Power Density: From 5 kW to 130 kW per Rack

Rack power density has roughly tripled in a decade and then broken its own curve in the last two years: from about 5 kW per rack typical in 2015, to a survey average of 7.8 kW in 2026, to 130+ kW for a single NVIDIA GB200 NVL72 cabinet running in production today. The driver is GPU thermal design power, not general IT growth — and it is now the single biggest variable in how data centers are designed, priced, and sold.

Key takeaways

• Fleet-wide average density is still climbing slowly. Uptime Institute’s 2026 survey puts the excluding-outliers average at 7.8 kW per rack, up from 7.5 kW in 2025. The modal figure including high-density facilities tops 11 kW.
• AI racks are on a separate, much steeper curve. A GB200 NVL72 rack draws 130-132 kW measured; the successor GB300 NVL72 lands around 135-142 kW, per TrendForce and integrator data.
• Air cooling caps out around 30-40 kW/rack. Above that, direct-to-chip or immersion liquid cooling becomes mandatory — see our liquid cooling guide for the technology comparison.
• The colocation market has split into two tracks. Legacy racks run 3-7 kW; AI-ready racks run 20-100+ kW, and they are priced, built, and sold differently, per CBRE and Data Center Knowledge 2026 reporting.
• Structural floor loading is now a hard constraint. A populated liquid-cooled AI rack can impose roughly 1,333 kg/m² — several times what a standard office floor or older data hall was engineered for.
• Power distribution, not total facility power, is the newer bottleneck. Stranded power — capacity that exists on the utility bill but can’t reach a specific rack — is emerging as a bigger planning risk than stranded floor space.

Live per-kW benchmark pricing by market is in the colocation price index; the underlying facility data is in the data center catalog.

The density timeline: 2000 to 2026

Era	Typical density	What changed
Late 1990s-early 2000s	1-3 kW/rack	Dot-com colocation, single-purpose servers
Mid-to-late 2000s	5-10 kW/rack	Blade servers, early virtualization (up to 32:1 consolidation)
2010s	8-20 kW/rack	Cloud consolidation shifts load from enterprise to colocation
2015-2020	Up to 30 kW/rack	Hyperscale and HPC clusters push peak densities
2021-2025	7 → 16 kW/rack (average)	AFCOM State of the Data Center Report tracks the average roughly doubling
2026	7.8 kW average / 130+ kW for AI racks	GPU clusters decouple from the general-IT average entirely

Sources: Ramboll historical analysis; AFCOM State of the Data Center Report 2025; Uptime Institute 2026 Global Data Center Survey.

The first four eras moved together — one curve, driven by server consolidation and virtualization ratios. Since 2021, the curve has forked: general enterprise and cloud racks are still climbing gradually, while GPU racks have jumped an order of magnitude in under five years.

What a modern AI rack actually draws

Platform	Power per rack	Cooling	GPUs per rack
Legacy enterprise rack	5-7 kW	Air	N/A
Standard cloud/virtualized rack	8-16 kW	Air	N/A
NVIDIA HGX-class air-cooled server	10-15 kW per server	Air	8
NVIDIA GB200 NVL72	120 kW nominal, 130-132 kW measured	Direct-to-chip (liquid) + air	72
NVIDIA GB300 NVL72	~135-142 kW	Direct-to-chip (liquid) + air	72

The GB200 NVL72 splits its load: roughly 115 kW is removed by the liquid loop and 17 kW by facility air, per TrendForce’s rack-level teardown. That single rack draws more power than 15-20 legacy enterprise racks combined, in the same floor footprint as one.

Why density is climbing

Two separate forces are at work, and conflating them leads to bad planning:

1. General IT density creep. Consolidation, virtualization, and denser server SKUs push the fleet-wide average up a fraction of a kW per year — the move from 7.5 to 7.8 kW/rack over 2025-2026 is this force. It is gradual and predictable.
2. AI cluster density. GPU thermal design power has grown faster than any prior server generation, and NVLink-scale rack architectures (72 GPUs acting as one compute domain) concentrate that power into a single cabinet rather than spreading it across a row. This is a step change, not a trend line, and it is the reason a single rack can now outdraw what an entire small colocation cage used a decade ago.

79% of operators surveyed expect rack density to keep growing, driven overwhelmingly by AI and other high-performance workloads rather than general enterprise IT.

Electrical distribution: the design has to change, not just scale up

A 5 kW rack and a 130 kW rack are not the same engineering problem scaled up — they require different distribution topology:

• Circuit sizing. A 130 kW rack needs multiple high-amperage circuits (often three-phase 400V or higher) rather than the single 208V/30A whip that serves a legacy rack.
• Busway vs. whips. High-density halls increasingly run overhead busway with tap-off boxes sized per rack, since individual whip runs become impractical at this current draw.
• UPS and generator sizing. Facility-level backup capacity has to be engineered against actual concentrated draw per row, not an assumed even distribution — a handful of AI racks can represent a meaningful share of a hall’s total committed power.
• PDU intelligence. Rack-level power distribution units need real-time monitoring at this density; a tripped breaker on a 130 kW rack takes out 72 GPUs at once, not a handful of servers.

Floor loading and structural limits

Weight, not just power and heat, becomes a binding constraint at high density.

Load type	Typical capacity
Standard office floor	300-500 kg/m²
General data center minimum (host room)	~1.0 t/m² (1,000 kg/m²)
China GB 50174-2017 standard (host room)	8 kN/m² (~800 kg/m²) minimum
GB 50174-2017 heavy-equipment zones	12 kN/m²+
High-density facility design target	1.5-2.0 t/m²
Populated liquid-cooled AI rack footprint	~13 kN/m² (~1,333 kg/m²)

A facility built to older 500-750 kg/m² standards cannot simply have AI racks rolled in — the slab, raised floor, or structural grid may need reinforcement before a 600-1,000+ kg cabinet with coolant manifolds and cabling can be sited safely.

The air-cooling ceiling

Air cooling remains reliable up to roughly 15-20 kW per rack with standard hot-aisle containment, and can be pushed to 30-40 kW with aggressive in-row cooling in well-designed halls. Past that point, thermal density outruns what air can remove economically, and direct-to-chip or immersion liquid cooling becomes the default rather than an option — every GB200/GB300-class deployment ships liquid-cooled. The technology tradeoffs between direct-to-chip and immersion, including retrofit cost and PUE impact, are covered in our liquid cooling guide; this piece focuses on the density trend and its facility-design consequences.

The two-track colocation market

Density has split colocation into two products that share a building but not much else:

Attribute	Legacy/standard colocation	High-density AI colocation
Typical rack density	3-7 kW	20-100+ kW
Cooling	Air, CRAC/CRAH	Direct-to-chip or immersion, CDUs required
Power distribution	Standard whips, single-phase	Busway or dedicated three-phase, per-rack monitoring
Floor/structural spec	Standard raised floor	Reinforced slab or floor rating verification required
Contract structure	Per-rack or blended per-kW	Committed capacity, often power-constrained not space-constrained
Market condition (2026)	More available inventory	North American primary-market vacancy near 1.6%, Northern Virginia near 0.7%

Roughly three-quarters of the power capacity currently under construction across North America is already pre-leased before a single cabinet ships, per Data Center Knowledge’s 2026 colocation coverage — a reflection of how power-constrained, not space-constrained, the high-density segment has become. Full market-rate benchmarks by segment and region are in our colocation pricing guide, and pricing for cross-connects and other line items that scale differently in high-density cages is in our cross-connect pricing guide.

Stranded power vs. stranded space

Two failure modes now compete for attention in capacity planning:

• Stranded space: floor area sits empty because the facility’s power or cooling budget for that zone is already committed elsewhere — common in older halls retrofitted for a handful of high-density cages.
• Stranded power: electrical capacity exists in the facility’s total draw but cannot reach a specific rack because PDUs, busway taps, or circuit breakers in that row are already at their limit — increasingly the more common failure mode as density rises unevenly across a hall rather than uniformly.

The practical effect is the same either way: paying for capacity you cannot use. The fix is design-time, not lease-time — provisioning distribution infrastructure for the density you expect to need in year three of a term, not just the density of the rack you are installing on day one.

Planning for future density: what buyers should do

1. State your density requirement in the RFP, not just total kW. A 500 kW deployment at 10 kW/rack and the same 500 kW at 100 kW/rack need entirely different cooling and floor specifications from the operator.
2. Verify structural rating before committing to high-density cages, especially in retrofitted (non-purpose-built) facilities — ask for the documented floor load rating, not just “it’s rated for high density.”
3. Ask how power is distributed to your row, not just what the facility’s total committed capacity is. A facility can have ample total power and still be unable to deliver it to your specific rack position.
4. Budget for liquid cooling infrastructure separately from rack rent if your density will exceed roughly 30-40 kW/rack — this is a distinct line item, not something legacy per-kW colocation pricing already includes.
5. Model growth, not just today’s load. Given that AI cluster density has moved from ~40 kW to 130+ kW per rack in roughly three product generations, a facility that only just meets today’s requirement offers little headroom for a refresh cycle.

For workload-specific compute pricing at these density tiers, see the GPU price tracker; for facility-level specifications by site, the data center catalog lists power and cooling capability per facility.