The Challenges of Fusion-Based Electricity for AI Datacenters

To get a sense of the energy needs of AI, look at the size of the datacenters Big Tech is building and at their common metric; not of bytes, but in the gigawatts of electricity that will drive them.

In late July, ChatGPT owner OpenAI and Oracle together said they plan to add 4.5 gigawatts of capacity to their joint AI infrastructure project called Stargate; an additional 500 megawatts (0.5 gigawatts) is already under construction in Abilene, TX. Japanese tech giant SoftBank is also a financial backer.

For perspective, the 5-gigawatt total is the combined capacity of about four large nuclear reactors. That is roughly the size of a power plant that would supply five million homes with electricity. OpenAI says Stargate will add a total of 10 gigawatts by 2029. Stargate participants also include Nvidia, Arm, Microsoft, and UAE-based investment firm MGX.

Stargate is just one of many examples. Amazon is building a 2.2-GW data site in Indiana which will house operations for AI company Anthropic. Meta is building a 2-GW plant in Louisiana.

These AI datacenters, with their industrial-scale energy appetites, are the steel mills of modern times. As reported in the first part of this two-part series on nuclear fusion as an energy source for AI, the U.S. Department of Energy (DoE) stated in late 2024 that “Datacenter load growth has tripled over the past decade and is projected to double or triple by 2028.” It made that assertion in its 2024 U.S. Datacenter Energy Usage Report, which noted that “Datacenters consumed about 4.4% of total U.S. electricity in 2023 and are expected to consume approximately 6.7% to 12% of total U.S. electricity by 2028.”

Big Tech is investing big dollars into fusion development companies, a few of which claim to be close to cracking fusion for commercial use. Microsoft has signed an agreement to receive fusion-generated electricity from Helion, an Everett, WA-based fusion firm.

Googling Fusion

Meanwhile, Google has focused its fusion eyes on Commonwealth Fusion Systems (CFS), Devens, MA, which spun out of MIT in 2018. In June of this year, Google followed in Microsoft’s footsteps and agreed to purchase fusion electricity from a plant that CFS will build in Chesterfield County, VA. (Google’s electricity contractually will come straight from that plant; for Microsoft, a middleman could source the electricity elsewhere).

“Fusion power is within our grasp thanks in part to forward-thinking partners like Google, a recognized technology pioneer across industries,” said Bob Mumgaard, CEO of CFS, in announcing the agreement in June.

The CFS/Google partnership has increased since then. In August, Google said it plans to spend $9 billion to build a new datacenter in the same county, and to expand two existing datacenters in two other Virginia counties. Around the same time, Google also strengthened its financial commitment to CFS, increasing its investment in the company as part of an $863-million funding round. Google first took a stake in 2021, as part of a $1.8-billion round. Neither Google nor CFS will reveal the size of Google’s stake.

Other tech players are backing CFS. Nvidia participated in the $863-million funding round, as did Japan’s NTT Inc., and financial services firm Galaxy Digital. Existing investors who re-upped in the round included Gates Frontier. Former Google boss Eric Schmidt was also among the backers.

CFS said the latest round will help it complete its demonstrator machine called SPARC in Devens, MA, and to start on its commercial fusion plant in Virginia, called ARC.

CFS is building a tokamak, a classic approach to fusion and one which the ITER project is pursuing in Cadarache, France. A distinguishing characteristic of CFS’ tokamak is the use of what CFS claims are the world’s strongest magnets; theoretically strong enough to lift an aircraft carrier. The magnets in a tokamak surround the donut-shaped structure to help confine a plasma of deuterium and tritium until they fuse and release neutrons for heat that drives a turbine. Because the magnets are so strong, CFS says the tokamak can be much smaller than classic tokamak designs, such as the roughly 100 foot by 100 foot tokamak planned by the repeatedly delayed ITER.

Material Matters

In a conversation with Communications, a Commonwealth spokeswoman characterized SPARC as 65% complete and said the company hopes to demonstrate it in about two years. At that point, SPARC will not include at least one crucial part: the internal lithium blanket that will absorb neutrons and collect the heat to be transferred to a steam turbine.

The blanket challenge is a common one for any fusion developer taking the thermal approach of using fusion heat to drive a turbine, the strategy for the majority of fusion projects. The blanket not only collects heat that is transferred for turbine use, it also protects the magnets on the outside of a tokamak. And, crucially, the lithium in the blanket combines with neutrons to form the tritium that allows for a continual deuterium-tritium fusion process. Given the extreme 100-million-degrees C and higher temperatures needed for fusion, it’s no surprise that fusion engineers are still looking for a blanket material that will stand up to the neutron bombardment.

“Right now, it is a challenge across our entire field, finding sufficient material to even serve as that blanket, or first wall,” said Tammy Ma, who leads the inertial fusion energy program at the DoE’s Lawrence Livermore National Laboratory’s National Ignition Facility (NIF) in Livermore, CA. “Because the conditions are so extreme, the temperatures are so high, you’re being bombarded by all kinds of energetic particles; neutron damage is enormous. We have yet to find materials that can withstand these materials for very long. Maybe it can last a year, or five years, but if you’re trying to build a commercially viable power plant, you can’t be replacing those components every couple of years.”

NIF’s fusion machine is not a tokamak. Rather, it’s shooting 192 lasers from a mile away at a tiny pellet of deuterium and tritium to coax them into fusing and maintaining a reaction. The approach is called inertial confinement. (ITER’s is called magnetic confinement; Helion’s is called magneto-inertial.)

Ma noted that other general engineering and safety challenges remain across the fusion industry. Those include approaches for dealing with radioactive waste. Unlike fission, fusion does not leave the sort of high-level radioactive elements that last for thousands of years, but it does leave waste that has to be managed over the 50-plus years during which it decays. Ma said the industry also needs to develop robots to maintain fusion reactors. And electricity grid operators will have to figure out how to mix fusion energy in with other sources, she said.

Money Matters

One thing could hasten the completion of these monumental tasks.

“It’s very much a function of funding,” observed Ma. “It’s mostly funding. Time is money.”

Big Tech is now pouring in big development dollars. With rarefied valuations, they also appear to have the wherewithal.

“Their investments in fusion energy are relatively small for the potentially enormous payoff of fusion energy,” Ma said.

Another aspect to the convergence of IT and energy production is fusion companies’ use of AI to expedite their own research and development. NIF is working with OpenAI and Nvidia, for example.

“It’s a very symbiotic relationship,” noted Ma. “We need AI to make fusion energy happen, but we need fusion energy to make AI go. It should be a very strong partnership.”

Mark Halper is a freelance journalist based near Bristol, U.K. He covers everything from media moguls to subatomic particles.