What Is an AI Data Center? Cost, Power, and Scale
An AI data center explained: how it differs from a traditional data center, what it costs per megawatt, and why power is the binding constraint.
By Capital & Compute
An AI data center is a building designed to run dense clusters of GPUs for training and serving artificial intelligence, and the single number that separates it from a traditional data center is power density. A conventional server rack draws under 10 kilowatts. One Nvidia GB300 NVL72 rack draws about 135 kilowatts. Everything else that is strange about these buildings, the liquid plumbing, the gigawatt land deals, the fights over the electric grid, follows from that roughly twentyfold jump in power per rack.
That is the whole story in one sentence: an AI data center is a traditional data center rebuilt around the problem of getting power into, and heat out of, a very small amount of floor space.
What is an AI data center?
A data center is a purpose-built facility that houses computer servers, storage, networking, power distribution, and cooling. A traditional or “cloud” data center is optimized for many small, independent workloads: web apps, databases, email, virtual machines. Racks are filled to a modest power budget, servers are cooled by moving air, and the facility is designed for high availability across thousands of unrelated tenants.
An AI data center is optimized for the opposite: a small number of enormous, tightly coupled workloads. Training a frontier model means running tens of thousands of GPUs as a single machine, synchronized over high-speed networking, for weeks at a time. That changes every design decision. The GPUs must sit physically close together to keep network latency low, which concentrates power and heat. The networking is exotic and expensive. Utilization is pushed as high as possible because idle GPUs are the most expensive thing in the building.
The result looks less like an office of servers and more like a factory built around a furnace. Nvidia even markets its rack-scale systems as “AI factories.”
AI data center vs traditional data center
The clearest way to see the difference is per rack. A rack is the standard steel cabinet that holds servers. In a traditional facility, the average rack draws under 6 kilowatts, and by the industry’s own survey most operators run nothing above 20 kW, according to the Uptime Institute Global Data Center Survey. An AI rack built on Nvidia’s latest hardware, the GB300 NVL72 documented in Lenovo’s product guide, packs 72 Blackwell Ultra GPUs into one cabinet and draws roughly 132 to 135 kilowatts, with peaks near 155 kW.
| Tool | Cost per task | Multiple of baseline |
|---|---|---|
| Typical enterprise rack | 6 kW | 1.0x |
| Cloud / colocation rack | 15 kW | 2.5x |
| Air-cooled 8-GPU rack | 40 kW | 6.7x |
| GB200 NVL72 rack | 120 kW | 20.0x |
| GB300 NVL72 rack | 135 kW | 22.5x |
That density cascades through the whole building:
- Cooling. Air stops working somewhere around 50 kW per rack. Above that, operators move to direct-to-chip liquid cooling, where coolant runs through cold plates sitting on the chips. In the GB300 NVL72, roughly 90% of the heat is captured by liquid and only 10% by air.
- Power delivery. A traditional facility runs power to a rack over a normal circuit. An AI rack needs a dedicated bus bar and a wall of power shelves. Vendors advise provisioning busways well above the nominal draw to handle peaks.
- Efficiency. Overhead is measured by PUE (power usage effectiveness), the ratio of total facility power to power that reaches the computers. The global average sits around 1.5 to 1.6, meaning half again as much energy goes to cooling and losses as to compute. Liquid cooling is one of the few levers that pushes that number down at high density.
For the economics of owning this hardware rather than renting it, the same density math drives the breakdown of self-hosted cost per token and the decentralized GPU versus cloud cost comparison.
How much power does an AI data center use?
At the fleet level, the numbers stop sounding like buildings and start sounding like national grids. The International Energy Agency’s Energy and AI report (2025) estimates global data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of the world’s total, and projects that figure roughly doubling to about 945 TWh by 2030. The agency attributes about half of that net increase to accelerated (AI) servers, which it expects to grow around 30% a year while conventional servers grow about 9%.
Goldman Sachs Research frames the same trend as capacity: global data-center power demand of roughly 55 gigawatts in 2023 rising to about 84 GW by 2027 and 122 GW by 2030, a 165% increase, requiring on the order of $720 billion in grid spending along the way.
| Step | Change | Running total |
|---|---|---|
| 2023 baseline | 55 GW | 55 GW |
| Growth to 2027 (reaches 84 GW) | +29 GW | 84 GW |
| Growth to 2030 (reaches 122 GW) | +38 GW | 122 GW |
| 2030 demand | 122 GW | 122 GW |
This is why power, not chips or capital, is now the binding constraint. Grid-connection queues for large new loads stretch three to five years in many US markets. A campus can have its GPUs on order and its financing closed and still wait years for a substation. That constraint is what pushes hyperscalers toward their own gas turbines, nuclear power-purchase agreements, and remote sites chosen for spare grid capacity rather than proximity to users.
How much does an AI data center cost to build?
Cost scales with power, so the industry increasingly quotes it per megawatt of IT capacity. A conventional hyperscale build has historically run around $10 million per megawatt. Industry estimates put next-generation AI halls well above that, in the rough range of $15 to $20 million per megawatt, because the liquid cooling, electrical gear, and high-density fit-out cost far more than a standard cloud row. (These per-MW figures come from real-estate and analyst estimates reported across 2026 industry coverage rather than a single audited source, so treat them as a band, not a quote.)
The concrete campus numbers are easier to pin down because operators announce them. Meta’s Hyperion site in Louisiana is a roughly $27 billion development, built through a joint venture with Blue Owl Capital, starting at 2 gigawatts and scaling toward 5. OpenAI’s Stargate program is pitched at up to 10 gigawatts and hundreds of billions of dollars, with its first Abilene, Texas campus targeting about 1.2 GW and hundreds of thousands of GPUs. At gigawatt scale, a single campus costs more than most companies are worth.
Where that money comes from is its own story, increasingly private credit and off-balance-sheet vehicles rather than corporate cash. That capital structure is the subject of AI data center financing, and the operators who rent this capacity by the GPU-hour are covered in what CoreWeave does. The build cost also flows downstream into the price of using AI, which is the throughline in what it costs to train AI models.
Water and power: the resource cost
The two resources an AI data center consumes at scale are electricity and water, the latter used mostly for cooling. Evaporative cooling towers are cheap and effective but consume large volumes of water. Google’s 2025 Environmental Report disclosed roughly 8.1 billion gallons of water withdrawn across its operations in 2024, up about 28% year over year, tracking its AI-driven electricity growth.
Liquid cooling changes this equation in both directions. Direct-to-chip and immersion cooling can cut direct water use sharply compared with evaporative systems (vendor estimates commonly cite reductions of 70% to 90%, though the real figure depends heavily on design and climate). But the electricity those chips consume still has a water and carbon footprint upstream at the power plant. The honest summary is that liquid cooling moves the resource problem more than it eliminates it.
Bottom line
An AI data center is a traditional data center turned inside out by power density. Once a rack jumps from 6 kW to 135 kW, air cooling gives way to liquid, ordinary circuits give way to bus bars, and a building’s appetite for electricity outgrows the local grid. That single physical fact explains the gigawatt land deals, the multibillion-dollar campuses, the nuclear contracts, and the three-year waits for a grid connection. The chips get the headlines. The power is the constraint.
Frequently asked questions
- What is an AI data center?
- An AI data center is a facility built to run dense clusters of GPUs for training and serving AI models. Its defining feature is power density: an AI rack can draw about 135 kW versus under 10 kW for a traditional rack, which forces liquid cooling and gigawatt-scale power.
- How is an AI data center different from a normal data center?
- A normal data center runs many small, independent workloads on air-cooled racks drawing well under 20 kW each. An AI data center runs a few enormous, tightly coupled GPU workloads at over 100 kW per rack, requiring liquid cooling, exotic networking, and far more power and space per unit of floor.
- How much power does an AI data center use?
- A single modern GPU rack draws roughly 135 kW. At the fleet level, the IEA projects global data-center electricity roughly doubling from 415 TWh in 2024 to about 945 TWh by 2030, with AI servers driving about half of that increase.
- How much does it cost to build an AI data center?
- Conventional hyperscale builds have run around $10 million per megawatt; industry estimates put next-generation AI halls roughly in the $15 to $20 million per MW range. At campus scale, projects run tens of billions of dollars, such as Meta Hyperion at about $27 billion for 2 to 5 GW.
- Why do AI data centers need so much water?
- Water is used mainly for cooling. Evaporative cooling towers consume large volumes; Google reported about 8.1 billion gallons withdrawn in 2024. Direct-to-chip liquid cooling can cut direct water use substantially, but the electricity the chips consume still carries an upstream water and carbon footprint.
Sources
- International Energy Agency (2025). Energy and AI: Energy demand from AI. IEA report. Grade: PRIMARY; global data-center electricity 415 TWh (2024) to ~945 TWh (2030) and AI share of growth. Verified 2026-07-23. iea.org
- Goldman Sachs Research (2024). AI to drive 165% increase in data center power demand by 2030. Grade: INDEPENDENT; 55 GW (2023) to 122 GW (2030) demand and $720B grid spending. Verified 2026-07-23. goldmansachs.com
- Lenovo Press (2026). Lenovo Nvidia GB300 NVL72 Rack Scale AI Product Guide. Vendor documentation. Grade: PRIMARY; ~135 kW rack power, 72 Blackwell Ultra GPUs, liquid cooling. Verified 2026-07-23. lenovopress.lenovo.com
- Uptime Institute. Global Data Center Survey. Industry survey. Grade: INDEPENDENT; average rack under 6 kW, most below 20 kW, PUE ~1.5-1.6. Verified 2026-07-23. uptimeinstitute.com
- Google (2025). 2025 Environmental Report. Corporate sustainability disclosure. Grade: PRIMARY; ~8.1 billion gallons of water withdrawn in 2024, up ~28% year over year. Verified 2026-07-23. sustainability.google