Ask a colocation operator or a utility planner what limits AI capacity in 2026 and the answer is rarely "we cannot get GPUs". It is grid connections that take years, buildings designed for racks drawing a tenth of what an accelerator rack draws, and cooling plants that cannot absorb the heat. Accelerators are scarce, but power, cooling and density decide where and when they can be switched on.

Power is the binding constraint

Gartner's press release of November 12, 2024 put a number on it: "Gartner predicts 40% of existing AI data centers will be operationally constrained by power availability by 2027." Bob Johnson, VP Analyst at Gartner, added that "short-term power shortages are likely to continue for years" and that this "will limit the growth of new data centers for GenAI and other uses from 2026". By June 10, 2026 Gartner had moved from prediction to description: alongside a forecast of 565 TWh of worldwide data center consumption in 2026 (up 26% from 447 TWh in 2025), Gartner economist Linglan Wang stated that "AI capacity is now constrained by power availability".

Primary data on scale:

  • Lawrence Berkeley National Laboratory's 2024 Report on U.S. Data Center Energy Use put US data centers at 176 TWh in 2023, about 4.4% of US electricity, and projected 325 to 580 TWh by 2028.
  • The IEA's Energy and AI report (2025) measured 415 TWh globally in 2024 and expects around 945 TWh by 2030. It calls grid connection queues for data centres "long and complex", says wait times for transformers and cables "have doubled in the past three years", and estimates "around 20% of planned data centre projects could be at risk of delays".

Interconnect queues are measured in years

Vendor blogs quote 24 to 36 month approvals in major US and EU markets; that is a vendor estimate, but primary sources agree on direction:

  • CBRE's H1 2025 North America report described "sites offering power access within 18 to 36 months" as highly sought after; that is what now counts as fast.
  • ERCOT CEO Pablo Vegas told the Texas Senate on July 29, 2026 that ERCOT is "tracking approximately 474 GW of Large Loads seeking interconnection, of which ~90% are data centers"; about 205 GW was eligible for the first batch study, due to begin by September 2, 2026.
  • PJM's new generation interconnection Cycle opened in April 2026 "with a one- to two-year timeline for reviews, depending on the impact to the system".
  • On the generation side, LBNL's Queued Up 2024 edition found the median time from request to commercial operation reached five years for plants built in 2023, versus under two years for those built in 2000-2007.

Dell'Oro's October 14, 2026 market update added permitting as a brake: Texas ordered an audit of data center projects in the ERCOT queue, delaying the first large-load studies, and "power, long-lead equipment, and community opposition are all stretching project timelines".

Density: 120 kW racks in 8 kW buildings

The NVIDIA DGX GB200 user guide states that for an NVL72 rack "The rack power consumption is approximately 120kW." Vertiv's reference architecture for the platform, co-developed with NVIDIA, "supports up to 132kW per rack", and TrendForce quotes 130 to 140 kW per rack for GB200 and GB300 NVL72 systems.

Uptime's Global Data Center Survey 2026 found the average modal rack density passed 11 kW for the first time, pulled up by a few new high-density sites; without them it is 7.8 kW, up from 7.5 kW in 2025, and the 4 to 5 kW design density "remains the most common rack density in deployment". One NVL72 rack draws roughly fifteen typical racks' worth of power, and Dell'Oro projects GPU TDPs above 4,000 W by 2029.

The cooling ladder

Each step up changes the mechanical plant, not just the rack.

Cooling method Rack density it supports Status in 2026
Air (CRAC/CRAH with aisle containment) 7.8 kW typical modal rack (Uptime 2026, excluding high-density sites) Majority of installed capacity
Rear-door heat exchangers Up to 80 kW per rack (Vertiv 360AI 1x80 reference design); vendor blogs cite roughly 70 kW Retrofit path for mixed-density halls
Direct-to-chip liquid (cold plates plus CDU) GB200 NVL72 at about 120 kW (NVIDIA) to 132 kW (Vertiv design) "The dominant architecture for AI clusters" (Dell'Oro, January 2026)
Immersion No primary figure confirmed "Selective adoption" (Dell'Oro)

Adoption, from primary sources:

  • TrendForce (August 17, 2026): liquid cooling penetration among AI chips rising from about 33% in 2025 to 53% in 2026 and approaching 60% in 2027.
  • Dell'Oro (January 8, 2026): the liquid cooling market roughly doubled in 2025 to close to $3 billion, heading for $7 billion by 2029; "single-phase direct liquid cooling accounts for the vast majority of liquid-cooled capacity coming online".
  • Uptime Institute (July 2025): among general operators, DLC adoption "remains slow and steady".

Liquid cooling does not fix facility efficiency by itself: Uptime's 2026 industry average PUE is 1.52, flat for seven years, though the largest facilities average 1.45.

Upstream: HBM and packaging

GPU supply is constrained too, but by memory and packaging rather than logic wafers. TrendForce's September 29, 2026 release "expects the HBM market to remain supply-constrained in 2027", forecasts blended HBM prices up 121% year on year in 2027, and reports GPU and ASIC makers "considering reducing HBM capacity per device" via 8-Hi stacks. Its April 30, 2026 release says "CoWoS has remained in short supply since 2023" and expects the shortage to "begin to ease slightly by 2027" as TSMC adds more than 60% CoWoS capacity. Those limits raise prices; a delivered rack still needs 120 kW and a liquid loop.

What it does to colocation pricing and lead times

CBRE's H1 2026 North America report: primary-market vacancy at a record low 1.4% (0.2% in Northern Virginia); 7,481 MW under construction, 80.4% preleased; under 1,500 MW of future capacity available, "approximately six months of demand at the current absorption rate". Asking rates rose in H1 2026 in every size band: 4.3% for 250-500 kW, 7.9% for 500 kW-3 MW, 8.3% for 3-10 MW and 6.7% for 10 MW+. In H1 2025 CBRE expected pricing to "remain elevated, comparable to levels last seen in 2011-2012 at $200+ per kW per month". Lead times are stretching: "RFS fit-out dates are expanding from weeks to several months", and liquid-cooling conversion "delays can extend beyond six months".

If you only need one or two GPUs

A 100 kW rack and a dedicated server with one or two GPUs are different problems. Vendor estimates put an eight-GPU H100 node at roughly 10 to 12 kW; a one- or two-GPU server is a fraction of that and fits in a conventional air-cooled rack, with no CDU or 132 kW bus bar. What you inherit is pricing pressure on space and power in primary markets, longer fit-out times and HBM-driven card prices.

Questions to ask any provider:

  • Deliverable power per rack in kW, on A and B feeds, at what voltage, and what happens if you add a second server.
  • Cooling type (air, rear-door or direct-to-chip) and its rated kW per rack, not the marketing ceiling.
  • Measured annual PUE, which decides how much of the per-kW rate is overhead; the Uptime average is 1.52.
  • Power-expansion and interconnect timeline: is utility capacity energized, is expansion contracted, and what is the lead time.

Why "GPU hosting" quotes vary so much: one is a card in a 5 kW air-cooled rack in a secondary market; another includes a liquid loop, 2N power, a primary-market lease signed at 1.4% vacancy and a GPU whose memory is doubling in price. Ask for kW per rack, cooling type, PUE and power timeline and the quotes become comparable.

Sources