- Power and cooling constraints should be validated before GPU density.
- East-west and north-south networking solve different problems.
- Capacity planning should include growth and workload behavior.
AI workloads do not simply need more of a traditional data center’s power, cooling and connectivity. They change the design assumptions behind all three.
Direct Answer
Planning AI data-center infrastructure means evaluating power density, cooling method, connectivity and interconnection, and compute planning as one system. AI racks can draw far more power than traditional enterprise racks; dense deployments may require liquid cooling; the compute fabric and facility connections solve different network problems; and the planned GPU or accelerator density must fit the power and cooling the site can actually deliver. Getting these decisions right up front helps avoid an expensive retrofit later.
Power: the constraint that shapes everything else
Traditional enterprise racks often operate around 5–10 kW, while GPU-dense AI racks may reach 40 kW, 80 kW or more depending on the hardware and configuration. That difference changes the conversation from “How much space do we need?” to “How much power can this site actually deliver, and how soon?”
Power availability and the lead time to secure more of it are frequently among the biggest constraints in AI data-center planning. Sites near available utility capacity, or with committed upgrades already underway, can be more valuable than raw square footage alone would suggest.
Cooling: air often is not enough anymore
Traditional air cooling becomes harder to use efficiently as rack power density rises. Direct-to-chip liquid cooling, rear-door heat exchangers and immersion cooling have therefore become increasingly common for dense AI deployments.
Evaluating a facility for AI workloads means confirming which cooling methods it can support in practice—not only what a specification sheet suggests. Retrofitting liquid cooling into a facility that was not designed for it can be expensive and slow.
Connectivity: two different problems, not one
East-west: compute to compute
East-west connectivity is the fabric linking GPUs and compute nodes within and across racks. Bandwidth, latency and consistency can directly affect distributed training performance, so this layer typically requires a very high-bandwidth, low-latency interconnect designed for the workload.
North-south: facility to the world
North-south connectivity moves traffic into and out of the facility—to data sources, users, other sites and the cloud. This resembles traditional enterprise and data-center interconnection: dedicated circuits, wavelengths, dark fiber or cloud on-ramps sized for the workload’s actual data-movement patterns.
Planning both layers together, while sizing each independently, reduces the risk of under-building the part that ultimately constrains performance.
Compute planning: match density to what the site can deliver
It is possible to specify more GPU density than a site can power and cool. Working backward from confirmed electrical and cooling capacity—rather than forward from a desired GPU count—helps avoid committing to a facility that cannot support the deployment.
A practical framework
- Confirm available power. Validate what the site can deliver now and the lead time to add more; this frequently drives the schedule.
- Confirm supported cooling methods. Determine whether liquid cooling is available or would require a build-out.
- Separate the connectivity requirements. Size east-west compute fabric and north-south facility connectivity independently.
- Fit compute to the site. Base GPU or accelerator density on confirmed power and cooling capacity—not the reverse.
- Build in growth headroom. AI infrastructure demand can grow faster than traditional enterprise capacity plans assume.
Related planning guides
- What Data Centers Should Consider Before Expanding into AI →
- Private AI Infrastructure: When Shared Cloud Compute Is Not the Answer →
- Governed AI Infrastructure: Building Oversight In From the Start →
Frequently asked questions
Why does AI infrastructure need so much more power than traditional enterprise infrastructure?
GPU and accelerator hardware used for AI training and inference can draw substantially more power per rack than traditional enterprise compute. That makes utility capacity, electrical distribution and delivery timelines central site-planning constraints.
Do all AI data centers require liquid cooling?
No. The answer depends on the planned rack density and hardware. Liquid cooling becomes increasingly important—and may be necessary—as power density rises beyond what the facility’s air-cooling system can support efficiently.
What is the difference between east-west and north-south connectivity in an AI data center?
East-west connectivity is the high-bandwidth, low-latency fabric between compute nodes. North-south connectivity carries traffic between the facility and data sources, users, other sites and cloud services. The two requirements should be sized separately.
Should compute density or site infrastructure come first in planning?
Confirmed site infrastructure should generally come first. Planning GPU density before validating available power and cooling can create a deployment the facility cannot support.
Related services
Related articles
Share the locations, applications, performance objectives, risks and timeline. SpartaLink can help frame and compare the right options.
Discuss the Requirement