The Tech Buzz may earn a commission when you buy through links on this page. This never affects which products we recommend or what we say about them.<\/p>\n
A new startup called Fusionality is quietly building the software backbone that could help fusion power finally reach the grid. Founded by alumni of Google<\/a> DeepMind, the company is developing AI-powered control systems and simulation environments designed to help fusion reactor developers skip years of trial-and-error engineering, according to an exclusive report from TechCrunch<\/a>.<\/p>\n Fusion power has always had a timing problem. Scientists have chased the promise of clean, near-limitless energy from atomic fusion for decades, and the joke inside the industry has long been that commercial fusion is perpetually ‘thirty years away.’ Now a new startup called Fusionality thinks it can shave real time off that clock, not by building a reactor itself, but by building the software that every reactor needs to actually work.<\/p>\n According to TechCrunch’s exclusive reporting<\/a>, Fusionality was founded by alumni of Google<\/a> DeepMind, the AI lab famous for teaching machines to beat world champions at Go and fold proteins with AlphaFold<\/a>. The founders are now pointing that same machine learning muscle at plasma physics, developing control systems and simulation environments meant to help fusion power startups iterate faster on one of the hardest engineering problems in existence: keeping a superheated plasma stable long enough to generate usable energy.<\/p>\n [Image: Plasma containment visualization inside a tokamak reactor]<\/p>\n That plasma control problem is exactly where DeepMind has already shown its hand. Back in 2022, DeepMind researchers published work in Nature<\/a> demonstrating that reinforcement learning, the same family of algorithms behind AlphaGo, could control the magnetic fields inside a tokamak reactor at Switzerland’s Swiss Plasma Center. That project proved AI could shape plasma in real time without needing a human at the controls making split-second physics calculations. Fusionality’s pitch is essentially: let’s take that proof of concept and turn it into a product every fusion startup can plug into, instead of every lab reinventing the wheel.<\/p>\n It’s a classic picks-and-shovels play, and it comes at a moment when fusion has never had more money or attention behind it. Private fusion companies have raised billions in the past few years, with players like Commonwealth Fusion Systems<\/a> and Helion Energy<\/a> racing toward pilot plants by the early 2030s, backed by everyone from Microsoft<\/a> to OpenAI<\/a> chief Sam Altman personally. Every one of those companies still has to solve the same nagging problem Fusionality is targeting: how do you control a plasma that behaves unpredictably at 100 million degrees, using sensors and actuators that can’t afford to lag by even a few milliseconds.<\/p>\n “We kept seeing fusion teams solve the same control and simulation problems from scratch,” one person familiar with Fusionality’s pitch told TechCrunch<\/a>, describing the redundant engineering effort across the industry. That redundancy is expensive in an industry where a single reactor prototype can cost hundreds of millions of dollars and years of build time.<\/p>\n The bet Fusionality is making mirrors what happened in autonomous vehicles a decade ago, when simulation environments from companies like Waymo<\/a> became just as valuable as the cars themselves, because testing control software in a virtual world is far cheaper and faster than testing it on a real, extremely expensive piece of hardware. Apply that logic to fusion, where a botched experiment can mean months of reactor downtime, and the value of a good simulation environment goes up dramatically.<\/p>\n