Here's a fun paradox: AI data centers are on track to double their power consumption by 2030. The solution a lot of very smart people are betting on? Fusion energy. The process that powers the sun, recreated on Earth, producing limitless clean electricity.
But here's the twist. The company seemingly closest to making it happen, Commonwealth Fusion Systems, is using AI to get there faster. Google DeepMind helps them control plasma. NVIDIA helps them simulate entire reactors. AI-powered image recognition catches defects on their magnet factory floor.
AI needs fusion. Fusion needs AI. They're basically building each other.
We sat down with Brandon Sorbom, CFS's Chief Science Officer and co-founder, on The Neuron AI Explained podcast to understand how this works. Then Brandon did a detailed Reddit AMA answering dozens of technical questions from the fusion community. Between the two, we got a remarkably clear picture of where fusion actually stands, how AI is accelerating it, and why the old joke about fusion being "always 30 years away" might finally be dead.
The TL;DR: AI Is Helping Build the Thing That Will Power AI
CFS has raised nearly $3B and has about 1,200 employees working to recreate the process that powers the sun, on Earth, in a machine the size of a natural gas plant.
The key breakthrough? A new type of superconducting magnet that lets them build fusion machines roughly 50x smaller than the massive international ITER project in France. Brandon's team at MIT figured this out in 2012, published the design, and thought... maybe we should actually build this.
Here's where AI enters the picture:
- Google DeepMind built TORAX, an open-source plasma simulator that lets AI systems learn directly from fusion physics. The goal: control 100-million-degree plasma in real time (humans adjusting knobs won't cut it).
- NVIDIA and Siemens are creating "digital twins" of the machine, speeding up simulations by 100x so CFS can plan experiments before running them.
- On the factory floor, AI-powered image recognition catches defects in superconducting magnets, with robots fixing issues automatically.
Brandon put the circularity perfectly: "We're using AI tools to build something that may eventually provide power to a data center." CFS already has a direct power purchase agreement with Google. The company helping them control plasma is also a customer waiting for the electricity.
The timeline: SPARC (their demonstration machine) targets first plasma in 2027. ARC (the commercial power plant in Virginia) aims to put electricity on the grid in the early 2030s.
Fusion's default state is off, meaning it can't melt down (unlike fission). The fuel is derived from water (deuterium) and lithium (for breeding tritium), with initial tritium supplied by existing fission reactors. And it produces zero carbon emissions. If CFS delivers, the electricity will power the very AI systems that helped build it. The flywheel is spinning.
Now let's dive into all that more in depth below.
The 30-Year Myth (And Why It Was Wrong)
"People have a right to be skeptical," Brandon told us on the podcast. Fusion has earned its reputation as the technology that's perpetually over the horizon. But the actual data tells a different story.
Since the 1960s, progress in what physicists call the "triple product" (temperature, density, and confinement time, the three things you need for a fusion plasma to self-sustain) actually slightly outpaced Moore's Law. That's the same Moore's Law that took us from room-sized computers to smartphones. Fusion was matching that pace.
So why didn't anyone notice? Brandon put it simply: "At every step along the way of Moore's law, you could sell those chips." A computer from 2005 was slower than one from 2015, but people still bought it. Fusion has a threshold effect. You need to cross a minimum triple product before you can generate net energy. Everything below that line, no matter how much progress it represents, looks like nothing happened.
"Even though there was a lot of progress, you didn't necessarily see it because there wasn't a product that came out of all of that progress," he explained. "That led to this perception of fusion being always 30 years out."
The other thing that happened: around the year 2000, the machines maxed out the magnetic field they could generate with existing superconductor technology. The only remaining knob to increase performance was size. That gave us ITER, the massive international fusion project in France, a reactor with a vacuum chamber you could "fit an elephant inside" (Brandon's words) that's cost tens of billions of dollars and taken decades to build.
The Magnet Breakthrough That Changed Everything
This is where the story pivots. While ITER was being designed in the late 1980s and 1990s using the best superconductors available at the time (which maxed out around 12 Tesla of magnetic field), a new class of material was discovered: high-temperature superconductors (HTS).
The name is a bit misleading."It's higher temperature than the previous ones, but it's all very cold." Brandon joked. You still need cryogenics, it's just not actually high temperature as far as us humans are concerned.
What HTS is is much, much higher magnetic field. And higher field means you can make the reactor dramatically smaller. How much smaller? When Brandon and his MIT advisor Dennis Whyte ran the numbers in a 2012 design course, the answer was stunning: roughly 50 to 100 times smaller in volume than devices like ITER.
The catch? In 1986, when HTS was discovered (winning the Nobel Prize the year it was announced, which almost never happens), you could only make single crystals a millimeter across. It took 30 years of thin-film fabrication R&D, borrowing techniques from the semiconductor and LCD screen industries, to figure out how to manufacture it as usable tape. By around 2010, the industry was producing meter-long strips. CFS bet that trend would continue.
"We said, people can make a meter now. They can make 10 meters, then a hundred, then a thousand." Brandon recalled. What if we had a fusion concept that removed the constraint of low magnetic field?
That bet became the ARC power plant paper, and the paper became Commonwealth Fusion Systems.
Where Things Stand Right Now
CFS has raised nearly $3 billion since it was founded in 2018 and grown to about 1,200 employees. Here's the current state of play:
SPARC is their demonstration machine, currently being assembled at CFS headquarters in Devens, Massachusetts. Its job is to prove the physics by achieving Q > 1, meaning more fusion energy out than heating energy in. The first magnet was installed earlier this year, with the second already following. First plasma is targeted for 2027, with Q > 1 to follow as quickly as possible.
Next up: testing the poloidal field (PF) coils, a completely different magnet architecture. While the TF magnets run at steady current, the PF magnets pulse, ramping current up and down to shape and control the plasma. CFS has already finished building the first full-scale PF magnet for SPARC and is gearing up to test it.
Brandon also shared a fun logistics detail: one half of SPARC's vacuum vessel was flown from Europe on an Antonov cargo plane (speed matters when you're racing the clock). The other half? Currently on a boat, because the timeline allowed for the cheaper option.
Brandon told us the thing that keeps him up at night isn't the science anymore. "At this point, it's not really the science and technology that keeps me up at night. It's those 10,000 unique parts. It's the execution challenge of having an organization bring all these things together in the right order."CFS maintains an enormous Gantt chart (an "integrated master schedule") with thousands of tasks. When one supplier slips a delivery, the whole schedule ripples. They have roughly 100 people dedicated to supply chain alone, many of them with technical PhDs who visit vendor sites to troubleshoot production bottlenecks.
One supply chain story stands out. When CFS launched in 2018, they estimated they needed a roughly 40x scaleup in the global HTS tape industry just to get enough material for SPARC. "Frankly, a lot of people laughed this off as impossible," Brandon wrote in his AMA. The company adopted an internal mantra: "no project waits on tape." It held. They worked so closely with suppliers that they effectively bootstrapped the industry alongside their own machine. Brandon says nothing on the horizon scares him as much as that early tape challenge did.
Now, ARC is the commercial power plant that follows SPARC. In his Reddit AMA, Brandon confirmed it's planned for Chesterfield County, Virginia, with a target of putting "watts on the grid in the early 2030s." The ARC physics basis papers have been submitted to the Journal of Plasma Physics and are currently in peer review.
A key design philosophy: even on SPARC, which isn't a commercial product, CFS refused to make design choices that wouldn't scale to a cost-competitive power plant. You can't cheat yourself forward on a system that's going to "cost a hundred times more than your final product." Brandon also said they believe ARC can eventually be cost-competitive with combined-cycle natural gas and renewables, on a non-subsidized basis.
How AI Is Actually Helping
This is where things get really interesting for our audience. CFS has three major AI collaborations, each tackling a different problem.
Google DeepMind: Controlling the Plasma
A fusion plasma is a collection of something like 100 billion billion charged particles, all bouncing around at hundreds of millions of degrees and interacting electromagnetically. Controlling this thing is, to put it mildly, a fast-moving target.
These plasmas are very wiggly, Brandon explained on our podcast. "The control is on pretty fast time scales... You can't have a human adjusting a knob. You really need a fast control system that's learning from the plasma."
CFS's collaboration with DeepMind centers on TORAX, an open-source tokamak transport simulator built entirely in JAX (Google's framework for high-performance machine learning). TORAX can simulate how energy, particles, and current move through a fusion plasma, and because it's built in JAX, it's fully differentiable, meaning AI systems can learn directly from the physics simulations rather than treating the plasma as a black box.
The practical application: building reinforcement learning systems that detect when a plasma is about to "disrupt" (the technical term for when the plasma goes haywire and extinguishes itself) and adjust the magnetic fields in real time to keep it stable. DeepMind has already demonstrated AI-powered plasma control on existing tokamaks, and the TORAX collaboration with CFS extends that work to SPARC-scale machines.
Importantly, Brandon stressed this is additive, not load-bearing. "SPARC is going to get Q greater than one with or without fast control. However, with fast control on top of the high magnetic fields, that's all additive." The AI makes a good system better; it's not papering over a design flaw.
In his AMA, Brandon elaborated that even before LLMs, researchers were using machine learning for disruption prediction. The new collaboration extends that to active, real-time plasma shaping.
NVIDIA + Siemens: Digital Twins of the Reactor
The second collaboration tackles a different problem: simulation speed.
Classical physics simulations of what's happening inside a fusion plasma require supercomputers and take enormous amounts of time. CFS, working with NVIDIA and Siemens, is building "surrogate models," AI systems trained on the results of slower classical simulations that can then make predictions orders of magnitude faster.
"The traversing of the space of options, you can greatly optimize by one or two orders of magnitude using AI techniques," Brandon told us. Think of it as a digital twin of the plasma. Instead of running every experiment on the real machine (which takes hours of setup and cooling between shots), you can explore the design space virtually.
MIT researchers have been leading the charge on these surrogate models, and CFS plans to use them to plan SPARC experiments before they run them, then feed real results back into the models to make them even more accurate. In his AMA, Brandon noted these models are also speeding up ARC's design process.
Factory Floor AI: Robot Quality Control
The third application is more grounded but arguably just as impactful. CFS built an actual magnet factory, with robotic assembly lines, to manufacture the 18 toroidal field magnets for SPARC. Each magnet is built from 16 near-identical sub-modules called "pancakes" (about 300 total).
They went from the first pancake taking months to produce to cranking out two per day at peak rate.
AI plays a role here too. Brandon described using ""Image recognition for some pieces of the process where you could either have a person sit and look at micrographs of a non-destructive scan, or you could have an AI image recognition tool trained on what a particular type of defect looks like." A robot then goes in and fixes the defect with minimal human intervention.
This is the kind of unsexy AI application that quietly changes everything. When you need 300 precision superconducting components and any defect could compromise a $3 billion machine, automated quality control isn't a nice-to-have; it's a necessity.
The Flywheel
Brandon captured the circular logic perfectly on our podcast: "We're using AI tools to build something that may eventually provide power to a data center."
CFS has signed a direct power purchase agreement with Google, meaning they have to deliver. The company that's helping CFS control plasma is also a customer waiting for the electricity.
But there's a deeper point here. When we asked Brandon what AI capability he most wishes existed, his answer was surprising: "Nobody has really applied these tools because there's a lack of hardware to apply them to." The constraint isn't better AI models. It's having a real fusion machine to train them on. SPARC will be that machine.
Once SPARC is running, CFS will feed real experimental data back into DeepMind's control systems and NVIDIA's surrogate models. That learning will inform ARC's final design and operation. The flywheel spins: AI helps build fusion, fusion data trains better AI, better AI builds better fusion.
But Will It Actually Be Affordable?
This is the question that separates fusion hype from fusion reality. Technically, JET achieved brief fusion in the 1990s. But nobody cares about fusion that costs a million dollars per kilowatt-hour.
Brandon makes a first-principles case for why fusion power can compete on cost. The physics works in fusion's favor: the fuel (deuterium and tritium) is millions of times more energy-dense than chemical fuels. Once you've built the plant, fuel costs are essentially nothing. Deuterium is filtered from ordinary water (roughly 1 in every 6,000 water molecules is heavy water). Tritium is bred from lithium, and the amount needed is tiny.
In his AMA, Brandon put a finer point on it: the lithium needed for tritium breeding works out to less than 20 electric vehicle batteries per year of ARC plant operation. (On the podcast, he'd estimated "about 100 Tesla batteries over the lifetime of the plant," which, depending on plant lifespan, roughly checks out.) Either way, fusion will be a tiny fraction of global lithium demand, even as it scales to a significant fraction of the energy market.
Then there's the blanket, arguably the most important piece of engineering between SPARC and a commercial power plant. Here's why.
Fusion burns two fuels: deuterium and tritium, both isotopes of hydrogen. Deuterium is easy; roughly 1 in every 6,000 water molecules is heavy water, and you just filter it out. Tritium is the hard one. It barely exists naturally on Earth. So where do you get it?
For the first plants, CFS will source startup tritium from existing CANDU fission reactors, which produce it as a byproduct. But after that, the reactor breeds its own. This is where FLiBe comes in.
ARC's blanket uses FLiBe molten salt (a mixture of lithium fluoride and beryllium fluoride) that surrounds the reactor chamber. When the fusion reaction fires, it produces neutrons that fly out of the plasma (they're neutral, so the magnetic fields can't hold them). Those neutrons slam into the lithium atoms in the FLiBe. That collision transmutes the lithium into tritium. You extract the tritium from the salt, pipe it back into the reactor, and burn it as fuel.
So FLiBe is doing four jobs at once: breeding tritium (closing the fuel loop), shielding components from neutrons, cooling the machine, and heating water to generate electricity. The blanket isn't just infrastructure; it's literally part of the fuel cycle.
CFS isn't starting from scratch on this either. The concentrated solar industry has decades of experience with corrosive molten salts, and national labs have done extensive work on molten-salt-cooled fission. CFS is adapting that knowledge, running lab-scale tests on salt chemistry and corrosion, and preparing to scale up to industrial-component testing.
Once the loop is running, the only external inputs are more deuterium (from water) and occasional lithium top-ups. In his AMA, Brandon put a number on it: the lithium needed for tritium breeding works out to less than 20 electric vehicle batteries per year of ARC plant operation. That's why he says the fuel is "effectively free."
The physical plant itself? "Roughly the size of a comparable natural gas plant," Brandon said. Steel and concrete, and not too many weird materials. Even the rare earth content in the superconducting magnets is modest: about 100 kilograms of rare earth material for an entire ARC power plant, spread as a thin layer across enormous lengths of superconducting tape.
And because fusion doesn't involve fissile materials, CFS is regulated like a particle accelerator, not a nuclear reactor. That's a massive cost advantage over fission. "A lot of the cost of fission is driven by the fact that you need to be very tightly regulated for safety," Brandon noted. Fusion's default state is off. If anything goes wrong, it just stops.
The Scaling Vision
In his AMA, Brandon laid out what the decades after Q > 1 could look like. To replace about 20-30% of global electricity generation, you'd need roughly 5,000 to 10,000 fusion power plants (each producing about 400 megawatts electric). That's a big number, but it's roughly proportional to the number of power plants that already exist globally.
"There's somewhere around 20,000 to 30,000 of these plants in the world already," he noted. We've built tens of thousands of airplanes, which are also complicated. And let's not forget: Humans have built all those power plants to begin with.
His long-term excitement extends beyond replacing existing generation. "History has shown that when humans have more energy available to them, they do really clever stuff with that energy," he told us. (this rhymes with the idea that there are no high income, low energy countries)
One application Brandon keeps coming back to: desalination. Tokamak-based fusion plants won't just produce electricity; they'll also generate significant process heat. You could use both to turn seawater into fresh drinking water. "Providing the world not just with clean energy, but with affordable, fresh drinking water could dramatically improve the lives of literally billions of people," he wrote in his AMA.
So, if CFS' system works, and fusion hits the grid before the 2040s and the infamous Limits to Growth collapse scenario kicks in, we'll not only be able to avoid civilization collapse, but also avoid any so called "water wars" that could precede or proceed it! My friend Augie always claims this is what we have to look forward to in the 2040s, so I like to point out ways to avoid it, because, y'know, no one wants a water war.
Also, you can see how both desalination and fusion plants, and fusion plants and AI data centers, both have some obvious synergies: fusion is powered by seawater, and could power desalination to turn seawater into drinking water. Fusion plants could power AI data centers, and could use the AI inside the data centers to help control the fusion power. Then, the desalination plants could create usable water for the data centers to use for their liquid cooling (since seawater is horribly corrosive for that use-case atm). Just wanted to point that out (and also maybe that one could consider buying coastal land so you can sell it to datacenter and fusion plant providers? But that's just a joke!).
Now, one other signal that CFS's magnet technology has legs beyond their own reactors: they've already licensed their HTS magnet technology to Type One Energy, a company building stellarators (a different fusion machine geometry). If CFS's magnets become the standard for high-field fusion applications across multiple reactor designs, that's a business model on top of a business model.
What to Watch For
Here's the timeline to bookmark:
- Now: PF coil testing underway (different magnet architecture from the TF coils already installed)
- 2027: First plasma in SPARC.
- After 2027: Campaign to achieve Q > 1, then Q ~ 11 (up to 140 MW of fusion power).
- After Q > 1: SPARC campaigns targeting Q~11, with up to 140 MW of fusion power (this is the real performance target, not just Q > 1).
- Early 2030s: ARC power plant puts electricity on the grid in Virginia.
- Coming soon: ARC physics basis papers published in the Journal of Plasma Physics.
- This spring: New video tour of SPARC construction (follow Alex Creely's updates).
The ARC power plant is designed as both a first-of-a-kind and pilot plant. It'll connect to the grid and generate electricity that customers actually buy.
A word of calibration: Brandon is notably more measured about AI's role in his AMA than in casual conversation. "It's probably too early to say exactly what the best and highest leverage use of AI will be," he wrote. When asked directly whether AI had delivered any "protein folding"-level breakthroughs for fusion, his answer was simply that CFS thinks AI "could be useful" and has "begun some work in that direction." The tools are promising. The results are still early.
Brandon also didn't shy away from acknowledging the risks of building something this complex. When asked about ITER's stress corrosion cracking issue (which delayed the project significantly), he called it "a sobering example of the complexity of these systems." His answer for how CFS manages that risk was practical, not hand-wavy: they picked tokamaks because over 200 have already been built worldwide, giving them a deep base of lessons learned. They hired people from adjacent industries (rockets, racecars, oil rigs, power plants) who bring "a healthy sense of paranoia." And they built internal systems with multiple checks on everything they build and test.
"No silver bullets, but a lot of hard work," he wrote.
Why This Matters
Fusion is one of those technologies where, if it works, it changes everything. Not in a vague, hand-wavy way, but in the concrete sense that you could replace every fossil fuel power plant on Earth with something that produces zero carbon, generates essentially no long-lived waste, can't melt down, and runs on fuel extracted from seawater.
The AI angle makes this story relevant right now, not in some abstract future. Tools that are familiar and/or adjacent to the ones you use every day (models built by Google, compute powered by NVIDIA) are being applied to one of the hardest engineering problems in history. And if CFS delivers, the electricity that flows back will power the next generation of those same AI tools.
It's the most elegant flywheel in technology. And the magnets are already going in.
To quote Brandon (quoting Top Gun) "I feel the need.... the need for speed."
This article is based on our interview with Brandon Sorbom on The Neuron AI Explained podcast and his Reddit AMA on r/fusion. For the full technical details, check out CFS's published research.
