Can We Build a Five Gigawatt Data Center?

We access AI through our screens as part of the ephemeral digital world that makes up the internet. But AI doesn’t actually live in the cloud. 

AI lives in data centers. Tens of thousands of computers racked in rows hum away in buildings the size of several football fields. It's louder than you would expect. Industrial fans push in a constant breeze of chilled air, funneling away the waste heat. Thick bundles of fiber optic cables snake along ceiling tracks like mechanical veins, carrying unfathomable streams of data. 

This is a 100-megawatt data center — a facility that consumes as much electricity as a small city, all to keep the digital world spinning. It’s impossible to say exactly how many exist today — companies prefer to keep their data centers private — but estimates put the number of hyperscale data centers worldwide at over 1,000. Yet, despite their footprint, they are already being outclassed in the constant need for more compute capable of training future generations of AI. We have reached the point where, if we don’t build bigger centers soon, tech giants will be forced to stretch training runs over multiple years. In short, AI has already outgrown its starter home.

GPT-4 was reportedly trained with 30 MW of power. Forecasts predict that in the next five years large training runs will require 5-gigawatt data centers ¹ — facilities at least 10 times the size of today’s largest data centers. That is roughly the average energy needed to power all of New York City. 

Tech leaders seem confident that they’ll be able to build centers of a size that even a few years ago would have seemed unprecedented. Mark Zuckerberg said a 1-GW data center is “only a matter of time.” Meta has broken ground on their largest data center yet, where they hope to bring one gigawatt online in 2025. Microsoft and OpenAI are reportedly planning a 1- to 5-GW “Stargate” facility supposedly launching in 2028. Sam Altman even pitched the White House on the construction of multiple data centers that each require up to 5 GW. Tech, in short, is betting on a YIMBY future for AI training. 

But how realistic are these plans? For all of the big talk, most GW proposals are still in the planning and permission stages. And actual sites in the United States that could support 1-GW — let alone 5-GW — projects are scarce for several reasons.

To understand those reasons, let’s first go over the constraints that would need to be overcome to build a GW-plus data center. 

Physical space

A data center is essentially a large warehouse. Each warehouse — the largest of which reach over 1 million square feet, or more than 20 football fields, in their footprint — can work together with nearby warehouses (approximately 20 miles) to form a “compute cluster.”

Scaling compute clusters — either through adding more chips or increasing the distance between data centers — poses problems, especially for AI training. Whereas an average data center may be running many independent tasks (e.g., video streaming or web hosting), an AI training center serves one purpose. Since all the chips need to constantly share and sync data with each other, it helps if they’re close. According to Konstantin Pilz, a researcher at the Special Competitive Studies Project,  “If you want to train efficiently, these chips have to talk to each other at a very high rate and with a very high bandwidth. The more chips you have, the more space you need. So it makes it more complicated to connect them efficiently.” Research is currently being done on “distributed training,” which would ease this bottleneck, but the current large-scale training runs have to happen mostly within a small physical distance.

Money

Microsoft and OpenAI’s proposed Stargate facility reportedly could cost up to $100 billion, and North Dakota was recently approached by two companies about developing $125 billion data center clusters that could scale to 5-10 GW. 

These numbers seem reasonable compared with naive scaling of current costs. It currently costs around $1 billion to build a 100-MW center, which implies $10 billion for a 1-GW center. However, this is just for the facility. The chips to operate such a center are estimated ² to cost another approximately $20 billion, for a total of $30 billion. Naively, this would imply a cost of $150 billion for a 5-GW center. 

And that’s just the cost for the data center itself. Operating costs are estimated to run into the millions of dollars per MW per year, mainly due to the large amount of power consumed. Additionally, Epoch AI estimates that it would cost about $3.5 billion to build a big enough gas power plant to run a 1-GW data center. 

Can anyone afford such a high price tag?

Last year, the capital expenditures for the hyperscalar tech giants — Google, Meta, Microsoft, Amazon — ranged from $28 billion to $55 billion. Those numbers are climbing. This means that a $100 billion data center built over five years would take half to two-thirds of any of the tech giants’ current capex altogether. 

We shouldn’t assume that’s the entire pot, though. Dedicated data center funds have been announced already. The expanded Stargate program claims it will invest $500 billion over the next four years, though there have been uncertainties whether the backers can raise the promised funds. The new Global AI Infrastructure Investment Partnership (GAIIP) between BlackRock, Microsoft, Global Infrastructure Partners, and MGX claims it plans to spend $80 billion to $100 billion to build data centers. Meanwhile, Bloomberg reports that Amazon has committed to spending $150 billion on data centers over the next 15 years (though mostly on many already-planned smaller projects).

If we take these funds at face value, then the Stargate Project could fund multiple 5 GW data centers. Amazon and GAIIP could each probably fund a single center, if that was all they attempted to do. And any of the companies could do so if they were willing to devote half of their capex to a single project. 

If the cost turns out to be lower, those funds could go even farther. If we assume a 5-GW data center requires only $25 billion, the combined $250 billion from Amazon and GAIIP could fund 10. However, this assumes business as usual. Additional investment from companies, foreign investors, or the US government could substantially increase large data center construction capacity over the next five years. 

Chips

Currently, H100 chips — each of which carries 80 billion transistors and costs some $25,000 — are state of the art for AI. The Llama 3.1 405B model used a cluster of 16 thousand H100s for its 4e25 FLOP training run. In contrast, Epoch predicts that 20 million H100-equivalent GPUs will be required to power the 2e29 FLOP training runs forecast to be required at the end of the decade — a thousand-fold increase from today.

Chip production is bottlenecked on a few components. The most important are the subcomponent high bandwidth memory (HBM) chips and packaging capacity to assemble the transistors and memory chips onto the base wafer. These require both expensive infrastructure and high-skilled workers, making it difficult for production to rapidly scale. 

It’s difficult to predict how many H100 equivalents will be produced by 2030 given these limits. Epoch’s median estimate is 100 million, but with a confidence interval of 20 to 400 million. Competition heightens the need for chips, as each company will need its own to power a large training run. Still, it’s not unrealistic to assume that a single company is able to obtain 20% of the available chips — enough to support a training run of the size predicted for 2030. Chip production would have to fall at the extreme low end of predictions for chips to become the biggest constraint. 

Hence, while chips are theoretically a bottleneck, it is more likely that power availability will be a bigger problem for companies seeking to build even bigger data centers. 

Power

AI chips need steady power available around the clock to work. Typically, a transmission line connects the data center to a nearby power plant, which supplies energy. To avoid costly service outages, data centers also include backup systems such as power generators. 

Power is the key choke point in building larger data centers. Availability of power reportedly “comes up in every conversation” as a major concern of AI company leaders, according to Brian Potter, senior infrastructure fellow at the Institute for Progress. Mark Zuckerberg claimed Meta would probably “build out bigger clusters than we currently can if we could get the energy to do it.”

While data centers currently utilize only about 1% of overall electricity, semiconductor and AI research organization SemiAnalysis predicts that “AI will propel datacenters to use 4.5% of global energy generation by 2030.” In areas that host large numbers of data centers demand will of course be vastly higher. For example, data centers currently use 18% and 24% of electricity in data center hubs in Ireland and Virginia, respectively. The United States produces only around 1,200 GW in total, so a 5-GW facility or even a 1-GW facility would consume a large fraction of the local energy wherever it was built. 

Or you can look at power requirements from another perspective. If a company wants to perform a 6-GW training run (Epoch’s estimate required power for 2030), they would need to build a power facility comparable to the largest coal power plant in the world and dedicate it solely to the data center. Securing that much available power in one location is — obviously — no easy feat. 

Still, tech companies seem confident they can get enough power for the foreseeable future of training runs. They have two strategies available to them: draw more energy from the power grid or build on-site power. Are they realistic?

From the grid

Major electricity consumers secure power through long-term contracts, so it’s likely that most of the available power, at least in the short-medium term, is already locked down. This leaves fewer sites that could support new large-scale demand, though it’s possible that data centers could buy out existing contracts. 

However, power companies are expanding their production capacity. Epoch predicts that Northern Virginia will increase data center power capacity from 5 GW to 10 GW by 2030. Dominion Energy, the sixth-largest provider in the United States by revenue, says it has received customer orders that could double the amount of data center capacity in Virginia by 2028, with a projected market size of 10 GW by 2035. Texas is also increasing energy production to meet a spike in demand: One company, Oncor, reported receiving 59 GW of new service requests just from planned data centers. The total projected load in the region is projected to peak at approximately 152 GW by 2030, though it’s uncertain how much of that will go to data centers. 

A 5-GW site would consume the entirety of predicted new power in Virginia, which means that predicted expansions sound unlikely to support anything bigger than a handful of 1-GW data centers. A utility company spokesperson agreed, telling Bloomberg that “finding a site that could accommodate five gigawatts would take some work, but there are places in the US that can fit one gigawatt.”

Additionally, data centers would typically require transmission lines installed to connect the power plant to the center. Transmission lines typically take around 10 years to build, and investments in new construction are currently dropping. (As I’ll address later, transmission line construction is hampered by significant permitting bottlenecks, which make routing energy to the data center a significant challenge.) 

Building power plants on site could potentially secure significantly more on-site energy and avoid transmission line construction delays — as long as tech companies are willing to plan three to five years in advance, shell out a decent amount of money, and get a bit lucky with permitting. Epoch estimates that building on-site power would add 20%-40% to the cost of purchasing GPUs.

But timelines might be tight for building power by 2030. While there are several options, in theory, for building on-site power, most come with significant drawbacks that would limit their feasibility in the next five years. 

Nuclear power plants 

Traditional nuclear power plants take over a decade to build, so if we’re looking only at the next five years, they’re out. The most recent nuclear construction in the United States, Unit 4 of the Vogtle Electric Generating Plant in Georgia, began construction in 2013 and came online in 2024. As of this writing, there are no full-scale nuclear plants currently in progress in the United States.

There is another nuclear option: Hyperscalers could pay to reopen previously closed plants. Most notably, Microsoft is reportedly paying to reopen Three Mile Island, which can generate 835 MW. For the reported $1.6 billion to upgrade the facility, the deal would be substantially cheaper than building a new power plant from scratch. However, the number of other sites amenable to reopening is perhaps as few as three.

Small modular nuclear reactors 

Companies are also investing in small modular nuclear reactors, the more quickly constructed cousins of traditional nuclear power plants. Oracle recently announced they were building a data center larger than 1 GW that will be powered by three small nuclear reactors. Amazon and Google are both investing in early stage development of additional reactors. 

However, the technology is still very early stage. Only a couple modular reactors have come online, despite more than 80 commercial SMR designs currently being developed, according to IFP. There’s still a ways to go before these will be a reliable source for scaling data centers, though some experts think the first ones could come online around 2030.

Solar and wind

Solar and wind power generators can be built in large quantities relatively quickly, but their major downside is intermittent supply. Data centers require extremely reliable energy to prevent costly outages, so depending on solar or wind would also require extensive use of batteries. Tim Fist, senior technology fellow at IFP, says that simply isn’t feasible yet. To maintain reliability during low-output periods, a data center would need to install battery storage far in excess of its power requirements. At that scale, they would be limited by battery supply chains and extensive cost. 

According to one company’s estimate, a 1-MW battery setup would cost in the range of $775 to $1,400 per kW, and so naively might cost upwards of $1.4 billion per GW. However, the need for extended periods of reliable power means the facility would require many times the energy to have a stable backup supply. At a conservative estimate of 10 times operating power, the batteries alone would cost around $10 billion for a 1-GW facility. 

With that, a data center could try to run partially on solar or wind, but it would still need energy from other sources to step up when those aren't available. In practice, this means that a data center would still need enough power to run with no solar or wind energy available at all. While this could be desirable for reducing carbon emissions, it doesn’t solve the issue of building sufficient power in one location. 

Gas 

All of this brings us to the most common form — at approximately 40% — of energy in the United States. Gas power plants are relatively quick to build. The construction is routine, they have established supply lines, and modular gas turbine generators are readily available on the market. The permitting requirements are also less cumbersome than with many other power types. ³  

There is also likely abundant natural gas to power new data centers: America is the world’s largest producer of natural gas, though increased demand may require more drilling.

An IFP report estimated that a facility using natural gas with carbon capture would take a minimum of three years to build. Other estimates for large-scale facilities put construction time at five years or more.

The West County Energy Center, a 3.6-GW natural gas facility in Palm Beach, took just five years to finish from the time it received its permit. This is strong evidence that it’s feasible to build a >3-GW power plant in only five years, but it still remains an exception — only one other facility of similar size has been built in the past 25 years.

Overall, natural gas power plants seem like the most likely primary power source for a 5-GW data center. And indeed, Goldman Sachs, for instance, predicts that natural gas will provide 60% of the expanded power demand for data centers by 2030.

Regulation

The United States has the physical ability to generate more power and build out transmission lines over the next five years. It’s not happening because there are permitting and regulatory bottlenecks at each step of the process: building power plants, building transmission lines, and building the data centers themselves. As Pilz told me, “My sense is that, really, these permitting challenges are the main reason why the US isn't building power plants much faster.” 

Yusuf Mahmood, a RAND researcher studying permitting, agreed, saying, “We have the capacity to build up the grid. We have the capacity to increase power. … However, we have a lot of permitting restrictions that make this really difficult.” 

But just what exactly is required to get a data center up and running?

Permits for building power

Permits for building power plants can be extensive. Atop all of the local and state approval is the National Environmental Policy Act (NEPA). NEPA review is required anytime the federal government allows infrastructure projects, and the review process alone can take years — or even decades — to complete. Then there are states and local permitting on top of that. Sometimes local communities can also block construction. 

Would-be builders often apply for permits at several sites at once because they are counting on many of those permits being denied. Then, they build a fraction of these sites where the permits are granted. (Predictions of solar capacity, when based on permits filed, have been consistently overoptimistic — less than half those permits are granted.) 

Power sources like solar, which covers more land area, are also likely to run into review issues with the Endangered Species Act. For example, Meta’s attempt to power a data center with a nuclear plant recently fell through after a rare species of bee was discovered nearby. 

Buying access to existing power also isn’t necessarily easier. Regulators recently blocked Amazon’s attempt to purchase a data center site with a power contract for nearly 1 GW from the nearby nuclear power plant. 

Transmission lines

It’s more common to hear objections to transmission lines than to the data centers themselves. Transmission lines cross many jurisdictions, and building one takes an agreement from all of the individual properties that the line will run through, compliance with state and local ordinances, and, if building across state lines, approval from the Federal Energy Regulatory Commission (FERC). And there can be huge numbers of local landowners involved: One transmission line between Kansas and Indiana has been stalled for over a decade because of holdouts among the 1,700 landowners it would impact.

Permits for running data centers

Apart from the power sources, data centers themselves must go through permitting review, which often isn’t simple. For example, the Clean Air Act is an environmental regulation intended to restrict the release of pollution. However, the Environmental Protection Agency hasn’t given clear guidance on how states should interpret the Clean Air Act, causing some to regulate minor polluters such as data centers similarly to how they regulate major polluters. For example, xAI recently ran into trouble with their new data center in Memphis over a dispute regarding what permitting they are required to obtain under the Clean Air Act. 

If we stick to the current course, available funding, chips, and energy suggest that the number of possible sites for a 1-GW data center by 2030 will range from a handful to a few dozen. It’s possible we will see at least one 5-GW facility, but it’s highly unlikely we see more than 3 — and it’s not certain we’ll have any. 

But there are numerous unknowns that could change this trajectory. 

First, we might no longer need big data centers. If the AI bubble bursts and demand plummets, then we might not even get 1-GW centers. Or, if we master distributed training, we’ll no longer require such large data centers, because training could take place across several data centers drawing from different power sources. 

Second, technological breakthroughs could change many of the above calculations. Breakthroughs in modular nuclear plants, cheaper or more effective batteries, cheap carbon capture, or advanced geothermal could make it substantially easier to build sufficient clean power for large data centers. 

Third, we could see significant escalation of investments in AI from a variety of sources, including the US government, foreign investors, or the companies. The coalition of backers (including Japanese and Middle Eastern funders) for the Stargate Project indicate that these options are being actively pursued. Fist suggested that the government could support data center development, for example by authorizing additional appropriations for the Department of Energy from Congress or providing better access to early stage financing for clean energy projects. 

Fourth, hyperscalers could build data centers abroad. The UAE and Saudi Arabia have already indicated their interest in building such facilities, and new US government export regulations might allow that. While there are significant concerns within the US government over relinquishing physical control of compute, increasing compute pressure might force compromises. For example, US data capacity could focus on training runs so that the United States maintains control of models while allowing data centers abroad to power inference.

Fifth, and most importantly, permitting reform could remove the most significant bottlenecks to building more power in the United States. If the United States wants to ensure that substantial AI compute power remains within its jurisdiction, permitting reform may be its best bet. For example, the United States could standardize permitting requirements for minor sources of air pollution under the Clean Air Act or reform NEPA. According to Mahmood, if the government really wanted to pull out all the stops to powering data centers in the United States, a combination of executive and legislative action could dramatically speed up the process. How the new administration might act is anyone’s guess.

On our current trajectory, the United States isn’t likely to have the power, funding, or permitting required to meet the forecasted AI data center infrastructure demands, though there will likely be at least a few 1-GW data centers. But with permitting reform and substantial investment, we may yet see the first 5-GW data center in America by 2030.