1.4%
Vacancy at a record low
North America overall; Northern Virginia near 0.3%. About 92% of capacity under construction is pre-leased.
CBRE Q1 2026
Big names turned you down?
Get 2–4 GPU colocation options for AI and HPC in 48 hours — power and cooling confirmed.
From one GPU server to 100+ kW per rack, air or liquid.
Free to you. Providers pay our fee from their existing sales budget.
Half the facilities that say liquid-ready mean it's on the data sheet, not in the floor. I check whether the CDUs are installed before anything reaches your shortlist.
5,000+
Clients helped since 2004
500+
Worldwide providers
Free
to Buyers
Quotes
in 48 h
Recently placed
8× H200 NVL, 18 kW
single rack, air-cooled, 12-month term
300 kW cluster
10 racks × 30 kW, Colorado secondary market
B200 at 80 kW/rack
DLC, CDU manifolds confirmed, 60–90 day deploy
15 kW inference rack
true metered per-kWh billing, IPMI/OOB
8–16 racks, 20–35 kW
phased enterprise AI, SOC 2 Type II
GPU and AI colocation pricing by market
| Market | $/kW/month | Density | Notes |
|---|---|---|---|
| Colorado (secondary) | $78–$140 | 10–30 kW | Recent 300 kW quote at $78/kW base. Up to 1 MW from some regional operators |
| North / South Dakota | $100–$170 | 20–60 kW | Lowest US power cost. Limited carrier depth |
| Pacific Northwest | $110–$200 | 15–50 kW | Cheap hydro. Washington at the lower end |
| Columbus / Ohio | $130–$210 | 10–30 kW | Cloud adjacency, lower costs |
| Phoenix / Arizona | $130–$220 | 15–40 kW | Good power availability. The common California Plan B |
| Dallas / Texas | $140–$240 | 10–40 kW | Best cost-to-availability balance in the US right now |
| Atlanta | $140–$230 | 10–25 kW | Growing secondary hub. ~1.0% vacancy |
| Chicago | $160–$260 | 10–25 kW | Strong carrier-neutral ecosystem. ~2.2% vacancy |
| New Jersey (NYC metro) | $170–$280 | 8–20 kW | Low latency to NYC financial sector, expensive power |
| Northern Virginia | $180–$300 | 10–30 kW | Tight power, 6–12 month queue. ~0.3% vacancy |
| Silicon Valley / Bay Area | $200–$350 | 10–25 kW | Very limited capacity. Consider Phoenix or PNW instead |
Cheapest to most expensive is roughly 4.5×, at comparable density. Power availability, not rack space, is what sets the spread.
Prices and availability change every month.
$/kW/month = colocation MRC only. Power billing model, cross-connects and remote hands typically add 25–30% to the real monthly cost (full range 20–60% depending on connectivity requirements). Request a quote to get the most up-to-date all-in price options.
Typical requests we handle
These come from real inquiries. If yours looks like one of them, you're in the right place.
| Request | Technical detail |
|---|---|
| "We bought 8× H200s. A major provider said no. Who will take us?" | 1 rack, 12–18 kW peak (~8 kW sustained), H200 NVL or HGX. Air-cooled, 12-month term. Rejected by 2+ Tier 1 operators citing 100 kW minimums. |
| 300 kW AI cluster, Colorado / secondary market | 10 racks × 30 kW. N+1 power (dual A/B, 208V, 3-phase). 100G uplink + 2× 10G cross-connects. Committed kW billing. 36-month term. |
| "Deploying B200/GB200. Need 80 kW+ per rack, DLC required." | 2–4 racks, 80–100+ kW/rack. 800G InfiniBand or RoCEv2. CDU manifolds confirmed installed. Timeline 60–90 days. |
| AI inference, idles 99% of the month, spikes to 15 kW | 1 rack, 10–15 kW peak / 2–3 kW average. True metered per-kWh requested. Consumer or prosumer cards. IPMI/OOB access required. |
| Multi-GPU HPC analytics box, ~10 kW, Oregon, no local staff | 1–2U custom chassis, ~10 kW peak. Air-cooled. Remote hands required. Pallet ship-in. 10G uplink. |
| Enterprise AI: 8–16 racks, 20–35 kW/rack, hybrid cloud | 160–560 kW total, phased. 100G, AWS Direct Connect or Azure ExpressRoute. SOC 2 Type II. A/B feeds + N+1 cooling. |
Don't see yours exactly? Send your specs. We've placed single nodes, consumer cards, Blackwell clusters, and everything between.
Why most teams struggle to find GPU colocation
Search and AI answers
Generic provider names, no real pricing, no capacity data. Most replies lead to a "schedule a call" form. Facilities that take a 1–2 rack GPU deployment often don't appear at all — they fill through referrals.
Directories
Outdated inventory, no verified power availability, no indication of who accepts sub-100 kW. Your contact details land in a general sales queue.
Consultants / formal RFP
Right for a 50+ MW multi-market procurement. Usually unnecessary for a single-rack H200 cluster with a 45-day deadline.
QuoteColo
Referral networks are built around operators who pay to be in them, so regional facilities are underrepresented by design. Send requirements once; we filter 500+ providers including off-market operators and return matched options.
Your price is the same as going direct.
Send us your specs
We will provide options with real pricing, confirmed power and lead times.
No company details needed to start.
Some operators demand full qualification before they'll talk; others evaluate on technical requirements first. We know which is which.
Email first.
No sales call until you've reviewed the options.
48 hours
Turnaround time.
2026 GPU Colocation Market
+62%
AI leasing year over year
Inference is now the majority of AI compute demand, increasingly in colocation rather than public cloud.
$196/kW
Wholesale asking price
For 250–500 kW deployments, up roughly 6.5% year over year.
CBRE
Power
Not space, is the bottleneck
The most competitive deals close off-market before capacity appears in any public listing.
40–100+ kW
Blackwell readiness is the new bar
B200 and GB300 clusters need liquid cooling most legacy facilities can't provide without significant upgrades.
69%
Reconsidering cloud-first
More GPU workloads moving from hyperscale cloud into dedicated colocation.
Broadcom 2026
What cooling and power does your deployment actually need?
Find your hardware. The cooling column is what disqualifies most facilities.
| Deployment | Generation | Rack power | Cooling | What we match to |
|---|---|---|---|---|
| 1× GPU server (RTX 6000 Pro, 4090/5090) | — | 0.8–7 kW | Air | Facilities that accept single nodes and consumer cards |
| 8× H100 / H200 / HGX node | 2023–25 | 12–20 kW | Air or RDHx | HD retail colo, 1–2 cabinets, air to ~20 kW with containment |
| DGX B200 / HGX B300 rack | 2025–26 | 20–40 kW | Liquid preferred (DLC) | Liquid-ready cabinets with installed CDU manifolds, metered power |
| GB200 NVL72 full system | 2025–26 | 80–100+ kW | Liquid (integrated) | Wholesale and powered-shell providers; purpose-built AI facilities |
| Multi-rack B300 / GB300 cluster | 2026+ | 100 kW+ | Liquid (DLC or immersion) | MW-scale — see MW Infrastructure for powered land and build-to-suit |
Average AI rack density is now around 50 kW. Air handles ~25–30 kW reliably; direct-to-chip is the default expectation above that. "Liquid-ready on the sales deck" doesn't qualify — ask whether the CDU manifolds are installed.
The rack rate is not the bill
Budget 25–30% above the quoted rate. A recent 300 kW Colorado quote came in at $78/kW — here's what got added.
| Cost component | Colorado 300 kW example | What to know |
|---|---|---|
| Base rack rate (MRC) | ~$78/kW/month | The headline number. Lower than primary markets at equivalent density |
| Cross-connects (2 × 10G) | +$700/month | $350 each, quoted separately from the rack rate |
| Power billing model | Committed kW | Idle inference boxes still billed at the full committed rate |
| Remote hands (after-hours) | $150/hr | Standard requests included; extended and after-hours billed separately |
| 12- vs 36-month term | ~12% premium | 36-month is the base rate |
| Real bill vs base rate | +25–30% | The typical gap between quoted rack rate and month-one cost |
The number that matters when comparing providers is MRC after power model, cross-connects and remote hands — not the headline $/kW.
Don't need a Tier 3 badge?
Tier III costs more because of the redundancy architecture. If your workload tolerates brief maintenance windows and you're not in a regulated environment, we'll show you what comparable infrastructure costs without the certification markup — so your margins actually work.
Most people who call me have already been turned down twice. They think the problem is that their deployment is too small. Usually they asked the wrong facilities.
Tell us what you're deploying. We know who takes it.
Specs go to operators that take your density and have the power free.
No company details needed to start.
Technical requirements are enough to begin.
Off-market included.
Regional operators that fill through referrals rather than search.
Real monthly cost, not just the rack rate.
Quotes normalised after power model, cross-connects and remote hands.
Back in 48 hours. A real person reviews it, not an auto-response.
GPU, AI and HPC colocation: FAQ
Will they accept consumer GPUs (RTX 4090, 5090, 6000 Pro)?
Many won't, and they won't tell you upfront — some have AUPs that exclude consumer cards, others have infrastructure built only for server-grade hardware. We know which facilities accept them, so you don't spend a week getting rejected after the intake form. Say so in your specs and we route accordingly.
Air or liquid — and does it matter at my density?
Air with containment or rear-door heat exchangers handles up to ~25–30 kW per rack reliably. Above that you need liquid. B200 and GB300 clusters typically require direct-to-chip at 40–80+ kW. Give us the GPU model and quantity and we'll tell you what cooling you actually need.
Do I have to pay for power I'm not using?
Committed kW is the default — you reserve an amount and pay for it whether you use it or not. For inference that idles most of the month and spikes occasionally, that's a significant overpay. We specifically look for facilities offering true metered per-kWh billing. They exist; they're just not the majority.
I only need one server or one rack. Do I qualify?
Yes. Most Tier 1 operators gate small deployments behind 50–100 kW minimums. We route to mid-tier and regional facilities that accept single-node and 1–2 cabinet deployments. Those providers usually don't publish available capacity — they fill through referrals, which is the inventory we access.
Can you support Blackwell — B200, GB200 NVL72, B300, GB300?
We match on what the hardware actually requires. GB200 NVL72 is a single-rack system at ~100+ kW with integrated liquid cooling — very few facilities outside purpose-built AI campuses can host it. When you give us the SKU we check usable kW at rack level rather than breaker nameplate, CDU manifold installation status, and 800G availability before shortlisting.
What are the lead times?
Two to six weeks for air-cooled single-rack inference in a secondary market with available power. Three to six months for primary-market liquid-cooled deployments in power-constrained metros. Colorado, Atlanta, Phoenix and Ohio typically run four to eight weeks. When we send a quote, the power is confirmed rather than estimated.
How does the cost compare to cloud?
An 8×H100 node on AWS P5 runs about $3.90/hr per GPU — roughly $22,500/month at high utilisation. A comparable colocation deployment typically runs a fraction of that all-in, with hardware break-even usually around 12–15 months.