The nuclear fusion experimental device, Honghuang 70, built by Energy Singularity

In the summer of 2022, hundreds of millions of venture capital were directed towards two Chinese fusion energy startups.

Two years later, a fusion experiment device was constructed in Lingang, Shanghai, completing its initial technical validation. It does not appear grand, with its main structure standing only 3 meters tall, and is surrounded by numerous pipes used for delivering gases, coolants, and electricity for the experiments.

Before activating the device, it takes several weeks to create a vacuum environment by extracting air from the device. Subsequently, facilities around it inject liquid nitrogen and liquid helium into the area outside the vacuum chamber over several weeks, cooling it to around -240 degrees Celsius. Then, current is injected into superconductors within, forming a powerful helical magnetic field. Maintaining such an environment costs 300,000 yuan in electricity every month.

The primary function of the device is to test whether these facilities can successfully ignite plasma—the fourth state of matter beyond gas, liquid, and solid—an essential condition for achieving nuclear fusion.

Upon clicking the “Start Experiment” button, engineers witness an instantaneous bombardment of electrons along the helical magnetic field, transforming the pre-injected helium gas in the vacuum chamber into rapidly rotating plasma. Simultaneously, surrounding devices emit electromagnetic waves with the same rotational frequency as the plasma, heating it to 5 million degrees Celsius.

This entire testing process lasts only a few milliseconds—less than the blink of eye. However, to achieve this fleeting test, Energy Singularity, a Chinese fusion company funded by MiHoYo, NIO Capital, and HongShan, invested nearly 200 million yuan over two years to construct this experimental device named Honghuang 70.

MiHoYo’s game Genshin Impact has 44 million overseas monthly active users and $800 million in global cumulative user spending

Yang Zhao, the chief executive officer of Energy Singularity, stated that this test validated the company’s technical approach. They plan to invest tens of billions in developing Honghuang 170, the next-generation device, aiming to achieve an energy gain greater than ten times by 2027—generating 100 units of electricity for every 10 units input. Currently, no controllable nuclear fusion device has reached this goal.

Energy Singularity’s new developments are part of the progress made by various controllable nuclear fusion startups. Another Chinese fusion startup, Startorus Fusion, also completed and successfully operated its first-generation experimental device in July last year and is now developing the next-generation device.

In the United States, fusion startups are even more ambitious. Helion plans to build a fusion device capable of generating electricity by 2028 and has already signed a power supply agreement with Microsoft. TAE Technologies announced plans to commercialize fusion by 2030. And at least four companies have set goals to generate electricity using nuclear fusion by 2030.

The Principles of Controllable Nuclear Fusion are Clear, but the Engineering Challenges Remain

Before undertaking the Manhattan Project which produced the first nuclear fission bomb, scientists had already grasped the principles of nuclear fusion: Combining two light atomic nuclei (e.g., deuterium and tritium) together can release enormous energy.

The first artificial nuclear fusion was achieved in 1952, when the first hydrogen bomb exploded over Bikini Atoll in the Pacific, with a force equivalent to 500 times the atomic bomb dropped on Hiroshima.

Light atomic nuclei all carry a positive charge, naturally repelling each other. To make two atomic nuclei collide and combine, suitable conditions are required. First, they must be converted into plasma (a state of matter distinct from liquid, solid, and gas) and then heated to at least 100 million degrees Celsius. This is necessary to overcome the repulsive force and allow the atomic nuclei to combine and produce nuclear fusion—the temperature in the region of the sun’s surface where energy is generated through nuclear fusion is only about 15 million degrees Celsius.

All hydrogen bombs achieve fusion by directly detonating an atomic bomb. However, using fusion for electricity generation requires a gentler method to heat the plasma.

Currently, some devices can heat plasma to over 150 million degrees Celsius, but they can only maintain this for a short duration, typically measured in seconds. This is because plasma is extremely unstable, akin to a constantly turbulent high-temperature ionized gas. Reactors need to confine the plasma stably in a limited space (Plasma Confinement) to enable frequent nuclear collisions and sustained energy release.

A long-time fusion researcher noted that studying plasma is a bottomless pit, involving many complex physical phenomena. Currently, no model can accurately predict how plasma move, so external forces are used to confine it.

The Tokamak, a device invented by Lev Artsimovich and his colleagues in the late 1950s, and the Inertial Confinement Fusion (ICF) device developed by the Lawrence Livermore National Laboratory in California represent the magnetic confinement and inertial confinement approaches, respectively, in nuclear fusion.

The internal structure of the Tokamak device

In the Tokamak device, atoms are injected into a doughnut-shaped vacuum channel and microwaved into plasma. Each direction of the channel is wrapped by different shapes of magnetic coils which form a magnetic field when electrified to compress the 100-million-degree plasma to a certain density, turning it into a high-speed helix.

In the ICF device, laser or particle beams impact fuel sealed in a specific space, creating a high-temperature and high-pressure environment to achieve fusion, simulating the process of an atomic bomb triggering a hydrogen bomb explosion.

The longest-lasting artificial nuclear fusion so far was achieved by the JET Tokamak device, jointly funded by multiple European countries. At the end of last year, it achieved 5.2 seconds of nuclear fusion, generating 69.26 megajoules of heat—equivalent to 19.24 kWh. However, it consumed more energy than it produced, rendering it commercially unviable. After this test, JET was dismantled, as the 1983-built device was no longer ideal for continued research.

The ICF device built by the Lawrence Livermore National Laboratory in California produced more energy with less input in two fusion experiments in 2022 and 2023, but the energy consumed by the laser starter was not included in the calculations. Otherwise, the energy produced by this device was only 3.15 megajoules.

If high temperatures can be maintained and the plasma retains high density as envisioned, controlled nuclear fusion can continuously occur, leading to the creation of a super nuclear power plant equivalent to existing nuclear power stations but with almost no radioactive pollution risk.

Government agencies are actively promoting fusion research along this direction, but efficiency is low. The ITER (International Thermonuclear Experimental Reactor) is a typical example, standing nearly 30 meters high, equivalent to a 10-floor building.

The device, initiated by the Soviet Union and the U.S. government in 1985, represents the hope of cooperation between the two major parties in the Cold War, and the most ambitious investment in nuclear fusion research so far. ITER’s goal is to heat plasma to a maximum of 300 million degrees Celsius and sustain fusion for 500 seconds, using 50,000 kWh of energy per hour to release 500,000 kWh. Once achieved, controlled nuclear fusion will be within reach.

However, after the Cold War, Russia’s financial capabilities were limited, and the U.S. government also cut fusion research funding. It wasn’t until 2006, when China, the EU, Japan, India, and South Korea joined, that ITER’s construction plan was finalized.

While multi-country cooperation can share costs, it significantly reduces efficiency. ITER is expected to be completed next year, then undergo a decade of adjustments before formally operating in 2035.

The ITER project’s decades of twists and turns led to the birth of the joke, “controlled nuclear fusion is always 30 years away,” reflecting the fatigue of many hopefuls.

Like Space X, Repeating What the Government Has Done Before

In 2021, a turning point emerged with significant progress in commercial fusion research, attracting private capital.

In June of that year, Helion, an eight-year-old fusion startup, announced that it had heated plasma to 100 million degrees Celsius, a feat previously achieved only by government projects. Five months later, Helion received a record $500 million investment from Silicon Valley notables and venture capital firms, including OpenAI CEO Sam Altman and PayPal co-founder Peter Thiel, equivalent to how much the US government funded in fusion research. They pledged an additional $1.7 billion if Helion continued to make breakthroughs.

In November, Commonwealth Fusion Systems (CFS), a fusion startup spun out of MIT three years earlier, announced over $1.8 billion in funding from Bill Gates, George Soros, Google, DFJ, and Emerson Collective, among others.

Investors believed CFS had unprecedented breakthroughs. They developed the world’s strongest high-temperature superconducting magnet in collaboration with MIT, producing a magnetic field exceeding 20 Tesla, 1.5 to 2 times stronger than ITER’s magnetic field.

A stronger magnetic field means better plasma confinement, enhancing fusion efficiency. CFS’s research revealed another path for fusion development: building smaller Tokamak devices that can generate energy more efficiently without the massive materials and time required by ITER. This is the direction Energy Singularity chose. New technological advancements and substantial funding rounds have fueled the fusion startup enthusiasm. According to the Fusion Industry Association, as of June 2023, over 40 fusion startups worldwide have raised nearly $6 billion from investors.

Leveraging decades of government research and new technological and material advancements, startups now only need a few hundred million dollars to validate their technical direction and reduce construction cycles to 3-5years, when building small fusion devices.

“Achieving fusion with Tokamak devices is already very feasible. We don’t need extensive scientific validation or research, just focus on solving engineering challenges,” said Energy Singularity COO Ye Yuming. Cost has become crucial in this process. He mentioned that during the design of the first-generation device, they considered outsourcing some designs to a research institute involved in the ITER project but turned to in-house design due to the high cost. “Whether it’s copper conductors or high-temperature superconductors, the physical principles are the same,” Ye said. After finalizing the design, they collaborated with several suppliers with years of nuclear power industry experience to produce the first-generation device:

• Shanghai Electric Nuclear Power Group Co.,Ltd. produced key equipment for the Tokamak system of Honghuang 70, including the vacuum chamber, Dewar, and cold shield.

• Shanghai Superconductor Technology Co., Ltd. provided all high-performance superconducting magnet materials for Honghuang 70.

• China Nuclear Industry Fifth Construction Co., Ltd. helped assemble Honghuang 70.

Energy Singularity is assembling the Honghuang 70

In Energy Singularity’s plan, building Honghuang 70 was just a step to validate the technical direction, “using relatively short time and low cost to verify the technical feasibility, avoiding risks, and then developing a larger investment and higher performance device.”

Most commercial companies follow this approach. Different fusion startups chose over 20 different schemes to build fusion devices, relying on decades of research results and iterating step by step.

For example, Helion has iterated its device to the seventh generation in over ten years, with the fusion supplying power to Microsoft in 2028 being their eighth-generation device, currently in the design stage.

“(The cooperation with Microsoft) is a binding agreement. If we fail, we will face economic penalties,” said Helion CEO David Kirtley. In the same year, Helion also reached an agreement with Nucor to build a fusion reactor to help it smelt steel by 2030.

Using low-cost solutions to build devices previously constructed with significant government investment, and then upgrading and iterating repeatedly, is a development path verified by SpaceX in its rockets and spacecrafts. The difference is that when SpaceX was founded in 2002, government-engineered rockets had already sent humans into space, landed on the moon, and established the International Space Station. Most of SpaceX’s progress involved solving problems encountered by NASA before at lower costs to stimulate commercial applications. However, nuclear fusion remains an unsolved challenge that major governments have strived to overcome for over half a century. Even if commercial companies catch up with government research, challenges still lie ahead.

Driven by AI and Driving AI Forward

Part of the confidence in commercial fusion companies comes from AI technology advancements.

Current experiments can heat plasma to 100 million degrees Celsius. To achieve controllable nuclear fusion, the key is to sustain fusion, ensuring the energy produced far exceeds the energy consumed.

In each fusion experiment, scientists must prepare parameters for controlling magnets in advance, based on principles and intuition, adjusting voltage thousands of times per second to change the magnetic field and prevent the high-temperature plasma from touching the walls inside the vacuum chamber. Otherwise, the plasma temperature will drop, or the device will break down, ultimately halting fusion.

AI can learn from historical data (including simulated data) how to better control plasma. Its learning process is similar to how DeepMind’s AlphaGo learned Go, setting a goal—precisely controlling plasma—and receiving rewards for achieving it and penalties for failing. Through numerous experiments, AI might find ways to control plasma for extended periods, sustaining fusion.

In February 2022, a research paper by Google DeepMind, after peer review, was published in Nature. It demonstrated that AI trained through reinforcement learning could control 19 magnetic coils simultaneously in a Tokamak device at the Swiss Plasma Center, releasing voltage tens of thousands of times per second, surpassing even experienced scientists.

The AI algorithm developed by Google DeepMind controlling the plasma

Subsequently, research on using AI to monitor and control plasma became more prevalent. In March this year, Princeton University published a paper introducing an AI algorithm that could predict plasma disruptions 300 milliseconds in advance, aiding scientists in real-time parameter adjustments to extend fusion reactions.

DeepMind continues to optimize its algorithms. In a paper published last July, they introduced a new method that reduced the time to train AI algorithms for controlling plasma to one-third and improved control precision by 65%.

A founder of a Chinese fusion company mentioned that after DeepMind’s first paper was published, they immediately began replicating its AI model for their projects. It is becoming a standard for fusion startups.

As large models become popular, AI and fusion, two technologies born in the 1950s, have formed deeper connections.

Compared to traditional algorithms, running large models consumes more energy. In a report released by the International Energy Agency (IEA) earlier this year, it was noted that a single Google search consumes about 0.3 Wh of electricity, while one use of ChatGPT consumes 2.9 Wh. They estimate that if algorithm power consumption is not reduced, the electricity consumption of data centers running large models could double by 2026, exceeding 1,000 TWh annually—roughly equivalent to Japan’s entire electricity consumption.

As the world transitions to electrification while aiming to reduce carbon dioxide emissions, the promotion of large models increases electricity consumption, putting more pressure on global power systems. Amazon’s data center in Ireland, due to excessive GPU power consumption, has already limited user access.

Some companies are focusing on nuclear energy. For instance, Microsoft has started hiring nuclear energy experts to use small nuclear reactors to power data centers. In March this year, AWS purchased a data center built next to a nuclear power plant.

“Future AI needs an energy breakthrough. The electricity consumed by AI far exceeds people’s expectations,” said Altman earlier this year. He believes that nuclear fusion is the fundamental solution, “It motivates us to increase our investment in nuclear fusion.” Altman is the chairman of the board of the fusion company Helion.

“When society needs fusion technology, fusion will be realized,” said Lev Artsimovich, a Soviet nuclear physicist who developed the Tokamak device in the 1950s. It is evident that now, more than ever, society needs the miraculous technology of controllable nuclear fusion.