We believed that the US was facing a major energy shortage problem for AI. The data says the opposite

To win the AI ​​race you need several things, but two are very important. The first, have the best technology and the best chips. The second, having enough energy to power those chips. The US has the first, but everything pointed to it having a major energy bottleneck. That is no longer so clear.

China has plenty of energy. The China’s strategic visionwhich once again has been investing in the energy field for decades, is bearing fruit and the country has considerable room for maneuver in terms of energy supply. That is a factor that seems to tip the balance in its favor: Jensen Huang, CEO of NVIDIA, already warned that China can win the AI ​​race. According to him, China has more flexible regulation and its companies have government subsidies for the energy their data centers need.

But the US has another philosophy. A deep study from the startup Epoch AI—responsible for FrontierMath AI benchmark— serves as a counterpoint to these pessimistic theories. In recent months we have seen how the US seems to have a real problem with the energy needed for AI data centers.

China has not stopped increasing its energy generation capacity, but the US has not for a simple reason: until now it did not need it. Source: Epoch AI.

However, Epoch AI explains that it is not that the US is not capable of creating more energy capacity: it simply has not needed it until now. While China has prepared for the future—even if that future does not come—the US has maintained a more conservative attitude: as long as there was no demand, it would not make any move. The immediate question, of course, is whether you can move it now or is it too late? And no, it doesn’t seem like it is.

Forecast of necessary energy capacity for data centers in the US until 2030 according to different scenarios. In the worst of all of them (pink color), almost 80 GW of capacity will be needed. Source: Epoch AI.

The demand is going to be huge. There is a reality: those ambitious plans to create more and more data centers throughout the US —with Project Stargate at the forefront—will cause data centers in the country to need between 30 and 80 GW of energy capacity in 2030. For those responsible for the study, it is perfectly possible that the US “gets its act together” – pun intended – and manages to increase its energy capacity. As? Various options.

The US has room for maneuver. In order to supply all that energy that all those data centers will theoretically need, there are several clear alternatives according to the Epoch AI study:

1. Natural gas: is relatively cheap and plants can be built quickly. There are three large companies that can cover this demand: GE Verona, Mitsubishi Heavy and Siemens. The plans of all of them point to a production of more than 200 GW in 2030. Even if they are not met, this supply (without being totally dedicated to AI) would already be an important part of the solution.
2. Solar energy: the other big part of the solution, especially because its costs have fallen drastically and because it is very, very scalable. We have already seen how the US has the capacity to install 1,200 GW solar for IA thanks to its deserts, but at the moment Big Tech does not dare to use them. Once again, estimates point to around 200 GW of installed capacity in 2030, but even if these expectations are not met, this infrastructure will also be a clear part of the solution.

Energy flexibility. The report also talks about a dynamic supply philosophy. Most of the time the US power grid is oversized for one simple reason: It is built to be able to supply power at peak peaks—like when everyone turns on the air conditioning—but most of the time there is plenty of power even to give to large AI data centers. This future infrastructure must be created with that same idea: oversized, but flexible.

And there are other alternatives. The country is turning to energy solutions that it thought were buried to power data centers. Among them are the fossil plants that were theoretically going to close but that are returning to operation due to the astonishing increase in demand. There is also talk of going to military solutions and even more unusual alternatives, such as energy under volcanoes. Not to mention, of course, the nuclear power plants and the small nuclear reactors (SMR) that are already being used by some of the Big Tech for your data centers.

Be careful with your electricity bill. The reality is that in the North American country data centers are growing faster than electrical infrastructure, and these facilities They are draining the country’s electricity. The situation is even causing electricity grid operators to ask be able to shut down data centers in times of high demand. And then there’s the other big side effect: AI data centers they are skyrocketing the electricity bill.

When starting up an AI data center, power costs a tenth of what chips cost. Source: Epoch AI.

There doesn’t seem to be a problem. Even with all those obstacles, Epoch AI’s conclusion is clear: “we doubt these challenges are significant enough to impede the scaling of AI.” In fact, they remember that what is actually expensive are the chips, not the energy, which represents a tenth of the investment in chips. The report concludes that China having an advantage is not necessarily true, and that the hypothetical US energy bottleneck “is much weaker than many people have indicated.”

Image | Andrey Metelev

In Xataka | Artificial intelligence has already reached nuclear power plants. And it’s going to change them forever