Industry Statistics

AI Infrastructure Statistics: Power, Cooling, Latency, and Networking

A living reference page of useful AI infrastructure ranges and bounds. Use it for orientation, then read primary sources before making design or investment decisions.

· 12 min read

By Seabase Editorial

← Articles index · Home · Infrastructure · Nori · About

Topics: Infrastructure · Cooling · AI Inference · Energy · Geography

People looking for subsea data centers, AI inference infrastructure, cooling, and coastal placement often need a short set of comparable numbers. This page collects commonly cited industry ranges, physics bounds, and deployment comparisons in one place.

It is designed to be cited and updated. It is not a substitute for vendor datasheets, utility studies, or site-specific engineering.

Treat every table as a planning range until you validate it against primary sources for your site and generation.

How to use this page

Canonical short URL: /statistics. This article is the same living reference.

Rack power evolution (typical planning ranges)

AI has pushed rack power far beyond classic enterprise densities. Exact values depend on accelerator generation, networking, liquid cooling coverage, and facility limits. The ranges below are planning envelopes commonly discussed in industry materials, not a single standard.

Era / class (illustrative)Typical rack power (planning range)Notes
Legacy enterprise / light density5–15 kWAir-cooled, limited GPU density
Early AI / mixed GPU racks20–40 kWOften still air-assisted
Dense AI racks40–80 kWLiquid cooling often required
High-density AI racks80–120+ kWFacility and CDU limits dominate

Related: Direct Liquid Cooling, Coolant Distribution Unit, and Can Underwater Data Centers Reduce Cooling Costs?.

Power Usage Effectiveness by Facility Type

PUE is total facility energy divided by IT energy. Lower overhead is better, but PUE is not a complete sustainability or cost score. Climate, utilization, electrical design, and cooling architecture all matter. See PUE.

Facility contextTypical PUE rangeCapital and operating context
Older enterprise data centers1.5–2.0+Performance varies widely by facility age, utilization, climate, and cooling architecture.
Modern hyperscale campuses1.1–1.4Achieving these levels typically requires substantial investment in optimized buildings, cooling plants, electrical systems, controls, and site infrastructure.
Leading high-efficiency campusesApproximately 1.1 or lowerThese results are generally associated with large, purpose-built facilities, favorable climates, high utilization, and significant capital investment.
Seabase design targetApproximately 1.05Seabase is designing for comparable or better infrastructure efficiency by taking advantage of modern direct liquid cooled AI systems, seawater as a stable heat sink, and a modular architecture intended to reduce the amount of cooling and facility infrastructure required.

PUE alone does not describe total infrastructure economics. Two facilities can report similar PUE while requiring very different levels of capital investment, cooling equipment, land, construction, and ongoing operations.

Seabase is modeling toward approximately 1.05 PUE while targeting lower infrastructure capital requirements and lower cooling-related operating costs than conventional high-efficiency campuses. These remain design objectives subject to engineering validation, site conditions, utilization, equipment configuration, and consistent metering boundaries.

Seabase is not pursuing low PUE by recreating a hyperscale cooling campus offshore. It is designing a simpler thermal architecture intended to reach approximately 1.05 PUE with less supporting infrastructure, lower cooling overhead, and a smaller physical footprint.

Related: Environmental Accountability.

Network latency and distance (physics-informed)

Light in fiber is slower than in vacuum. A practical planning rule of thumb is that one-way fiber delay is on the order of about 5 microseconds per kilometer of fiber path (path length, not straight-line map distance). Round-trip delay is roughly twice the one-way path delay, before equipment and queuing.

Approximate fiber pathRough RTT floor (order of magnitude)What it means
Same metro / nearby campus (~10–50 km path)low single-digit to ~0.5 ms path floorApplication and peering often dominate
Regional coastal corridor (~200–500 km path)~2–5 ms path floorStill strong for many interactive inference cases
Cross-country continental path (~4,000 km)~40 ms path floorBefore peering and congestion
Intercontinental pathstens to 100+ msRoute and undersea cable path matter

These are path floors, not measured user latency. Queuing, Wi-Fi, mobile cores, and multi-step agent calls add more. Deep dive: The Latency Tax and Why AI Inference Changes Infrastructure Geography.

Why multi-step AI makes geography matter

If one network round trip costs 60 ms and an agent pipeline needs four sequential model or tool calls, network delay alone can impose about 240 ms before counting inference time. That compounding effect is why metro and coastal placement appear repeatedly in inference planning.

Cooling methods by deployment type

Deployment typePrimary heat rejectionTypical tradeoff
Air-cooled land hallCRAH / CRAH + chillers / evaporativeFamiliar ops; limited at high rack density
Direct liquid cooling (land)Cold plates + CDU + facility loopEnables denser racks; adds fluid systems
Immersion (land)Dielectric fluid bathsHigh density; service model changes
Floating / barge marineSeawater exchangers + surface accessAccess easier; motion and weather exposure
Subsea sealed modulesModule-to-seawater heat rejectionStrong sink coupling; retrieval/service design critical

Related: What Is a Subsea Data Center?, Floating vs. Subsea Data Centers, Underwater vs Land Data Centers.

Cable landing stations and coastal networking

Most intercontinental internet traffic travels on subsea fiber. Cable landing stations (CLS) are where those systems come ashore and hand off to terrestrial networks.

Related: The Durable Geography of AI Demand and Metro AI Infrastructure.

AI power context (facility planning)

AI clusters are persistent electrical loads. Nameplate renewable capacity is not the same as firm delivered capacity for a continuous compute load. Batteries help with ride-through and smoothing. They are not a primary energy source.

Related: An Electron Is an Electron and power-flexible AI infrastructure.

Subsea thermal orientation

Ocean-adjacent and subsea designs use seawater as a stable heat sink for direct liquid cooled AI hardware. Modern DLC systems operate at higher coolant temperatures, increasing the temperature differential relative to surrounding seawater. That larger ΔT supports compact heat exchange and favors modular architectures that can follow successive hardware generations. Local thermal effects still depend on site conditions, flow design, equipment configuration, and monitoring practice.

See environmental research and Environmental Accountability.

Methodology and update policy

If you cite this page, include the access date. For industry narrative context, also see State of the Subsea Data Center Industry.

Frequently asked questions

What is a typical AI rack power density?

Planning ranges vary widely by generation. Dense AI racks often fall in the tens of kilowatts, and high-density designs can exceed 80–100 kW per rack. Always validate against the specific accelerator and cooling design.

What PUE should I expect?

Modern hyperscale campuses often report roughly 1.1–1.4. Leading high-efficiency campuses sometimes report approximately 1.1 or lower. Seabase is designing toward approximately 1.05 PUE as a modeled target, subject to engineering validation and operating conditions.

Next step

If you want help interpreting these ranges for a coastal deployment, contact Seabase.

Contact Seabase