You’ve probably heard that AI uses loads of electricity. Every time you ask a chatbot a question or stream a new game, a data centre somewhere is working hard. Now imagine powering those servers the way stars do. That’s the promise of fusion for data centres—but how close are we, really?
Why data centres need steady, clean power
Data centres don’t sleep. They prefer energy that is always there, not just when the wind blows or the Sun shines. Batteries help, and so does flexible demand, but the dream is a steady, low-carbon supply. Today, nuclear fission already provides reliable, clean power. Fusion is different: instead of splitting heavy atoms, it aims to join tiny hydrogen isotopes together to release energy—like a miniature star in a box. In theory, fusion could offer vast amounts of energy with no climate-warming emissions during operation and minimal long-lived waste.
Fusion in one minute: pellets, magnets, and other clever tricks
There are two big routes:
- Inertial confinement fusion (ICF): super-powerful lasers squeeze a pepper-grain-sized fuel pellet so hard and fast it fuses.
- Magnetic confinement fusion (MCF): tokamaks and stellarators use strong magnets to hold a super-hot plasma in a doughnut shape long enough for fusion to happen.
Some startups try other ideas, like field-reversed configurations (think of a plasma “smoke ring” held by magnets) or projectile-driven compression. Key word to know: gain (Q). If Q is above 1, the fusion output beats the energy put into the fuel. That’s exciting, but turning that into electricity on a wire is a much bigger engineering step.
What actually changed in 2025?
Laser fusion kept improving. At the U.S. National Ignition Facility (NIF), researchers reported a February 2025 shot that delivered about 5.0 MJ from 2.05 MJ of laser energy—a new target-gain record (~2.44)—and followed earlier multi-megajoule shots in 2024.
Meanwhile, Big Tech didn’t just watch. Microsoft signed the world’s first fusion power-purchase agreement with Helion (announced in 2023) and, in 2025, Helion started building a pilot plant in Washington state aimed at delivering power later this decade.
And in June 2025, Google announced a deal to buy 200 MW from Commonwealth Fusion Systems’ planned ARC power plant in the early 2030s—another signal that major companies want fusion in their long-term energy mix.
From lab sparks to plug sockets: the hurdles
Even with record shots, a power plant is a different beast. Here’s why:
- Do it again. And again. A power station needs many shots per second (ICF) or hours-long steady plasma (MCF). Lab wins are still occasional.
- Efficient hardware. Lasers must become far more efficient and durable. Targets (those tiny pellets) must be made by the millions, cheaply.
- Turn heat into electricity. You need robust systems to capture heat, run turbines, and handle materials that face intense particles and radiation.
- Fuel supply. Many designs use tritium, which needs to be bred inside the plant using lithium blankets. That cycle has to work reliably.
- Cost and uptime. Data centres need high availability and predictable prices. Early fusion plants may be expensiveand need lots of maintenance.
So… when could it power your apps? Realistic timelines
Short answer: not tomorrow—but not sci-fi either.
- Late 2020s: You might see demonstration units producing short bursts or small amounts of electricity to prove the concept and connect to a grid—useful for learning, not for running a hyperscale site.
- Early 2030s: If demos go well, a first wave of pilot plants could deliver limited, firm power. Big Tech deals suggest some companies hope to test fusion alongside other clean sources.
- Mid-2030s to 2040s: With several successful pilots, early commercial plants could supply a slice of data-centre demand in certain regions. Scaling beyond that will depend on cost, reliability, and regulation.
Here’s the honest take: fusion is moving faster than many expected, but it still faces tough physics and engineering. Data centres will likely run on a mix—wind, solar, grid-scale batteries, conventional nuclear, maybe small modular reactors—and, if milestones are hit, fusion joining the team later. Next time you use an AI app, ask yourself: if your favourite tech ran on star-style power, what trade-offs—cost, location, reliability—would you accept to make it happen?