs 30-50% more than standard colocation due to higher power density (20-50kW per rack vs 5-10kW), liquid cooling requirements, enhanced security for high-value equipment, and specialized engineering support. Budgeting for a “like-for-like” density upgrade at your current facility’s standard rate is how the overage bill becomes a surprise.
- Most operators are already behind the curve you’re planning against. Only a fraction of facilities report being ready for the density levels AI workloads now demand, which means “the provider will figure it out” is not a plan — it’s a bet on someone else’s capital expenditure timeline.
Implementation Checklist
- Map your 24-month compute roadmap to a per-rack kW forecast — not a vague “we might need more space” estimate.
- Pull your current contract’s committed power allocation, overage rate, and escalation clauses in writing; if they’re not itemized, ask for an amendment before renewal.
- Confirm your facility’s liquid cooling retrofit path and timeline, not just whether it’s “on the roadmap.”
- Negotiate expansion rights and lock overage pricing now, while you still have renewal leverage — not after you’ve already outgrown the allocation.
- Audit the portability of your dedicated fiber and cross-connects: what does a forced move actually cost in re-termination fees and cutover downtime?
- Score every renewal or new-build option against the Power Density Runway Matrix before you sign, using your roadmap density — not your current density — as the test case.