Definition: A data center is the purpose-built facility of racked servers, power distribution, and cooling that physically runs software — and for AI work it is measured in megawatts rather than square feet, because power density is what limits how much compute fits inside.
TL;DR: A data center is the building your AI request travels to. An AI-focused facility differs from a traditional one mainly in power density: a flagship AI rack draws about 120 kW against roughly 10 kW for a general-purpose rack, which forces liquid cooling and a much larger electrical system. Global data centres used about 415 TWh in 2024, roughly 1.5% of world electricity. Build an app free →
Why AI Data Centers Are a Different Building Type
The difference is not the software. It is watts per rack.
A general-purpose rack draws roughly 10 kW. An AI rack draws roughly 60 kW, and flagship rack-scale systems draw about 120 kW. That single ratio cascades into everything else: the electrical rooms grow, air cooling stops working, the floor needs plumbing, and siting is decided by access to megawatts rather than proximity to users.
One number worth keeping in view: the fleet-wide deployed average rack across all the world's data centers is still just over 11 kW (Uptime Institute). The 120 kW racks are the frontier, growing fast, and a minority of installed capacity.
| Dimension | Traditional | AI |
|---|---|---|
| Rack power | 5–15 kW | 60–120 kW+ |
| Cooling | Air | Liquid to the chip |
| Dominant cost | Real estate and power | Accelerators |
| Siting constraint | Latency to users | Access to megawatts |
| Network | Ethernet, north-south | High-bandwidth internal fabric |
How Power Reaches a Chip
Electricity arrives from a transmission line at 230–345 kV and steps down five or six times before it reaches silicon. Uninterruptible power supplies are not a power source — they hold the load up for the seconds it takes generators to start.
Six conversions between the grid and the chip. Each box is a room with its own multi-year lead time.
The Binding Constraint Is Time, Not Megawatts
The scarce resource in 2026 is not electricity. It is permission to connect and equipment to connect with. Interconnection queues in the United States exceed four years, and large power transformers moved from roughly one-year lead times to two and a half to four years.
Lawrence Berkeley National Laboratory's queue analysis adds the discount that most coverage omits: only about 13% of capacity that entered interconnection queues between 2000 and 2020 had reached commercial operation by the end of 2025, while roughly 75% withdrew. Queue figures are piles of applications, not incoming capacity.
Where the Power Goes Inside the Fence
The IEA's breakdown is the best-sourced version: servers take roughly 60%, cooling anywhere from 7% in an efficient hyperscale facility to over 30% in a less efficient enterprise one, with storage and networking around 5% each. That 4× spread in cooling share is why a single efficiency number tells you very little on its own.
And essentially all of it ends up in the same place. Better than 99% of the electricity a data center draws leaves as low-grade heat, typically at 45–70 °C. The tokens carry a negligible fraction of the energy — which is why cooling is close to half the capital cost of an AI facility rather than a support function.
PUE, and What It Buys on a 100 MW Site
The efficiency number the industry runs on is PUE — power usage effectiveness: total facility power divided by the power that reaches IT equipment. A PUE of 1.4 means that for every watt of compute, another 0.4 W goes to cooling, power conversion, and everything else. Typical published figures put air-cooled facilities at roughly 1.3–1.6 and modern liquid-cooled ones at 1.03–1.1.
That decimal is worth real money, because a site's grid connection is fixed long before its racks are chosen. Run the math on a 100 MW connection:
| Facility | PUE | Power reaching compute | ~GB200 NVL72-class racks at ~130 kW |
|---|---|---|---|
| Older air-cooled | 1.6 | 62.5 MW | ~480 |
| Typical air-cooled | 1.3 | 77 MW | ~590 |
| Liquid-cooled | 1.1 | 91 MW | ~700 |
| Best liquid-cooled | 1.03 | 97 MW | ~745 |
Same grid connection, same fence line — and the best facility fits roughly 260 more flagship racks than the worst, over a third more compute, purely by wasting less of the intake on overhead. Since each of those racks holds 72 GPUs, the spread is nearly 19,000 accelerators. This is why liquid cooling is an economic decision before it is a thermal one, and why operators chase hundredths of a point of PUE.
The same logic scales down to one request: Google reports a median Gemini text prompt consumes about 0.24 watt-hours end to end — small per question, multiplied by everything.
A Short History: From Machine Rooms to Megawatts
The building type has been reinvented three times. Mainframes lived in raised-floor "machine rooms" inside ordinary offices. The dot-com era moved servers into shared colocation facilities sold by the square foot and the rack. The hyperscale era, roughly 2006 onward, saw cloud providers design entire warehouse-scale buildings around one operator's workload — and made PUE the industry's scoreboard. The AI era is the fourth reinvention: the unit of account shifted from square feet to megawatts, the constraint from floor space to the grid connection, and the cooling medium from air to liquid. A modern AI campus is best understood as an industrial power facility that happens to emit tokens.
Related Concepts
- GPU — the processor the building is designed around
- Memory Bandwidth — the constraint inside the chip
- HBM — the memory supply chain the fleet depends on
- Inference — what the building spends its power on
- Tokens Per Second — the building's output, measured
- Inference Cost — what that translates to on a bill
- Model Training — the workload that tolerates being far from users
- Agent Infrastructure
Frequently Asked Questions About Data Centers
What makes a data center an "AI data center"?
Power density and cooling. A facility organised around accelerators runs racks six to twelve times denser than a general-purpose one, which requires liquid cooling, a larger electrical system, and a high-bandwidth internal network. The workload shape differs too: fewer, larger, more synchronised jobs.
How much electricity do data centers use?
About 415 TWh globally in 2024, roughly 1.5% of world electricity, according to the IEA, which projects that more than doubling to around 945 TWh by 2030. For the United States specifically, Lawrence Berkeley National Laboratory's 2025 Update puts 2024 consumption at 192 TWh, or 4.7% of national electricity — a downward revision of the lab's own earlier estimates.
Why do AI data centers need liquid cooling?
Because air cooling becomes impractical around 30 kW per rack and AI racks draw far more. Per unit of volume water carries about 3,500 times more heat than air, so removing 120 kW takes roughly 21,000 cubic feet per minute of air or about 46 gallons per minute of water.
What is PUE in a data center?
Power usage effectiveness: total facility power divided by the power delivered to IT equipment. A perfect score is 1.0. Air-cooled facilities typically run around 1.3–1.6 while modern liquid-cooled ones reach 1.03–1.1, and on a fixed grid connection that gap translates directly into how much compute fits on the site.
How much power does one AI rack use?
A flagship rack-scale system such as NVIDIA's GB200 NVL72 draws roughly 120–140 kW for its 72 GPUs — enough that a single rack outdraws a dozen general-purpose ones. That density is what forces liquid cooling and reshapes the building's entire electrical system.
Why are AI data centers built in remote places?
Because training workloads follow megawatts, not users. A training run does not care about latency to a city, so sites go where grid capacity, land, and cooling are available and cheap. Inference facilities sit closer to population centers, since interactive requests do feel the distance.
Does Taskade run its own data centers?
No. Taskade is a hosted product that runs on managed cloud infrastructure and buys model capacity from providers. The practical effect is that you rent a share of this machinery by the token rather than sizing a GPU fleet or waiting in an interconnection queue: every AI agent reply, automation run, and sync across 100+ integrations is a metered visit to a building like this.