Inference and training · 1 rack to multi-MW

AI Colocation Providers: Who’s Actually Ready for Your Workload (and Who Isn’t)

An AI-ready colocation provider can show usable power at your kW per rack, cooling that is installed today rather than planned, and a minimum you can meet. On our November 2026 price sheet a full 48U rack with 15 kW runs $4,500–6,750 a month and a 30 kW rack $9,000–9,999. High-density and GPU capacity is listed in 28 US states and Ontario, most of it for Q4 2026.

Updated November 2026 · Bob Spiegel, Founder, QuoteColo

Rear of a row of GPU server racks with liquid-cooling hoses, and a coolant distribution unit at the end of the row
Bob Spiegel

“Tell me what you’re deploying; I’ll come back with real options from a network built over 20 years.”

— Bob Spiegel, Founder, QuoteColo

From our desk, November 2026

$4,500–6,750
Full 48U rack, 15 kW, 1 Gbps
$9,000–9,999
Full 48U rack, 30 kW, 1 Gbps
28 states
with listed high-density capacity, plus Ontario
23 of 29
of those states and provinces list air and liquid cooling

From Bob’s November 2026 price sheet and capacity list (listings, not confirmed offers) and the October 2026 price index (50+ provider responses). Prices are per month, space and power, unless noted.

The AI colocation provider market: what you’re actually choosing between

Four types of operators take AI deployments. If you need 2–8 GPU racks soon, a Tier 1 operator in a primary market will rarely be the answer.

Provider typeWho they areWhat to know
Tier 1 hyperscale-adjacentEquinix, Digital Realty, Iron Mountain, QTS, CyrusOne, Vantage Data CentersUnmatched carrier density, but Ashburn and Santa Clara power queues run 6–12 months or more for new deployments.
Tier 2 mid-market regionalMid-market and regional operators, many with Uptime Tier III certified facilitiesA 5–10 rack AI deployment matters to them, and power is often easier to get. Few advertise availability.
Tier 3 specialty / power-firstHydro-powered Pacific Northwest sites, energy-advantaged Texas operators, Midwest wholesale-adjacent sites, purpose-built AI hostsBuilt or repositioned for dense AI, on shorter terms. Sites follow cheap, reliable power, and most publish no pricing.
Tier 4 managed / bare-metalBare-metal GPU rental and managed AI infrastructure, not traditional colocationPriced per node. For comparison, one 8-GPU H200 node on demand lists at about $25,000–62,000 a month running continuously (Crusoe to Azure and Google Cloud list prices, read 4 November 2026); see the NVIDIA page for the providers.

Provider tiers here are market segments, not Uptime Institute Tier ratings.

What “AI-ready” actually means and what to verify before you sign

Five checks show whether a facility can run your hardware at your size, on your date.

AreaWhat to verify
Power: three numbers, not oneCommitted (paid for), allocated (on your breaker) and available kW by your date. “Allocated” can mean six months out.
Cooling: installed vs. “ready”Installed cooling type for your exact system, in writing. “Liquid-ready” often means floor prep, not an installed CDU manifold.
Redundancy: A+B feeds, real generator runtimeDual-feed power and UPS runtime at your power draw, not the facility’s rated maximum.
Minimums and term flexibilityMinimum kW, minimum term and the contract path from 3 racks to 10. Many AI teams are turned away here.
Remote hands rate cardHourly rates, response SLAs and whether NVIDIA-certified staff are on site. Budgets often leave this line out.

Cooling is decided by the system, not the GPU generation: 8-GPU nodes such as DGX B300 are air-cooled in NVIDIA’s documentation, while rack-scale NVL72 systems need direct liquid cooling. Our NVIDIA colocation page covers rack planning by system.

Want these numbers for your own deployment? and we’ll reply with the providers that fit, with their prices.

AI colocation provider types: what each tier actually delivers

Typical footprints, minimums and cooling by provider type. Facilities differ, so we ask providers to confirm them for your hardware.

Provider typeTypical footprintPower minimumCoolingRight for
Tier 1 hyperscale-adjacent10+ racks100 kW–1 MW+Air and liquid optionsLarge enterprise and multi-MW AI. For 1 MW+, plan on 6–12 months.
Tier 2 mid-market regional2–20 racks10–50 kWAir and containment; some liquidStartups and enterprise AI teams that put speed and $/kW ahead of brand
Tier 3 specialty / power-first1–10 racks5–30 kWOften liquid-designedCost-driven training, for teams open to secondary markets
Tier 4 managed / bare-metalPer nodeNoneManaged by the providerTeams that haven’t bought hardware yet, short-term needs

Provider tiers here are market segments, not Uptime Institute Tier ratings. Prices depend on size more than tier. Planning bands for late 2026 and 2027 deployments, space and power: $225–250/kW/month under 100 kW, $200–250 for 100–500 kW and $185–250 for 1 MW and up (QuoteColo price index, October 2026, 50+ provider responses).

Inference colocation vs. training colocation

Inference and training need different things from a facility, so the right provider is often different too.

AI inference colocation

  • Low latency to your users: location matters more than power cost
  • Carrier neutrality and direct cloud on-ramps for hybrid setups
  • 10–20 kW per rack, often on standard air cooling
  • A 99.99% uptime SLA, because downtime is customer-facing
  • Room to grow with your user base

AI training colocation

  • 30–80 kW per rack, with liquid cooling where the hardware needs it
  • High-bandwidth node interconnect: InfiniBand, RoCE or 100G+ Ethernet
  • $/kWh is a primary cost at scale
  • Location matters less, since training doesn’t need low user latency
  • Jobs can checkpoint and resume

Fine-tuning and batch inference sit in between: 15–30 kW per rack, air or containment cooling and a moderate network. Tier 2 regional operators are usually the best fit.

The AI teams QuoteColo matches to providers

Five common starting points, and what we send back for each.

AI team profileTheir situationWhat we deliver
AI startup, 1–3 racks, own hardwareBought H100s, then hit a 100 kW minimum. Needs a home for a 15–30 kW rack without a multi-year contract.Mid-tier and regional providers that take 1–3 rack AI deployments on 12-month terms, with cooling that fits the hardware.
Enterprise ML team, 5–20 racksMoving AI inference off cloud. Needs a carrier-neutral facility with cloud on-ramps, predictable MRC and room to grow.Tier 1 and Tier 2 options compared like for like: power model, bandwidth, cross-connects and term escalators.
AI training operation, power-firstCost per kWh is the decision driver. Willing to go to secondary markets for a lower rate per kW.Power-advantaged Tier 3 providers in Texas, the Pacific Northwest and the Midwest, most not findable through search.
Series B–C AI company, scaling fastStarted with 2 racks and needs to plan for 20 in 18 months, without migrating facilities mid-growth.Providers with room to expand and expansion terms written into the contract.
Enterprise IT, regulated AI workloadAI inference for a healthcare or fintech application. Needs a current SOC 2 Type II report, a BAA where the provider’s role calls for one, and documented physical security alongside the power density.Providers with current SOC 2 Type II reports that also support the required power density.

Why finding the right AI colocation provider is harder than it looks

Every operator has a GPU landing page now. Few will tell you about usable kW, installed cooling and minimums up front.

THE DIY WAY

You search “AI colocation providers”

The same six names come up, all with “AI-ready” pages. Many want 100 kW minimums in markets with 6–12 month power queues. Others mention provisioning backlogs only after you sign the LOI.

THE DIY WAY

You ask ChatGPT or Perplexity

You get the same well-known providers and no availability data. Mid-tier and regional operators that take 2–10 rack AI deployments often don’t publish specs, so AI tools don’t see them.

THE DIY WAY

You contact providers directly

You spend four weeks on calls. Providers go quiet, quote hyperscaler minimums or claim “liquid-ready” infrastructure that turns out to be a floor drain.

How it works

You send the specs. We find the options. You decide.

STEP 1

Send your requirements.

Tell us what you're deploying, where you want it, and when you need it. Send the specs you have; we'll help with the rest.

HardwareRacks & kWLocationStart date

Estimates are fine. No perfect RFP needed.

STEP 3

Take the next step. On your terms.

Like the options? Contact the providers directly. Need more detail? Call or email us. We're here to help, not push you.

  • Go directly to the providers you like
  • Ask us questions whenever you need

Get AI Colocation Provider Options

Free for buyers

Your price is the same as going direct.

All-in estimates by email

From 500+ providers. No sales calls, no long forms.

Trusted service since 2004

5,000+ clients helped.

Your request will be sent securely to QuoteColo.

The AI colocation provider market going into 2027

Power is the real constraint, not rack space.

In Ashburn and Santa Clara, facilities have cabinets but don’t always have power on your timeline. We ask providers to confirm power availability before they make your shortlist.

The best options aren’t on Google or ChatGPT.

Tier 2 and Tier 3 providers that take 2–10 racks, offer 12-month terms and have capacity now rarely advertise it. They fill space through referrals, not search.

“AI-ready” means nothing without verification.

Since 2024 it’s the default claim. It takes installed cooling for your system, A+B power at your rack density and minimums that fit. We ask providers to confirm all three.

Plan B markets win on speed going into 2027.

Secondary markets such as Salt Lake City, Columbus and Reno, plus primary markets with heavy new construction such as Dallas, Phoenix and Atlanta, take deployments primary markets can’t place in time. For training and batch inference, latency rarely matters.

Case studies

See how we helped teams secure colocation with the right power, pricing, and providers.

Leonard Roszkowski

CEO Elemental Energy Solutions

Problem: Needed to source ASIC servers domestically. Skeptical about working with a new organization.

What we did: Connected them to buyers, sellers, and infrastructure partners across deployments. Removed procurement and logistics overhead entirely.

Result: Rapid organizational growth in a short period. Now uses QuoteColo as an ongoing procurement interface across hardware and infrastructure.

Noel Frutasio

CEO FSN Services and Communications

Problem: Regularly needed colocation space across multiple US states and time zones. Difficult to navigate the market without deep industry knowledge.

What we did: Defined requirements in one meeting, then delivered proposals with multiple carrier and data center options across every target market.

Result: Placed in 4 cities. Contracts signed quickly. Multi-state colocation needs resolved through a single point of contact.

Tim Mandell

DC Mining / MDI Access

Problem: Needed reliable options for crypto mining containers and equipment, along with industry partners who could support real deployment requirements.

What we did: Connected the project with relevant mining infrastructure and equipment resources through the QuoteColo network, simplifying the search and evaluation process.

Result: Gained a clearer path to suitable mining infrastructure, equipment and industry contacts without having to source every option independently.

5,000+

Clients helped since 2004

No Cost

to Buyers

500+

Worldwide providers

Quotes

in 48 h

FAQ: AI colocation providers

What makes a data center actually AI-ready vs. just claiming it?

Four things: confirmed power delivery at the density your GPUs require, available by your deployment date; cooling that supports your exact hardware and kW per rack (air, rear-door heat exchangers or direct-to-chip liquid, depending on the system); A+B dual-feed redundancy; and minimum commitments that fit your footprint. A real AI-ready facility has all four. A GPU landing page alone means nothing.

How do I find AI colocation providers that accept small deployments (1–5 racks)?

Tier 1 operators like Equinix and Digital Realty often require 100 kW+ minimums in primary markets. Providers willing to support 1–5 rack AI deployments are usually Tier 2 regional operators or Tier 3 specialty sites.

Most do not publish pricing or live availability publicly. The fastest path is usually a referral service with active provider relationships and current inventory.

What’s the difference between GPU colocation providers and AI colocation providers?

They overlap in practice. “GPU colocation” is about hosting the hardware itself, such as H100, A100 or L40S clusters, in a data center. “AI colocation” covers the full stack an AI workload needs: power density, distributed networking, cooling, compliance, room to scale and deployment planning. For the hardware side, see our GPU colocation page.

How long does it take to deploy AI infrastructure in a colocation facility?

For 1 MW and up, plan on 6–12 months, and 2027 blocks are being reserved now. In Ashburn and Santa Clara, power queues for new deployments can run 6–12 months or more. Smaller deployments where power is already available, often in Plan B markets such as Salt Lake City, Columbus, Dallas or Phoenix, can go live much sooner. Liquid-cooled systems add CDU and manifold installation and commissioning, so ask each provider for the power date and the install date separately.

Is AI colocation cheaper than AWS or Azure for GPU workloads?

For steady, always-on workloads, usually yes: on-demand H200 lists at $4.29–10.60 per GPU-hour (Crusoe to Azure and Google Cloud list prices, read 4 November 2026), roughly $300K–740K a year for an 8-GPU node running continuously. Owning the hardware and colocating it usually wins over time. Cloud still makes sense for bursty or short-term workloads where flexibility matters more than long-term economics.

What should I ask an AI colocation provider before signing?

The questions that matter most:

1) What is the usable kW per rack?
2) What cooling infrastructure is actually installed?
3) Is power available now or queued?
4) What are the minimum kW and contract commitments?
5) What does the remote hands rate card look like for GPU-specific work?
6) Is liquid cooling operational or only “liquid-ready”?
7) Can the deployment scale from 3 racks to 15 without moving facilities?

Which US markets have the best availability for AI colocation right now?

Going into 2027, Dallas/North Texas, Phoenix, Salt Lake City, Columbus, Atlanta and parts of the Pacific Northwest offer the best mix of power availability and realistic deployment timelines. Ashburn and Santa Clara still have stronger carrier density, but sub-100 kW deployments there often struggle with long queues and aggressive minimum commitments. Our secondary markets guide compares the secondary options.

How much does AI colocation cost per rack per month, all-in?

For late 2026 and 2027 deployments, QuoteColo’s price index (October 2026, 50+ provider responses) puts space and power at $225–250 per kW per month under 100 kW, $200–250 for 100–500 kW and $185–250 for 1 MW and up. These are planning bands, not market averages. Bandwidth, cross-connects, IPs, setup and remote hands can be extra. At $225–250/kW, a 10–20 kW GPU rack comes to about $2,250–5,000 a month for space and power, before bandwidth and cross-connects. See our NVIDIA colocation page for rack planning by system.

Can QuoteColo help with AI colocation outside the US?

Our primary focus is the US and Canada, where we maintain the deepest provider relationships and market visibility.

We also support selected international markets for clients expanding beyond North America. Share your target region and deployment goals, and we’ll tell you directly what coverage is realistic.

Get 2–4 qualified options in 48 hours

Free for buyers

Your price is the same as going direct.

You will get all-in price estimates by email

From 500+ providers. No sales calls, no long forms.

Trusted service since 2004

5,000+ clients helped.

Your request will be sent securely to QuoteColo.