The fusion industry's impossible triangle

For decades, fusion energy has been the ultimate promise of clean energy. It is the process that powers the sun, capable of producing enormous amounts of energy without the carbon emissions associated with fossil fuels. Scientists have spent generations trying to recreate it on Earth, convinced that if they can make it work at scale, it could fundamentally reshape the world's energy future. The problem is that fusion is incredibly difficult, not just from a scientific perspective but from an economic one. Building and testing experimental reactors costs vast amounts of money, and progress often comes through a frustrating cycle of trial and error. Researchers develop a theory, build hardware to test it, gather data, tweak the design, and repeat the process. Sometimes that cycle takes years.

Now, a Chinese startup called VeloAlpha believes artificial intelligence could help break that pattern. Founded earlier this year by fusion scientist Xie Huasheng, the Beijing-based company is developing FusionAlpha, a simulation platform that lets researchers test fusion reactor designs digitally before committing to expensive physical experiments. It may not sound as exciting as a giant reactor generating limitless clean power. But if VeloAlpha's technology delivers on its promise, it could end up solving one of fusion's most expensive and persistent challenges.

According to Xie, fusion researchers have long been stuck with an uncomfortable trade-off. The most advanced simulation software available today can model plasma behavior with remarkable accuracy. Plasma — the superheated, electrically charged gas that fuels fusion reactions — is notoriously difficult to control, and understanding its behavior is critical to designing a viable reactor. The catch is that these simulations require substantial computing resources and can take a long time to complete.

At the opposite end of the spectrum are newer AI-driven systems that can process calculations much faster. While attractive from a speed perspective, researchers often remain cautious because these tools can struggle with reliability and extrapolation beyond the data they were trained on. Then there are simplified physics models, which are computationally efficient but often too crude to accurately guide the design of next-generation reactors. Xie describes this as fusion software's 'impossible triangle': speed, accuracy, and predictive capability. Historically, researchers have had to sacrifice one to gain another. VeloAlpha's entire business is built around the idea that this trade-off no longer has to exist.

How FusionAlpha works

The company claims that advances in artificial intelligence, combined with new mathematical techniques, can dramatically accelerate simulations without sacrificing the underlying physics. According to Xie, some parts of FusionAlpha can run anywhere from 100 to 10,000 times faster than today's state-of-the-art fusion codes while maintaining benchmark errors below 5%. Those claims still need independent validation, but if they hold up, they would represent a significant leap forward for the industry.

FusionAlpha integrates machine learning models trained on vast datasets from existing experiments and high-fidelity simulations. These models learn the complex relationships between plasma parameters — such as temperature, density, and magnetic field configuration — and the resulting stability and reaction rates. Instead of solving the full set of plasma equations from scratch each time, the AI uses pattern recognition to approximate the outcome, dramatically cutting simulation time. The platform also allows researchers to explore a much wider parameter space than previously possible, potentially uncovering novel reactor designs that would have been overlooked using traditional methods.

Building a star is expensive

To understand why software matters so much, it helps to understand what fusion researchers are trying to accomplish. Fusion occurs when the nuclei of light atoms collide and merge, releasing huge amounts of energy. That is exactly what happens inside stars. Replicating those conditions on Earth requires heating fuel to temperatures hotter than the sun's core, creating plasma that must then be confined and stabilized long enough for fusion reactions to occur. Most researchers attempt this using machines called tokamaks — massive doughnut-shaped devices that use powerful magnetic fields to contain plasma. Others are experimenting with alternative approaches, including stellarators, linear devices, and laser-driven fusion systems.

Every design comes with its own engineering challenges. Researchers must figure out how to sustain reactions, withstand extreme heat, manage radiation, secure fuel supplies, and ultimately generate electricity cheaply enough to compete with existing energy sources. None of those problems are inexpensive to solve. A single experimental facility can cost hundreds of millions or even billions of dollars. Even smaller design changes often require extensive testing and validation. That is why simulation software has become increasingly important. The more accurately researchers can predict outcomes before building hardware, the less money they waste on dead-end pursuits.

Fusion's EDA moment

Xie compares FusionAlpha to electronic design automation (EDA) software, a technology that transformed the semiconductor industry. Modern chip companies do not build a physical processor every time they want to test a new idea. Instead, they use sophisticated software tools to model, simulate, and optimize designs before sending them to fabrication plants. Without EDA software, the pace of semiconductor innovation would be dramatically slower.

VeloAlpha believes fusion is approaching a similar turning point. Rather than relying primarily on physical experimentation, future fusion companies could use advanced simulation platforms to virtually test thousands of design variations, identify promising approaches, and dramatically reduce development costs. So, the next generation of fusion reactors may be built twice: first in software, then in steel.

This analogy is particularly apt because the fusion industry, like semiconductors in the 1970s, is still in its infancy. Many design choices — such as whether to use a tokamak or a stellarator, what materials to line the reactor walls with, or how to extract heat — remain open questions that require extensive exploration. AI-powered simulation could compress decades of iterative experimentation into months.

Why the timing matters

The startup's emergence comes at a particularly interesting moment for China's fusion industry. For years, fusion research was largely driven by governments and national laboratories. That is beginning to change. China has identified nuclear fusion as a strategic future industry, placing it alongside fields such as quantum computing, embodied AI, biomanufacturing, brain-computer interfaces, and 6G communications. Investors have taken notice, pouring money into a growing ecosystem of fusion startups, component suppliers, and supporting technologies.

Companies focused on reactor development are attracting increasingly large funding rounds, while businesses supplying magnets, materials, power systems, and software are also emerging. VeloAlpha sits at the intersection of two of the biggest technology trends of the decade: artificial intelligence and clean energy. The company recently secured seed funding from investors who appear convinced that fusion's future will not be determined solely by advances in hardware.

That does not mean commercial fusion is right around the corner. The industry still faces enormous technical and economic hurdles, and many experts believe that practical fusion power remains years, or even decades, away. But as the sector becomes more competitive, the companies that can iterate fastest may gain a significant advantage. And that is where software could become as important as the reactors themselves.

Beyond VeloAlpha, other startups are also applying AI to fusion. For example, researchers at MIT and DeepMind have used reinforcement learning to control plasma in tokamaks. However, VeloAlpha's focus on design-stage simulation rather than real-time control sets it apart. By enabling rapid virtual prototyping, the platform could lower the barrier to entry for new fusion companies and accelerate the overall pace of innovation.

Additionally, the Chinese government's strong support for AI and clean energy provides a favorable environment. China already leads in patent filings related to fusion, and the country has built several major experimental reactors, including the EAST tokamak in Hefei. VeloAlpha's software could help Chinese researchers optimize these machines more effectively, potentially giving the country a lead in the race to commercial fusion.

However, challenges remain. The FusionAlpha simulator must prove its accuracy on a wide range of reactor designs, not just the ones it was trained on. Fusion plasmas exhibit complex, nonlinear behavior that can be difficult for AI models to capture, especially in regimes not covered by training data. VeloAlpha plans to continuously update its models using new experimental data from partner facilities. The company is also exploring hybrid approaches that combine fast AI approximations with occasional high-fidelity runs to validate results.

For years, the fusion industry's biggest challenge has been figuring out what to build. If AI can help answer that question faster and more accurately, the path toward commercial fusion may suddenly look a little shorter. The promise of limitless clean energy remains tantalizingly close, and with tools like FusionAlpha, the journey to reach it may become far more efficient.

Source: Digital Trends News