VeloAlpha, a startup founded in Beijing in April 2026 by scientist Xie Huasheng, has developed an AI-powered simulator called FusionAlpha that promises to accelerate nuclear fusion reactor design between 100 and 10,000 times faster than current simulation codes, with a margin of error below 5% in benchmark tests. The company has already closed its first funding round to build a dedicated simulation center.
The problem FusionAlpha aims to solve is as much technical as it is economic. Xie calls it the “impossible triangle” of fusion simulation: existing tools are either highly accurate but computationally prohibitive, fast but unreliable, or simple but far too approximate to guide the design of next-generation machines. Each development cycle requires researchers to formulate a theory, build hardware to test it, collect data, revise the design, and repeat the process — sometimes for years. FusionAlpha aims to replace much of this physical experimentation with large-scale virtual simulation: thousands of design variants tested computationally before a single dollar is spent on steel or superconductors.
VeloAlpha’s approach draws on recent advances in applied mathematics and machine learning techniques. Xie has spent much of his career developing mathematical tools and software for fusion plasma modeling — the ionized gas at extreme temperatures that drives thermonuclear reactions and is notoriously difficult to control. According to the founder, the performance of more than a dozen physics models used for design and analysis has improved significantly through more refined mathematical frameworks and AI integration. The result is a platform that does not replace physicists, but compresses the time it takes to turn physics into verifiable engineering.
The context in which VeloAlpha has emerged is no coincidence. China has identified nuclear fusion as a strategic industry of the future, alongside fields such as quantum computing and 6G communications. Private investors are funding a growing number of startups in the sector — companies working on reactors, magnets, materials, power systems, and now software as well. Energy Singularity, another rising name in China’s fusion ecosystem, has already unveiled the HH70, a compact superconducting tokamak developed entirely through domestic supply chains to keep costs competitive with Western rivals. VeloAlpha adds a different piece to the puzzle: it does not build physical machines, but the tools to design them faster and more effectively.
The international race is already underway. Google DeepMind has worked on plasma control through reinforcement learning. The next generation of fusion reactors may effectively be born twice — first in software, then in metal — and whoever controls the simulation tools holds a structural advantage across the entire development chain. If FusionAlpha delivers on the promise of its early benchmarks, VeloAlpha could become an enabling infrastructure for dozens of fusion teams worldwide, regardless of where they choose to build their reactors.
