The race to build larger AI data centers is becoming a race over money, power, and physical infrastructure. A new study from researchers at Georgetown, Epoch AI, and Rand suggests that, if recent growth patterns continue, the leading AI data center could become a project on the scale of a major city electricity system.
The study looked at AI data center growth around the world from 2019 to this year. Its dataset covered over 500 AI data center projects and found that computational performance, power requirements, and capital expenditures are all rising quickly.
The scale is moving beyond ordinary data center planning
The central finding is simple: the biggest AI data centers are growing fast enough that the old assumptions around cost and electricity may no longer fit. The study found that computational performance is more than doubling annually, while power requirements and capital expenditures are also more than doubling.
That matters because AI data centers do not scale only as software businesses. They require chips, buildings, grid connections, water, land, and financing. The more ambitious the model training and AI service demand becomes, the more the underlying physical system has to expand with it.
By June 2030, the study projects that the leading AI data center may reach three striking thresholds:
- 2 million AI chips
- $200 billion in cost
- 9 GW of power, roughly the output of nine nuclear reactors
Those figures describe a possible future, not a guarantee. But they frame the size of the infrastructure challenge now facing companies that want to build and operate frontier AI systems.
Costs and power needs are rising together
The study highlights xAI's Colossus as an example of where the industry already is. Colossus has a price tag of around $7 billion and draws an estimated 300 megawatts of power, as much as 250,000 households.
For AI data centers like xAI's Colossus, hardware costs increased 1.9x each year between 2019 and 2025, according to the study. Over the same period, power needs climbed 2x annually.
This pairing is important. A more expensive AI data center is not only a financial commitment; it also becomes an electricity demand problem. If the largest facilities continue to require more chips, they also require much larger and more reliable power supplies.
That is why AI infrastructure is now tied closely to energy planning. The limiting factor may not be only whether companies can buy enough AI chips. It may also be whether they can secure enough power, connect to the grid, and do so in locations where the environmental and public finance tradeoffs are acceptable.
Efficiency gains are real, but may not be enough
The study does not say the industry is standing still on efficiency. In fact, it found that AI data centers have become much more energy efficient in the last five years.
One key measure, computational performance per watt, increased 1.34x each year from 2019 to 2025. That means each watt of electricity is delivering more computation than before.
But the study's concern is that efficiency is being outpaced by growth. If the total amount of computation demanded keeps rising faster than efficiency improves, overall electricity use still climbs. In plain terms, better chips and systems can reduce waste, but they do not automatically shrink the total power footprint when the industry keeps building larger facilities.
This is the core tension in AI infrastructure. Technical progress is helping, yet the scale of ambition is growing even faster. The result is a sector where efficiency improvements and higher power demand can happen at the same time.
The pressure extends beyond the grid
Electricity is the most visible constraint, but the source article points to several other issues around AI data centers. These projects can involve high water consumption, occupy valuable real estate, and affect state tax bases.
A study by Good Jobs First, a Washington, D.C.-based nonprofit, estimates that at least 10 states lose over $100 million per year in tax revenue to data centers because of overly generous incentives. That turns the AI data center buildout into a public finance question as well as a technology question.
There is also the energy mix to consider. A recent Wells Fargo analysis forecasts that data center energy intake will grow 20% by 2030. The source article notes that this could put pressure on renewable power sources that depend on variable weather and could spur a ramp-up in non-renewable, environmentally damaging electricity sources like fossil fuels.
For local governments and utilities, the issue is not just whether a data center can be built. It is whether the surrounding infrastructure, tax policy, and environmental resources can absorb the project without creating bigger costs elsewhere.
Expansion is not guaranteed to continue unchecked
The study's projections depend on current trends holding. The source article also notes that the time scales could be off, or that the projected growth may not fully happen.
There are already signs of caution. Some hyperscalers, including AWS and Microsoft, have pulled back on data center projects in the last several weeks. In a note to investors in mid-April, analysts at Cowen observed a "cooling" in the data center market in early 2025, pointing to concern about unsustainable expansion.
At the same time, major AI infrastructure commitments remain large. OpenAI, which recently said that roughly 10% of the world's population is using ChatGPT, has a partnership with SoftBank and others to raise up to $500 billion to establish a network of AI data centers in the U.S. and possibly elsewhere. Microsoft, Google, and AWS have also pledged major spending to expand their data center footprints.
The picture, then, is not a straight line. The industry is pushing toward bigger AI data centers, but the economics, electricity demand, and public impacts are forcing harder questions. By June 2030, the leading AI data center may show whether AI infrastructure can keep scaling at its recent pace, or whether physical limits begin to shape the next phase of AI development.