AIFEATURE

Why AI Data Centers Devour Power: Reading the AI Infrastructure Energy Ledger in Gigawatts

N
NathanTechnology Editor · Technical Lead
Published · Updated
AI data centers now draw power at the gigawatt (GW) scale — power-plant-level demand — because clusters of thousands of high-wattage GPUs, plus cooling and power-delivery losses, push total draw far past traditional megawatt-class facilities. OpenAI's Ohio data-center project illustrates the scale, planned around 10 GW of new power supply to support 4.25 IT-GW of computing capacity. Facing this bottleneck, operators are turning to self-built power plants, natural gas, and nuclear power including small modular reactors, while advanced process nodes, dedicated inference chips, and liquid cooling aim to extract more compute from every watt available.

Why is AI data-center power now measured in gigawatts instead of megawatts?

AI data centers have crossed into gigawatt (GW)-scale power demand, a threshold once reserved for power plants, rather than the megawatt (MW) scale that defined traditional data centersCITE:E1. The root cause is the GPU cluster: training and running large models requires thousands of high-power AI accelerators operating simultaneously, each drawing anywhere from hundreds to over a thousand wattsCITE:E2. Add cooling and power-delivery losses on top of that raw compute draw, and total electricity demand climbs sharply as model size and deployment scale growCITE:E2. In effect, a modern AI cluster's power profile has shifted from "data center" to "power-plant customer"CITE:E1.

What do 2026 financing and capex figures reveal about AI power demand?

In 2026, AI infrastructure deals have made electricity a core bargaining chip rather than an afterthoughtCITE:E3. OpenAI's planned Ohio data center is a case in point: the project is built around 10 GW of new power supply to support 4.25 IT-GW of actual computing capacityCITE:E3. That gap between gross power secured and usable IT capacity reflects the overhead described above — cooling and delivery losses eat into the raw supply before it reaches the chipsCITE:E2.

MetricValueSource
New power supply planned (OpenAI Ohio project)10 GWCITE:E3
IT computing capacity to be supported4.25 IT-GWCITE:E3
Power draw per GPUHundreds to 1,000+ wattsCITE:E2

Hyperscalers' capital-expenditure disclosures point the same direction: cloud providers have raised their spending plans for data centers and power infrastructure, and industry observers now describe the AI compute bottleneck as shifting from chip supply toward power and land availabilityCITE:E4.

How is the industry responding to the power bottleneck?

Operators are moving to secure power directly rather than wait on grid capacity, turning to self-built power plants, natural gas generation, and nuclear power including small modular reactors (SMRs)CITE:E5. This has pulled the AI industry and the energy industry into a tightly bound relationship, where compute buildouts and power-generation buildouts are now planned together rather than separatelyCITE:E5.

How do process nodes, dedicated chips, and liquid cooling raise compute per watt?

With power itself now a constraint, efficiency has become as important as raw capacity. Three levers are in play: advanced process nodes that lower power draw per chip, dedicated inference chips built for higher energy efficiency, and liquid cooling to manage heat from denser GPU clustersCITE:E6. Performance-per-watt has consequently become one of the central competitive metrics in AI chip design, alongside raw throughputCITE:E6.

What this means

The numbers line up into a single story: AI power demand is now denominated in gigawatts because GPU clusters draw hundreds to over a thousand watts each at massive scaleCITE:E2, and a project like OpenAI's Ohio data center needs 10 GW of new supply to deliver 4.25 IT-GW of usable computeCITE:E3. That gap — and the broader shift of the compute bottleneck from chips to power and landCITE:E4 — is exactly why operators are now building power plants and nuclear capacity alongside data centersCITE:E5, and why performance-per-watt has become a headline metric in chip designCITE:E6.

📊 Evidence

FAQ

Why is AI data-center power now measured in gigawatts instead of megawatts?

AI data centers have crossed into gigawatt (GW)-scale power demand, a threshold once reserved for power plants, rather than the megawatt (MW) scale that defined…

What do 2026 financing and capex figures reveal about AI power demand?

In 2026, AI infrastructure deals have made electricity a core bargaining chip rather than an afterthoughtCITE:E3.

How is the industry responding to the power bottleneck?

Operators are moving to secure power directly rather than wait on grid capacity, turning to self-built power plants, natural gas generation, and nuclear power i…

How do process nodes, dedicated chips, and liquid cooling raise compute per watt?

With power itself now a constraint, efficiency has become as important as raw capacity.

📎 Sources

  1. en.wikipedia.org
  2. effectstory.com
  3. effectstory.com
  4. en.wikipedia.org
  5. en.wikipedia.org

Related data

Author's TakeNathan

The number worth sitting with is the 10 GW-to-4.25 IT-GW gap in OpenAI's Ohio project: less than half of the power supply being secured actually reaches the chips as usable compute, with the rest absorbed by cooling and delivery overhead. That gap is precisely why process-node efficiency, dedicated inference chips, and liquid cooling have become competitive battlegrounds rather than nice-to-haves — every point of performance-per-watt gained narrows the distance between gigawatts procured and IT-GW delivered. The metric to watch going forward is whether future project disclosures narrow this GW-to-IT-GW ratio, which would be the clearest evidence that efficiency gains are outpacing raw power procurement rather than just supplementing it.

N
NathanTechnology Editor · Technical Lead

Related

BRIEF

SpaceX and NVIDIA Plan to Launch Orbital AI Data Centers Starting Next Year, Musk Says

SpaceX plans to launch its first NVIDIA-chip-powered AI satellites, Starmind AI1, in the fourth quarter of 2027, reaching large-scale deployment by 2028 — at least a year ahead of the original schedule. The company has also filed with the FCC for a network of up to 1 million satellites, while Taiwanese suppliers Unitech and Sesoda report rising satellite-related revenue and order shares tied to the buildout.

林紀旭 James Lin ·
BRIEF

NVIDIA Launches Jetson Orin Nano 2 Robotics Computer, Targets Entry-Level Edge AI

NVIDIA announced the Jetson Orin Nano 2 robotics computer on August 25, 2026, packing 78 TOPS of AI compute, 8GB of memory, and an 8-core Arm CPU while doubling inference performance and cutting power draw 40% versus its predecessor. Modules and developer kits ship in the first half of 2027, targeting a robotics developer base NVIDIA says already tops 3 million.

Nathan ·