Which AI capacity path should buyers compare first?
Start with the constraint most likely to break the plan: time to power, rack density, cooling readiness, network reach, lease term, or operating control. For near-term experiments or variable demand, GPU cloud and reserved GPU capacity usually reduce setup burden. For sustained GPU clusters, high-density colocation or a wholesale suite can improve control and unit economics if the buyer can manage hardware, network, and remote-hands workflows. Powered shell and owned campus paths are better suited to durable load, larger MW commitments, and teams that can carry development risk.
AI capacity should be compared as a portfolio decision rather than a single-provider search. The right path can change when the workload moves from training to inference, when cabinet density crosses air-cooling limits, or when the buyer needs a specific metro, cloud on-ramp, or utility queue position.
| Capacity path | Best fit | Buyer caveat |
|---|---|---|
| GPU cloud or reserved GPU capacity | Fast starts, uncertain demand, experiments, short runs | Sustained usage can be expensive and region availability can shift |
| AI-ready colocation | Owned GPU hardware, stable workloads, network control | Verify rack kW, liquid-cooling support, remote hands, and expansion rights |
| Wholesale suite or private data hall | Multi-rack or multi-MW deployments needing more control | Longer contracting cycles and heavier diligence on power and fit-out |
| Powered shell | Buyers with fit-out capability and longer planning windows | Schedule risk moves to the buyer, including MEP, cooling, and commissioning |
| Owned campus or build-to-suit | Durable load, site control, and strategic power access | Requires development expertise, capital planning, and utility delivery confidence |
How should buyers trade off speed, control, and cost?
Speed is rarely free. The fastest options usually ask the buyer to accept less site control, less hardware choice, or a shorter planning window for network and compliance. The most controlled options usually require more diligence around utility delivery, mechanical design, commissioning, insurance, and operating responsibility. A strong short list shows where the buyer is consciously paying for speed, certainty, flexibility, or asset control.
Use the first meeting to separate available capacity from usable capacity. A facility that can quote space may still fail the workload if the rack density, cooling architecture, network route, or expansion rights do not match the deployment. Buyers should also keep lease term and exit risk visible. A one-year GPU experiment and a five-year inference platform should not be evaluated with the same procurement scorecard.
| Constraint | Usually favors | Verify before signing |
|---|---|---|
| Fastest start | GPU cloud, reserved GPU clusters, or existing AI colocation | Actual cluster size, region, lead time, support model, and renewal terms |
| Highest hardware control | Colocation, wholesale suite, powered shell, or owned build | Remote hands, spare parts, firmware access, compliance, and change windows |
| Highest power density | AI-ready colocation, liquid-cooled suites, or purpose-built halls | Rack kW, CDU boundaries, water policy, commissioning load, and heat-rejection path |
| Lowest development burden | GPU cloud or managed colocation | SLA language, escalation path, support response, and cost at sustained utilization |
| Longest asset control | Powered shell, build-to-suit, or owned campus | Utility milestones, substation scope, permitting, interconnection, and contingency |
What should provider diligence prove before an RFP?
Capacity claims should be converted into evidence before a buyer issues an RFP. Ask for the committed electrical capacity, the density that has already been operated, the cooling architecture available today, the network fabric and carriers in the building, the expansion pathway, and the parties responsible for commissioning and operations. If the buyer owns the GPU hardware, also verify freight, staging, security, remote-hands procedures, spares, and failure replacement windows.
The biggest mistake is treating MW as interchangeable. A 5 MW deployment with 20 kW racks, a 5 MW deployment with 100 kW racks, and a 5 MW powered shell all create different cooling, operations, and contract risks. Buyers should request the evidence that matches the workload profile, not a generic data hall overview.
| Diligence area | Buyer question | Evidence to request |
|---|---|---|
| Power | Is the capacity committed, reserved, or still dependent on utility delivery? | Utility letter, energization schedule, one-line diagram, redundancy design |
| Cooling | Can the facility support the target rack density without a retrofit surprise? | Supported kW per rack, liquid loop status, CDU scope, heat-rejection plan |
| Network | Can the site support training, inference, storage, and cloud connectivity needs? | Carrier list, cross-connect process, cloud on-ramp access, latency data |
| Operations | Who handles incidents, hardware swaps, access, and compliance evidence? | Remote-hands SLA, security process, maintenance windows, audit support |
| Expansion | Can the buyer grow without renegotiating the whole footprint? | Expansion rights, adjacent capacity, phase schedule, pricing assumptions |
Which next pages should buyers use?
Use the lease-capacity guide when the core question is whether to use GPU cloud, colocation, wholesale, powered shell, or an owned build. Use the colocation provider comparison when the buyer already owns or plans to buy GPU servers and needs to evaluate rack density, liquid cooling, interconnection, compliance, and remote-hands fit. Use the market power page when geography, utility timelines, land, permitting, or fiber access can change the short list.
The construction cost guide and capacity lease RFP builder handle the next layer of diligence. The cost guide helps normalize shell, power, cooling, land, labor, and contingency assumptions before comparing build and lease options. The RFP builder turns those assumptions into a provider-ready brief so sales conversations start with density, MW, term, cooling, network, and expansion facts instead of a vague space request.