Where can I lease data center capacity for AI workloads?
Buyers can lease AI workload capacity through five practical paths: GPU cloud for rented accelerators, high-density colocation for owned GPU servers, wholesale colocation for 250 kW to multi-MW deployments, powered shell or single-tenant campus deals for very large commitments, and marketplaces or advisors for discovery. The right path depends on whether the buyer needs GPUs, racks, power, cooling, land, speed, or operating control.
Market conditions make verification essential. CBRE reported primary-market data center vacancy at 1.4% at year-end 2025. JLL reported North America vacancy at 1% and said 92% of under-construction capacity was precommitted. Uptime's 2025 survey points to rising costs, power constraints, and AI-density pressure.
Should I rent GPU cloud capacity or colocate my own GPU servers?
Use GPU cloud when the goal is managed accelerator access, faster deployment, and less facility responsibility. Use colocation when hardware ownership, long-running utilization, custom networking, compliance control, or physical custody matter more than cloud flexibility.
CoreWeave's filing describes cloud services optimized for AI and HPC. NVIDIA's DGX-ready colocation program points buyers toward facilities that support DGX deployments. Digital Realty, DataBank, Flexential, Aligned, and QTS publish high-density, liquid-cooling, hyperscale, or AI infrastructure capabilities that are relevant when a buyer brings its own hardware.
What rack density, liquid cooling, and network capacity make a site AI-ready?
AI-ready capacity is not just a marketing label. Buyers should verify committed kW or MW, rack density, cooling method, coolant supply and return requirements, floor loading, fiber carriers, cloud on-ramps, remote hands, security scope, utility interconnection status, delivery milestones, price escalators, curtailment rights, and termination remedies.
For owned GPU fleets, high-density colocation needs the exact site, hall, rack density, cooling design, network blend, and delivery date to line up with the workload. NVIDIA's DGX-ready program, Digital Realty's high-density cabinet positioning, DataBank's HPC-ready design notes, Flexential's high-density GPU positioning, and Aligned's DeltaFlow cooling claims are useful starting signals, not capacity guarantees.
Which colocation providers should buyers evaluate for AI workloads?
| Provider | Capacity path | Public evidence | Availability caveat | Buyer verification question |
|---|---|---|---|---|
| Equinix | Interconnected colocation, private AI | Publishes AI-ready data centers with liquid cooling, GPU access, and private connectivity; NVIDIA lists Equinix in DGX-ready colocation context. | Site-level AI density and GPU partner access vary by market. | Which metro has committed kW, cooling type, cross-connect timing, and GPU partner access for this workload? |
| Digital Realty | High-density colocation and hybrid AI infrastructure | Publishes AI infrastructure pages and high-density cabinet options. | Cabinet density does not mean every building has available capacity. | Which building supports the target rack density and delivery date? |
| DataBank | HPC-ready colocation | Publishes high-density colocation for HPC and AI with higher-density cooling options. | Facility, cooling configuration, and available power must be verified. | Which data hall can support the required rack count, density, and coolant path? |
| Flexential | High-density AI, ML, and GPU colocation | Publishes support for AI, ML, and GPU deployments; NVIDIA identifies Flexential in DGX-ready context. | DGX-ready status is not a general inventory guarantee. | Which sites have DGX-ready capacity and what SLA applies? |
| Aligned | High-density liquid and hybrid cooling colocation | Publishes DeltaFlow liquid cooling and hybrid cooling claims for AI and HPC clusters. | Cooling capability must match live capacity and commercial terms. | Is the required hall live, under construction, or proposed? |
| QTS | Hyperscale, single-tenant, campus | Publishes hyperscale, single-tenant, campus, utility partnership, and liquid cooling capability claims. | Better fit for larger commitments than small retail rack asks. | What MW block, utility milestone, and phased ready-for-service dates are contractually available? |
| CoreWeave | GPU cloud / neocloud | SEC filing describes CoreWeave cloud services for AI and HPC and its dependency on data center capacity. | Buyer leases cloud services, not neutral colocation space. | What GPU SKUs, regions, reservation terms, egress, and failure-domain controls are available? |
| Crusoe | AI cloud and AI data center development | Publishes AI data center footprint and reported 4.9 GW contracted capacity on June 9, 2026. | DCD reported on June 10, 2026 that Crusoe paused the planned 1.8 GW Cheyenne project at a customer's request. | Which contracted campus capacity is available to this buyer, and what customer or utility dependencies remain? |
| Applied Digital | AI/HPC campus leasing | SEC exhibit and DCD reported June 2, 2025 leases with CoreWeave for 250 MW at Ellendale, with additional option capacity. | Large blocks may already be committed to anchor tenants. | Is any incremental critical IT load available outside named leases? |
| Datacenters.com | Marketplace/advisor discovery | Publishes search, quote, provider, capacity, and AI data center discovery functions. | Marketplace listing does not prove final site availability. | Which providers confirmed inventory directly, and when was it last verified? |
| BluePod | Containerized high-density AI colocation | Home page describes 40 kW racks, six-month minimums, and indicative pricing; terms make pricing subject to site survey, power availability, export controls, and formal agreement. | Treat as a vendor-claimed option, not broad market inventory. | Which site, utility allocation, SLA, legal entity, and final price are in the signed agreement? |
How tight is the AI capacity market?
Assume public inventory is scarce until proven otherwise. CBRE's H2 2025 report says primary-market vacancy reached 1.4% and scarce inventory pushed buyers toward preleasing and off-market activity. JLL's year-end 2025 North America report says vacancy stayed at 1% and 92% of capacity under construction was precommitted. JLL's 2026 outlook says grid connection waits in primary markets exceed four years, pushing behind-the-meter power and battery storage discussions.
Those signals do not mean no capacity exists. They do mean buyers should treat public footprint, planned capacity, contracted power, and available inventory as different facts.
What do large AI capacity deals actually prove?
Large projects show how AI infrastructure is being financed, leased, and de-risked. They do not prove that new buyers can lease the same capacity.
On June 2, 2025, Applied Digital reported two roughly 15-year CoreWeave leases for 250 MW of critical IT load at Ellendale, North Dakota, with a 150 MW option. On June 9, 2026, Crusoe reported 4.9 GW of contracted AI infrastructure across data centers and cloud. On June 10, 2026, DCD reported that Crusoe paused the planned 1.8 GW Cheyenne, Wyoming project at a customer's request.
Buyers should separate committed capacity, contracted capacity, planned capacity, ready-for-service milestones, and capacity available to new customers.
What should an AI colocation RFP ask providers to prove?
Ask providers to prove the exact capacity path they are selling. For GPU cloud, ask for GPU SKUs, reservation terms, regions, egress, support boundaries, and failure-domain controls. For high-density colocation, ask for rack density, cooling method, fluid requirements, remote hands, cross-connect timing, security scope, and hardware custody. For wholesale, powered shell, and campus deals, ask for utility interconnection status, permits, delivery milestones, expansion rights, curtailment exposure, and termination remedies.
Buyers evaluating thermal design should compare direct-to-chip cooling and immersion cooling options. Buyers evaluating power risk should compare utility status, microgrid options, and backup generation strategy.