The leader of OpenAI has proposed a grand benchmark: building enough new AI infrastructure every week to generate one gigawatt (GW) of computing capacity by the end of this year. Translated to the ground, that amount of capacity (based on current hyperscale layouts) is equivalent to about 60 football fields of data center floor space each week. The aspiration captures the spirit of the gold rush around AI, but can the physics, supply chains and power grids of the industry keep pace?
What ‘1 GW a Week’ Really Means for Data Centers
One gigawatt is the constant output of a midsize power plant. Add that up every week and it would put another 52 GW of load on a yearly basis — a figure that matches or exceeds many current estimates for the entire United States data center footprint. It would also mean 168 gigawatt-hours in new weekly consumption for each of the tranches that do end up going live.
OpenAI also has teamed with Nvidia on what the pair are casting as the world’s largest AI infrastructure deal. Even with deep-pocketed partners, the pace suggested by a gigawatt-per-week cadence would be an unprecedented feat in the history of digital infrastructure.
The Space, Power and Water Math for AI Data Centers
How did we arrive at “60 football fields”? Take a Texas campus in development that’s been called about 4 million square feet and around 1.2 GW of total capacity by its developer. That ratio works out to around 3.33 million square feet per gigawatt — approximately the area of 60 regulation football fields, not including setbacks, substations, cooling yards and logistics space.
These sites are sprawling. The New York Times has mapped individual campuses of more than 1,000 acres. Meta has talked designs on the scale of Manhattan if you sewed together several facilities. The land is only one part of the puzzle: modern AI centers consume vast amounts of energy and water. Globally, the average PUE remains well above 1.4, according to data from the Uptime Institute, which measures overhead power consumption beyond the chips. In hot climates, the use of evaporative cooling can lead to high water draws during peak months, as has been revealed in local filings and reporting.
Communities have noticed. National reporting has documented higher utility bills, 24/7 noise and stressed water systems near big builds. These worries have translated into more stringent permitting and longer timelines.
Chips, Racks and Grid Hardware Are Bottlenecks
Then, even if the land and capital can be found, the one remaining hard constraint is the silicon supply chain. Elite accelerators rely on advanced packaging, and capacity for such technologies — CoWoS among them — has been acknowledged by manufacturers as a chokepoint. A single gigawatt dedicated mostly to AI may be on the order of hundreds of thousands of accelerators if we include overhead power — orders of magnitude more than can be fabricated each week today.
Power delivery is another limit. National laboratories and the U.S. Department of Energy have cautioned about multi-year lead times for large power transformers and substation equipment. Interconnection queues at grid operators often stretch three to five years from the application file date until the system is energized, Lawrence Berkeley National Laboratory analysis found. You can’t just wish a gigawatt onto the grid next Friday.
The Money and Permitting Math Behind 1 GW Weekly
Prices for AI-optimized systems are skyrocketing. Real estate companies that monitor the sector pin current greenfield costs often in the $8 million to $12 million per megawatt range, when liquid cooling, high-density power and supply contingencies are factored in. At those numbers, one gigawatt of capacity could cost $8 billion to $12 billion or more, not including the price tag for buying land and long-term power purchase agreements. Hitting that outlay on a weekly basis would suggest annual capital investment in the hundreds of billions.
Permitting isn’t trivial, either. Local resistance over noise, diesel backup fleets and water rights has already slowed projects from Virginia to the Pacific Northwest, according to reporting by The Washington Post, NPR and other outlets. Even in supportive communities, the workforce is scarce for special trades.
What Could Make It Believable at Meaningful Scale
If you’re looking for a steadier ride, they’d need to make a few moves. The first is industrialization: repeatable, factory-made modules — prefabricated power rooms, liquid-cooling plants and rack blocks — that minimize complexity on site. Hyperscalers and integrators are creeping in this direction, but the ecosystem isn’t fully there yet.
Second is siting near ample generation and transmission. Co-location with new renewables projects, grid-scale storage, or nuclear uprates can shorten the distance to power. Advanced nuclear and fusion could possibly help, in the long run; OpenAI’s leadership has backed companies in those areas, and big tech firms have pursued new power purchase agreements. None of this obviates the multi-year development cycles.
Third is efficiency. Improved model architectures, pruning and quantization lower the requirement for compute per unit of capability. House-side benefits with direct-to-chip liquid cooling and smaller PUEs can trim some of the fat. These are upgrades that won’t erase the scale, but they help you drive a jalopy of aspiration closer to reality without embarrassing yourself.
Bottom Line: Slogan or Schedule for AI Buildout?
“One gigawatt a week” is a potent rallying cry for an industry sprinting to feed AI’s ravenous demand. And, on today’s foundations — chips, transformers, interconnection queues, water, power, labor and capital — doing it literally right now isn’t credible. And as a multiyear pipeline target that its modular build strategy, large power deals and efficiency gains justify, pieces of it might emerge. But 60 football fields a week, in the near term, is more metaphor than master plan.