Colocation Shortage 2026: Where Can You Put AI Servers Now?

Colocation is in short supply in 2026: CBRE put primary-market vacancy in North America at a record 1.4% in the first half of the year, and over 80% of capacity under construction was already preleased. A business that needs a home for AI servers still has options, including retail colocation, high-density suites, on-premises power upgrades, smaller inference hardware and cloud. The right choice depends on how much power you really need. Measure that first, because power rather than floor space is what you are buying.

How bad is the colocation shortage in 2026?

Vacancy is at a record low on CBRE's measure. CBRE's H1 2026 release, published August 27, reported 1.4% vacancy across primary markets and 0.2% in Northern Virginia. Cushman & Wakefield reported 3.8% vacancy for the wider Americas in H1 2026. The two figures use different market definitions, so quote each with its source. Both firms describe the same direction: Cushman reports 91.7% of the 37.7 GW under construction in the Americas as precommitted before delivery. CBRE's market table shows how little is left in some places. These are reported H1 2026 figures for available capacity, and a provider's real answer for your size and date may differ.

• Northern Virginia: 10.8 MW available from 4,496.5 MW of inventory.
• Atlanta: 29.5 MW available from 1,803.2 MW.
• Dallas-Fort Worth: 39.6 MW available from 1,432.8 MW.
• Chicago: 20.0 MW available from 910.6 MW.
• Hillsboro, Oregon: 1.1 MW available from 491.4 MW.

How fast is colocation pricing rising?

Asking rates rose in every size tier in the first half of 2026. CBRE's H1 2026 report lists increases of 4.3% for 250-500 kW, up from 2.5% a year earlier, 7.9% for 500 kW to 3 MW, 8.3% for 3-10 MW and 6.7% for 10 MW and above. CBRE says the momentum has shifted toward smaller tiers, which have more available inventory. For a dollar anchor, CBRE's H2 2025 report gave a record of about $196 per kW per month for 250-500 kW requirements, and its midyear outlook expects that average to exceed $215 in 2026.

Those are asking rates in primary markets, not quotes, and the CBRE series starts at 250 kW. We did not find a reliable public benchmark for smaller retail deployments, so treat any per-rack price you see online with caution and get written quotes. JLL said in January 2026 that it expects Americas lease rates to grow 7% a year through 2030, according to DataCenter News.

Why can't new construction fix the shortage quickly?

Construction is at a record, but most of it is already spoken for. CBRE counted 7,481 MW under construction in H1 2026, with under 1,500 MW available to prelease, roughly six months of supply at the current pace. Power sets the timetable: JLL says average grid connection lead times in primary locations now exceed four years, according to DataCenter News. CBRE also says ready-for-service fit-out dates are stretching from weeks to several months. Local opposition adds risk, as our data center moratoriums update explains.

CBRE's Gordon Dolven, in the August 2026 release: "many occupiers are effectively competing for future inventory rather than existing availability."

How much rack power do AI servers need?

More than most existing racks were built for. The Uptime Institute's 2026 survey, released July 28, reports that the average of modal rack densities passed 11 kW for the first time, or 7.8 kW without a small number of new high-density facilities with typical racks above 30 kW (Uptime Institute). NVIDIA's user guide lists the 8-GPU DGX B300 at 14.5 kW, with a 15 kW system maximum. One such system therefore exceeds the average rack, and four would need up to 60 kW by simple multiplication. OEM builds vary, so use the vendor's power calculator for the exact configuration. Our B300 and GB300 guide covers the rack-scale step up.

How do I size rack power before signing a colocation contract?

Work from the servers up, not from the provider's rack product down. As a rough guide for North American three-phase circuits, usable power is volts times amps times 1.732 times 0.8, where 0.8 is the common continuous-load derating. By that arithmetic, a 208 V 30 A circuit gives about 8.6 kW, 208 V 60 A about 17.3 kW, and 415 V 60 A about 34.5 kW. Confirm the numbers with your provider and electrician.

• Add up maximum system power, not typical draw, for every server, switch and storage shelf.
• With redundant A and B feeds, one feed must carry the full load alone. Do not add the two together.
• Confirm that cooling matches the power per rack, not just the room average.
• Check server depth, weight and floor loading against the cabinet and the room.
• Match plug types and phases with the right rack PDUs.
• Reserve growth in the contract, because adding power later may mean a new cage or a new building.

Where can a business put AI servers in 2026?

• Retail colocation: a cage or a few cabinets suits small deployments, but many standard cabinets will not take a dense GPU node. Ask for the kilowatts per cabinet in writing.
• High-density suites: these suit 8-GPU systems and larger. Check whether the site supports the liquid cooling your hardware needs, using our liquid cooling guide.
• On-premises upgrades: new circuits and cooling in your own building can beat a long colocation wait if your utility service has headroom. Get the electrical survey before you commit.
• Smaller inference hardware: PCIe GPU servers such as our 4-GPU inference configuration or AI workstations can run many inference and development jobs in far less power than an 8-GPU training node.
• Cloud or GPU rental: this suits short projects, uncertain demand and bridging the gap until a site is ready. Our neocloud comparison covers the trade-offs.

What should I ask a colocation provider before signing?

• Is this power energized today, and if not, what contractual date and remedy apply?
• What is the maximum density per cabinet, and is it air or liquid cooled?
• How is power billed: committed kW, metered use or a mix?
• What escalator applies at renewal, and do I have a right to renew? CBRE says tenants of all sizes may seek to renew early.
• What credit checks apply? CBRE says owners are becoming more selective and favoring strong credit profiles.
• Can I expand in the same building, and is that option written down?

When is waiting or renting the better answer?

If you cannot yet say how many GPUs you will keep busy, rent first and buy once utilization is proven. If your only available site is a year away, do not let hardware sit in storage while its warranty may already be running: order against the real ready date and current lead times. If your workload fits a small inference server, you may not need colocation at all. Our build vs buy vs cloud guide walks through that decision.

Nexus Compute can help you fit the hardware to the power you can actually get. Tell us your kW per rack, cooling type and workload, and we will check compatibility against your platform and quote new hardware with the full manufacturer warranty within 48 business hours. Purchase orders and net terms are available for qualified organizations, and we ship worldwide. Request a quote or email sales@nexus-compute.com.