Prometheus' Fire: Controlled Nuclear Fusion — The Leap from Physical Feasibility to Engineering Realization
Abstract: Controlled nuclear fusion, the ultimate solution to humanity's energy and climate crises, is undergoing a historic transition from "is it physically feasible" to "can it be engineered," after more than seven decades of exploration. This review systematically surveys landmark breakthroughs in magnetic confinement fusion (MCF) and inertial confinement fusion (ICF) between 2024 and 2026. In MCF, China's EAST achieved 1,066-second H-mode operation; HL-3 reached a fusion triple product on the order of 10²⁰; Germany's Wendelstein 7-X stellarator sustained a triple-product peak for 43 seconds; and CFS's SPARC compact high-field tokamak targets Q > 1 by 2027. In ICF, the U.S. National Ignition Facility (NIF) has continued breaking records since its first ignition in 2022, reaching a fusion yield of 8.6 MJ with a capsule gain Q = 4.13. In artificial intelligence, China's first AI for Fusion startup, Xinzhu Shidai, founded in September 2025, completed a 60-million-RMB angel round in March 2026, building a fusion intelligence operating system driven by a dual "physics + data" engine — marking a pivotal step in the industrialization of AI for Fusion. Simultaneously, the explosive growth of global private fusion financing has spawned speculative excesses: Inertia Enterprises falsely promoted $450 million in financing by misappropriating NIF's public ignition results; TAE Technologies faced lawsuits from nine suppliers on the eve of a merger. These cases caution that enthusiasm must be tempered with prudence. Based on a systematic evaluation of six critical paths — physics, engineering, materials, fuel cycle, capital ecology, and regulation — this review reaches a measured conclusion: controlled nuclear fusion has crossed the threshold of physical feasibility; its commercialization now depends on overcoming engineering bottlenecks and purifying the industry ecosystem. An appendix explores, as a thought experiment, the possibilities and challenges of controlled nuclear fusion in the space environment. As Prometheus' fire draws near to being kindled by Chinese hands, ensuring it is held in a safe grip — one combining physical rigor, engineering robustness, and commercial integrity — is the shared responsibility of our era.
Keywords: controlled nuclear fusion; magnetic confinement; inertial confinement; tokamak; high-temperature superconductors; tritium breeding; artificial intelligence; capital ethics; space fusion
1. Introduction: From the "Eternal Fifty Years" to Crossing the Threshold
Ever since Soviet scientists Sakharov and Tamm first proposed the tokamak concept in the 1950s, the pronouncement that "nuclear fusion is always fifty years away" has hung like an inescapable shadow over every generation of fusion researchers. Yet this seemingly fatalistic rhetoric conceals the fact of six orders of magnitude of exponential progress in the triple product over six decades, and obscures the deeper physical complexities revealed by each breakthrough that brings us closer to the end.
On December 5, 2022, the U.S. National Ignition Facility (NIF) achieved the first controlled fusion ignition in human history, with a capsule gain Q = 1.5. By April 7, 2025, NIF had produced 8.6 MJ of fusion energy in a DT capsule using 2.08 MJ of laser energy, reaching a capsule gain Q = 4.13, and accumulating ten ignition events. In the magnetic confinement domain, China's EAST achieved steady-state H-mode plasma operation at 100 million degrees Celsius for 1,066 seconds on January 20, 2025, once again breaking the world record for H-mode operation on a tokamak. Germany's Wendelstein 7-X stellarator sustained a triple-product peak for 43 seconds in May 2025. Together, these landmark events point to a core conclusion: controlled nuclear fusion has shifted from "is it physically feasible" to "can it be engineered."
Equally significant is the fact that artificial intelligence is profoundly reshaping the paradigm of fusion research. Between 2025 and 2026, AI achieved a dense cluster of breakthroughs on China's two flagship tokamaks. At the same time, AI for Fusion is rapidly crystallizing as an independent industrial track. In September 2025, China's first technology company dedicated exclusively to AI for Fusion — Xinzhu Shidai — was founded, and in March 2026 it completed a 60-million-RMB angel round co-led by CAS Star and Dingfeng Sci-Tech Innovation, with Tsinghua Alumni Fund as co-investor. The company builds its technical foundation on reinforcement learning, generative models, self-evolving agents, and operator learning, constructing a dual "physics + data" engine. Its CTO, Wang Yue, previously worked at Microsoft Research Asia for nearly a decade, specializing in reinforcement learning and AI for Physics. Xinzhu Shidai positions itself as the "intelligent brain" of fusion devices: "Fusion companies are like Intel — focused on the chip hardware — while Xinzhu Shidai is like Microsoft, providing them with the operating system." This industrial practice, together with the trusted fusion AI research carried out by Phaenarete AI Technology Co., Ltd. under its mission of "empowering fusion safety control through AI," jointly constitutes China's industry-academia-research collaborative layout for fusion intelligence.
Yet beneath the capital flood, sediment swirls. The fusion domain has become fertile ground for speculators seeking arbitrage. A laser-fusion startup named Inertia Enterprises, riding NIF's ignition achievements, engaged in false promotion, claiming its team possessed "decades of experience in developing and operating the only fusion system to have achieved ignition," and in February 2026 announced a new $450 million investment round — in reality packaging the U.S. government national laboratory's public results as its own commercial endorsement. This episode reveals a deep tension in the fusion field: on one hand, genuine physical breakthroughs have ignited capital enthusiasm; on the other, the profit-driven nature of capital has provided fertile soil for false claims and over-promising. Yan Jianwen, CPPCC member and chairman of Fusion New Energy, observed that private fusion enterprises' timelines are a full twelve years ahead of the national team's — Startorus Fusion claims a commercial demonstration reactor by 2033, while CNNC chief scientist Duan Xuru gives an official timeline of 2045. The boundary between acceleration and bubble is becoming the subtlest tension in the fusion industry.
This review aims to systematically integrate key global advances in fusion between 2024 and 2026 — from physical fundamentals to engineering challenges, from traditional approaches to the new AI paradigm, from capital ecology to ethical reflection — providing a panoramic cognitive framework for domestic fusion researchers and decision-makers, and on that basis, soberly assessing when exactly "fifty years" will reach zero.
2. Magnetic Confinement Fusion: The Tokamak Advances as the Main Force
2.1 Chinese Tokamaks: From "Catching Up" to "Running Alongside" to "Leading"
The center of gravity of magnetic confinement fusion research is shifting eastward. On January 20, 2025, EAST achieved long-pulse H-mode plasma operation at 100 million degrees Celsius for 1,066 seconds, once again breaking the world record for steady-state H-mode operation of a tokamak. EAST had previously crossed major milestones of 60 seconds, 100 seconds, and 400 seconds.
HL-3 (Huanliu-3) achieved systematic breakthroughs in plasma temperature and triple product. Between March and June 2025, the Southwestern Institute of Physics successively announced that the device had, for the first time in China, achieved "double hundred-million-degree" high-parameter operation with ion temperatures of 117 million degrees and electron temperatures of 160 million degrees, and reached the milestone of a fusion triple product on the order of 10²⁰ under a plasma current of 1 MA. According to committee member Duan Xuru, HL-3 is expected to conduct burning plasma experiments in 2027. HL-3 has been open to the global community since 2023, inviting scientists worldwide to tackle challenges together.
The construction of the China Fusion Engineering Demonstration Reactor (CFEDR) has also been comprehensively accelerated. According to the integrated physics design published by Ding et al. in 2025, CFEDR is based on a major radius R = 7.2 m, a toroidal field B_T = 6.5 T, and a plasma current I_p = 15 MA, with a design fusion power of 1.5 GW and an energy gain Q = 14.9. The supporting CRAFT facility has also entered a critical phase. According to the plan, CFEDR is to begin construction in 2030, be completed in 2035, and demonstrate power generation around 2040.
In fundamental fusion physics research, the EAST team published a paper in Science Advances in January 2026, reporting the important result of plasma density entering the "density-free regime" with operational density exceeding the conventional Greenwald limit by 65%.
2.2 International MCF Progress: JET's Curtain Call and W7-X's Rise
JET completed its third deuterium-tritium experimental campaign (DTE3) on October 3, 2023, releasing a record 69 MJ of fusion energy with only 0.2 mg of D-T fuel in a 5.2-second sustained discharge, with a peak fusion power of 16 MW and an energy gain Q = 0.67. JET concluded scientific operations at the end of 2023; its more than twenty years of DT experimental legacy has been fully inherited by ITER and the DEMO programs of various nations.
Germany's Wendelstein 7-X stellarator announced its latest results in May 2025: sustaining the triple-product peak for 43 seconds in a long-pulse discharge, with plasma temperatures reaching 30 million degrees, a volume-averaged plasma beta reaching 3% for the first time, and an energy turnover of 1.8 GJ in a 360-second discharge. The result was reported by ITER Newsline as a "new triple-product peak." The stellarator's triple-product sustainment time under specific long-pulse conditions has surpassed that of existing tokamaks, directly attributable to its advantages of zero plasma current, intrinsic steady-state operation, and zero disruption risk. In May 2026, the U.S. Department of Energy and Germany's Max Planck Institute for Plasma Physics signed a ten-year cooperation agreement to jointly advance W7-X research.
2.3 Compact High-Field Tokamaks: SPARC and the HTS Magnet Revolution
The commercialization of high-temperature superconducting (HTS) REBCO tapes is rewriting the economic logic of magnetic confinement fusion. HTS magnets make 20 T-class fusion fields feasible, allowing equivalent fusion gain to be pursued at roughly 1/20 the volume of ITER. As of early 2026, CFS's SPARC device has made substantial construction progress: the first of eighteen D-shaped HTS magnets was installed in January 2026, with all magnets expected to be installed by summer 2026. SPARC is expected to be near completion by the end of 2026, produce first plasma in 2027, with the primary scientific goal of validating Q > 1.
Shanghai Superconductor Technology Co., Ltd. applied for an IPO on the STAR Market in June 2025, planning to raise 1.2 billion RMB for capacity expansion; after subsequent expansion projects, total capacity will reach 20,000 km/year. In the global REBCO capacity landscape, China already occupies a first-tier position.
2.4 Private Fusion: Capital Influx, the "Horse Race" Pattern, and Industrial Bubbles
According to the Global Fusion Industry in 2025 report released by the Fusion Industry Association (FIA) in July 2025, cumulative global commercial fusion financing reached 2.643 billion from 2024, the highest annual increase in three years. Major players have formed a "horse race" pattern with multiple technical routes in parallel: CFS has cumulative financing of approximately $3 billion, accounting for one-third of total global fusion industry financing; Helion Energy's Polaris prototype was commissioned in 2025; ZAP Energy's FuZE-3 device achieved total plasma pressure of 1.6 GPa (electron pressure 830 MPa), setting the highest pressure record for shear-flow-stabilized Z-pinch technology; Type One Energy and others also secured substantial funding in early 2026.
However, as noted in the introduction, there are opportunists in the capital tide. Inertia Enterprises "borrowing light from ignition" — using the public achievements of national laboratories to gild its business plan — reflects a pervasive contradiction in the industry ecosystem: the high complexity of fusion technology makes it difficult for investors to effectively distinguish genuinely scientifically feasible technical routes from carefully packaged commercial narratives, thereby creating the risk of adverse selection under information asymmetry. More broadly, as early as 2022, observers had noticed that fusion companies less than six months old obtaining funding was not uncommon, while commentators such as Zhao Ling bluntly stated in a series of critical essays that "any garbage that latches onto the prestige of top universities can harvest investors under the banner of 'controlled nuclear fusion energy entrepreneurship.'" TAE Technologies, on the eve of signing a merger agreement with Trump Media & Technology Group valuing the deal at over $6 billion, faced lawsuits from nine suppliers for unpaid bills. This reminds the industry: enthusiasm and prudence must coexist; capital should be grounded in scientific feasibility, rather than the reverse — a narrative-driven approach.
3. Inertial Confinement Fusion: NIF's Ignition and Beyond
On December 5, 2022, NIF produced 3.15 MJ of fusion energy in a DT capsule using 2.05 MJ of laser energy, achieving the first scientific energy gain Q = 1.5. By April 7, 2025, NIF set a historic record of 8.6 MJ yield with capsule gain Q = 4.13 using 2.08 MJ of laser energy. By October 2025, NIF had accumulated ten ignition events, validating the physical repeatability of the ICF pathway.
It must be noted that as a single-shot experimental facility, NIF's Q value is defined as the ratio of fusion energy released to laser energy injected, not accounting for laser generation efficiency (approximately 1%) and target fabrication energy consumption — fundamentally different from the calculation of steady-state Q in magnetic confinement fusion. The leap from scientific ignition to inertial fusion energy (IFE) still requires crossing an enormous engineering gulf: repetition rates must reach 1–16 Hz (NIF operates only 300–400 shots per year), electro-optical efficiency must exceed 10%, target costs must fall dramatically, and the tritium cycle must be closed.
Notably, NIF's publicly announced ignition results have objectively provided a "credibility endorsement" for speculative behaviors such as those of the aforementioned Inertia Enterprises. This phenomenon suggests that the dissemination of results from large-scale scientific facilities requires greater prudence, to avoid being taken out of context by commercial speculators to mislead the capital markets.
4. Fusion AI: The Paradigm Leap from Offline Analysis to Trusted Safety
4.1 AI Plasma Control: From Proof of Concept to Multi-Device Deployment
In 2022, DeepMind and the Swiss Plasma Center published a study in Nature demonstrating, for the first time, autonomous control of tokamak plasma magnetic configurations using deep reinforcement learning (DRL). In February 2024, Seo et al. reported in Nature that DRL achieved real-time suppression of tearing mode instabilities on the DIII-D device — the first successful AI control of an advanced physical process in a fusion reactor. The AI system was able to identify precursor signals milliseconds before plasma destabilization and adjust actuators in advance, a response speed far exceeding the cognitive limits of human operators.
On Chinese fusion devices, AI control has entered a rapid deployment phase characterized by "multi-device, multi-task." On HL-3, the Southwestern Institute of Physics, in collaboration with Zhejiang University and Zhijiang Lab, achieved "zero-shot" closed-loop magnetic control based on data-driven kinetic models. The EAST team used visible-light camera imagery and deep learning to perform real-time plasma boundary identification and shape control. In disruption prediction, EAST achieved real-time early warning of locked-mode disruptions using interpretable machine learning, with a success rate of 94% and an average warning time of 137 milliseconds.
4.2 Industrialization of AI for Fusion: Xinzhu Shidai's "Operating System" Practice
Xinzhu Shidai, founded in September 2025, positions itself as the "intelligent brain" of fusion devices — just as Microsoft provides the operating system for Intel chips, Xinzhu Shidai aims to build, for fusion devices, a suite of AI workflows and models that transform scattered, ambiguous, and tacit expert judgments into reproducible, quantifiable, and deployable AI processes. Its product core covers four capabilities: diagnostics, prediction, control, and design. The diagnostics module fuses multimodal external signals (magnetic, optical, high-speed camera) and infers the plasma state via the Grad-Shafranov equation; the prediction module models plasma evolution as a next-frame prediction problem; the control module converts prediction results into real-time decision commands at a control frequency of 10 kHz (one decision every 0.1 milliseconds).
Xinzhu Shidai has established deep collaborations with fusion enterprises such as Startorus Fusion, providing AI capabilities across diagnostics, prediction, control, and design. Its model's differentiation lies in this: the input is not textual tokens, but continuous high-dimensional physical fields; it does not rely on brute-force data scaling, but uses physical laws as core constraints. This technical approach, together with the trusted fusion AI research carried out by Phaenarete AI Technology Co., Ltd., differs in technical emphasis — the former focuses on "operating-system-level" full-stack intelligence, while the latter delves into rigid constraints and formal verification in safety-critical scenarios — yet both embody the technical philosophy of a dual "physical principles + data-driven" foundation, and both respond to the national policy direction of "combining the multi-physics coupling characteristics of controlled fusion devices and conducting research on intelligent control systems for controlled nuclear fusion based on AI technology."
4.3 Neural Operators and Surrogate Models: Six to Eight Orders of Magnitude in Computational Acceleration
Fourier Neural Operators (FNO) can predict MHD models with a normalized mean squared error of 10⁻⁵, six orders of magnitude faster than traditional solvers. Citrin's team's QLKNN surrogate model achieves a speed improvement of eight orders of magnitude over GENE gyrokinetic simulations. The NLT code independently developed by the Institute of Plasma Physics, Chinese Academy of Sciences, based on numerical Lie transforms, revealed the physical mechanism by which ITG turbulence near the magnetic axis in weak-shear configurations is suppressed by geometric curvature.
4.4 Trusted Fusion AI: Rigid Physical Constraints and Formal Safety Verification
Although AI performs excellently in plasma control, its safety and trustworthiness are becoming the critical bottleneck for fusion engineering. In 2026, a Princeton University team used interpretable AI on the DIII-D device to achieve automatic control of divertor detachment, and explicitly stated in Physics of Plasmas that the "black box" nature of AI controllers poses a fundamental challenge to their compliant implementation within nuclear safety regulatory frameworks. The IAEA, in its AI for Fusion Coordinated Research Project, has also emphasized that trust-building for fusion AI must rest on a dual "physical principles + data-driven" foundation.
This direction is precisely the core domain in which Phaenarete is deeply engaged. Based on its corporate mission of "empowering fusion safety control through AI," Phaenarete has conducted research on trusted AI architectures for tokamak plasma safety control — embedding, atop data-driven predictive capabilities, rigid physical constraints derived from long-term experimental regularities (such as the Troyon limit β_N ≤ 3.5 and the Greenwald density limit n_G = I_p/πa²), and exploring formal methods for mathematical guarantees of critical safety logic. Preliminary validation was completed on the HL-3 historical discharge dataset; experiments demonstrated that controllers embedding the rigid constraint layer do not output out-of-bound control commands in out-of-distribution scenarios. Due to confidentiality requirements regarding the company's research progress, detailed technical parameters are not disclosed in this paper. The aforementioned research is still undergoing continuous iteration, but has preliminarily demonstrated the engineering feasibility of "trusted AI" in tokamak control.
The industrialization practice of Xinzhu Shidai, the safety and trust exploration of Phaenarete, and the physical research of traditional fusion research institutions — these three are jointly weaving a complete picture of fusion intelligence from different entry points: the first addresses "can AI control it," the second answers "is AI control reliable and verifiable," while the national team's device operations provide the foundational experimental verification platform. Only with these three working in concert can the complete edifice of fusion intelligence be raised.
5. Fusion Engineering Bottlenecks: From Superconducting Magnets to Tritium Self-Sufficiency
5.1 HTS Magnets: Performance Ready, Costs Awaiting Reduction
A GW-class fusion power plant requires over 10,000 km of REBCO tape annually; current global annual capacity remains in the thousands-of-kilometers range, and the unit cost must drop from approximately 20/kA·m. Shanghai Superconductor's continuing capacity expansion plan — from 4,000 km/year by end of 2025 to a subsequent 20,000 km/year — provides the industrial foundation for closing this gap.
5.2 Material Irradiation: The Verification Gap for 14 MeV Neutron Environments
Reduced-activation ferritic/martensitic (RAFM) steels must maintain performance under 50–200 dpa of irradiation damage; the helium embrittlement and swelling effects induced by 14.1 MeV fusion neutrons cannot be fully simulated using fission neutron sources. IFMIF-DONES, located in Spain, began main building construction in 2025; the multilateral international agreement was signed in November 2025 by the European Fusion Energy Agency, Spain (contributing 55%), Japan (5.1%), Croatia (5%), and Italy (approximately 8%). The main building is expected to be completed by the end of 2029 with operations beginning in 2034; comprehensive material qualification cannot be completed until at least the mid-2040s.
5.3 Tritium Self-Sufficiency: The Most Subtle Yet Deadliest Bottleneck
A 1 GW fusion plant consumes approximately 168 kg of tritium annually; global commercial tritium production capacity is virtually zero. The CFEDR blanket design achieves a neutronic TBR of 1.21, but to date, no blanket has been tested in a real D-T fusion neutron environment. ITER's TBM program is the only verification opportunity, deferred due to the postponement of ITER D-T operations to 2039. The essence of the tritium self-sufficiency dilemma is a chicken-and-egg problem — without tritium there is no fusion, and the only large-scale source of tritium is precisely the fusion power plant itself.
5.4 Regulatory Framework: Technology May Outpace Licensing
In February 2026, the U.S. NRC published a proposed rule for a fusion device regulatory framework in the Federal Register, regulating fusion under the byproduct materials framework. The UK and Japan have each established their own regulatory pathways. However, no country possesses a complete licensing framework for D-T fusion power plants. Should future fusion devices be brought under nuclear safety regulation, the interpretability and formal safety assurance of AI control systems may shift from a "bonus feature" to an "entry requirement."
6. Fusion Commercialization Timeline and Measured Assessment
6.1 Triple Product Evolution: Sixty Years of Exponential Progress
From the 10¹⁵ level of the 1960s to the 10²⁰ of HL-3 in 2025, a leap of five to six orders of magnitude represents, by any objective standard, a normal rate of progress for a mature scientific field.
6.2 A Measured Assessment of the Commercial Timeline
Based on a systematic evaluation of six critical paths — physics, engineering, materials, tritium cycle, regulation, and AI control maturity — this review proposes a cautiously optimistic timeline: 2026–2027, SPARC first plasma and HL-3 burning experiments; 2030, CFEDR construction begins; 2034–2035, ITER first plasma and IFMIF-DONES operation; 2039, ITER D-T full-power operation and TBM verification; early 2040s, the first fusion power plant prototypes may be connected to the grid. A delay at any critical node could shift the timeline to the right by a decade or more.
It must be specifically noted that the aggressive timelines — such as "commercial demonstration reactor by 2033" — claimed by capital-driven private fusion enterprises currently lack independent physical and engineering feasibility justification. The FIA 2025 report has already pointed out that some startups' commercialization timelines are based on technical assumptions not yet verified by experiment. This review calls for the fusion industry to establish stricter technical milestone disclosure mechanisms to curb the erosion of industry credibility caused by over-promising.
7. Fusion Industry Landscape: Global Co-opetition from Laboratory to Market
The industrialization of fusion energy is forming a multi-actor, multi-pathway, transnational landscape. In January 2026, the "Power of Fusion · Creating the Future" Nuclear Fusion Energy Technology and Industry Conference was held in Hefei, with over 1,500 delegates in attendance. Sichuan has established an industrial innovation consortium covering the full spectrum of magnetic and inertial confinement technical routes, and has hosted an IAEA Fusion Energy Research and Training Collaboration Center. Anhui's CRAFT facility has entered a critical phase, and Zhejiang's private enterprises are deeply involved in the development of CFEDR core components.
In the emerging direction of AI for Fusion, Xinzhu Shidai's industrialization exploration is of landmark significance. Its "operating system" positioning effectively lowers the threshold for AI capabilities to enter the fusion domain, but also faces a realistic challenge: at the current stage where fusion devices have not yet achieved commercial power generation, just how large is the market size for AI for Fusion? Wang Yue's response is that the trial-and-error costs of fusion devices are growing exponentially — each generation of devices explores physical boundaries untouched by the previous generation, and the core value of AI lies in dramatically reducing the cost of trial and error. From a broader perspective, when the total financing of the fusion industry approaches the ten-billion-dollar mark and multiple fusion devices enter the engineering construction phase, "empowering fusion" itself is a track worth positioning for in advance.
8. Conclusions and Outlook
Based on the key advances in global fusion between 2024 and 2026, this review has systematically analyzed the physical pathways, engineering bottlenecks, AI control paradigms, capital ecology, and commercialization prospects. The principal conclusions are as follows:
First, controlled nuclear fusion has crossed the threshold of physical feasibility. NIF's Q = 4.13, JET's 69 MJ, EAST's 1,066 seconds, HL-3's 10²⁰ triple product, and W7-X's 43-second long pulse — together these constitute a complete chain of evidence experimentally confirming physical feasibility.
Second, engineering bottlenecks determine the commercialization timeline. Material irradiation verification depends on IFMIF-DONES, which begins operations in 2034; tritium self-sufficiency depends on ITER TBM testing after 2039. Any delay could shift the timeline to the right.
Third, Chinese fusion is entering the world's first tier, and the AI for Fusion ecosystem is taking shape first. The founding of Xinzhu Shidai marks the transition of AI for Fusion from an academic topic to an independent industrial track. Its differentiated positioning alongside enterprises such as Phaenarete — the former focused on "operating-system-level" full-stack intelligence, the latter deeply engaged in rigid physical constraints and formal verification in safety-critical scenarios — jointly propels China from "catching up" toward "leading" in the direction of fusion intelligence.
Fourth, capital acceleration must be accompanied by vigilance against bubbles. The speculative behaviors represented by Inertia Enterprises and the supplier lawsuits faced by TAE Technologies caution that the industry should establish stricter technical milestone disclosure and regulatory mechanisms, grounding capital on scientific feasibility rather than narrative-driven speculation.
Fifth, artificial intelligence is reshaping the fusion research paradigm, and safety and trustworthiness are preconditions for engineering. When fusion devices transition into nuclear engineering facilities, the interpretability and safety assurance capabilities of AI will rise from academic topics to rigid requirements for regulatory compliance.
"Nuclear fusion is always fifty years away" — after systematically examining the data, a more precise formulation is: "fifty years" measures not calendar time, but the collective will of the scientific community and the rationality of the capital ecology. The spark is already in human hands. And when this Promethean fire is finally kindled by Chinese hands, we must ensure it is held in a safe grip — one combining physical rigor, engineering robustness, and commercial integrity.
Acknowledgments: The author benefited from the rigorous comments of anonymous reviewers during the revision process and extends sincere thanks. The trusted AI research described in this paper received support from Phaenarete AI Technology Co., Ltd. Information on Xinzhu Shidai is drawn from public reporting. All omissions and shortcomings in this paper are the sole responsibility of the author.
Appendix A: Thought Experiment — Controlled Nuclear Fusion in Space
While the research community toils on Earth to overcome the bottlenecks of plasma confinement, material irradiation, and tritium self-sufficiency, a thought experiment naturally arises: What if controlled nuclear fusion were carried out in space? Space offers natural vacuum, microgravity, and limitless heat dissipation — conditions unattainable on Earth. Could this alter the global constraints of fusion engineering?
A.1 The Unique Physical Conditions of Space
In 1960, American physicist Robert Bussard proposed the concept of the interstellar ramjet — a spacecraft collecting hydrogen from the interstellar medium via electromagnetic fields as fusion fuel, enabling sustained acceleration without carrying large quantities of fuel. This concept reveals the first unique advantage of space fusion: in-situ fuel harvesting. Although interstellar hydrogen is extremely sparse (averaging roughly 1 proton per cubic centimeter), when a spacecraft reaches approximately 6% of the speed of light, an enormous collection funnel (tens of thousands of kilometers in diameter) can gather sufficient fuel for fusion.
More generally, space provides three conditions for fusion that cannot be replicated on Earth. First, natural vacuum — ground-based devices must maintain ultra-high vacuum at 10⁻⁸ Pa, which in space comes "free" — though deep-space high-energy particles may cause sputtering and activation via interaction with vacuum vessel walls, imposing additional requirements on material selection. Second, microgravity environment — current mainstream fusion devices do not face significant buoyancy-driven convection issues for the plasma itself under gravity, but for future devices employing liquid lithium blankets, microgravity facilitates uniform spreading and free-surface flow control of the liquid first wall. Third, heat dissipation capability — ground-based GW-class fusion plants require massive cooling towers, whereas in space, heat can be radiated into deep space via large radiator panels (the radiative rate scales with the fourth power of temperature), an attractive feature for compact fusion propulsion systems.
A.2 Two Pathways for Space Fusion
Pathway One: Propulsion Fusion. The Direct Fusion Drive (DFD) concept aims to provide both thrust and electrical power from a compact fusion reactor. In March 2026, the UK's Pulsar Fusion achieved first plasma in the fusion propulsion exhaust system of its Sunbird project, with an on-orbit demonstration of core components planned for 2027. The company is also collaborating with the UK Atomic Energy Authority (UKAEA) on neutron shielding and activation modeling, with UKAEA providing fusion materials modeling and radiation analysis, laying the engineering foundation for shielding and material selection for future fusion-powered spacecraft. However, the core engineering challenge facing fusion propulsion is the specific power target — 1 kW/kg is a baseline threshold, and no concept has yet demonstrated this at engineering scale.
Pathway Two: Space-Based Power Generation Fusion. Y-Combinator-backed Zephyr Fusion has proposed a space-based fusion power plant concept, aiming to provide energy for orbital industrial applications such as high-performance computing and advanced robotics. Freed from terrestrial gravity and volume constraints, space-based power plants could theoretically adopt more compact magnet designs. However, such concepts face a paradox: the current imbalance between launch costs and power generation revenue. Until reusable rocket technology further reduces launch costs, the economic viability of large-scale space power plants remains highly questionable.
A.3 Critical Bottlenecks of Space Fusion
Despite the enticing prospects, space fusion faces at least three core bottlenecks in physics and engineering.
First, particle balance and fuel replenishment. Ground-based devices can maintain plasma density through continuous fueling, whereas in deep space, interstellar hydrogen density is only on the order of 1/cc. A Bussard ramjet requires an initial velocity of roughly 6% of the speed of light to initiate an effective collection process — a speed that itself far exceeds the capabilities of current propulsion technology. A space fusion power plant without in-situ fuel sources must carry all of its deuterium-tritium fuel, negating some of the advantages of space fusion.
Second, thermal management and radiation safety. Although space heat dissipation has theoretical advantages, fusion reactor power density is extremely high (on the order of GW/m³), and arranging sufficient radiator panels within a minimal volume in a compact design is an engineering challenge. Meanwhile, shielding against 14.1 MeV fusion neutrons in space is more dependent on onboard systems (deep space lacks Earth's atmospheric protection), requiring additional mass budget for crewed missions, and the mass of the shield may dominate the total system mass.
Third, maintenance and reliability. Ground-based devices can periodically replace first-wall and divertor components, but in deep space, a fusion reactor must operate maintenance-free over its entire lifecycle. This is an unprecedented challenge in materials science — current RAFM steel in the first-wall position is expected to require replacement every 2–5 years, in stark contrast to the decades of maintenance-free operation required for deep-space missions.
A.4 Reasonable Projection and Outlook
In the foreseeable future (2026–2070), the most likely development path for space fusion is: propulsion first, power generation later; unmanned first, crewed later. In the short term (2026–2040), compact fusion propulsion concepts (such as Pulsar Sunbird) will continue advancing ground-based verification of key subsystems, with specific power gradually approaching the practical threshold. In the medium term (2040–2060), after ground-based fusion power plants (such as CFEDR and ARC) complete engineering demonstration, the engineering foundation for space fusion power plants will be substantially strengthened, and the first unmanned deep-space fusion propulsion demonstration mission may be realized. In the long term (2060–2100), if humanity possesses large-scale space industrial capability, space-based fusion power generation could complement space-based photovoltaics, providing round-the-clock power for orbital factories.
Elon Musk has bluntly stated that in space "the Sun itself is already a giant, free nuclear fusion reactor in the sky," making space-based photovoltaics the more rational choice. This criticism is forceful — yet it does not constitute a refutation of fusion. Photovoltaics and fusion solve different problems: photovoltaics are suitable for orbital power supply, while fusion propulsion aims to shorten interplanetary transit times. The two are more likely to form a complementary relationship in the future space energy landscape.
Ultimately, the thought experiment of space fusion reminds us: the significance of fusion lies not only in lighting the light bulbs of Earth, but in opening the sea routes to the stars. A civilization that has mastered controlled nuclear fusion truly possesses the energy foundation to become an interstellar species.
References
[1] Lawrence Livermore National Laboratory. Ten Ignition Results, Unprecedented Fusion Yields [EB/OL]. https://lasers.llnl.gov.
[2] Institute of Plasma Physics, Chinese Academy of Sciences. Fully superconducting tokamak achieves hundred-million-degree, thousand-second H-mode plasma operation [EB/OL]. 2025-01-21. http://lssf01.cas.cn.
[3] Oak Ridge National Laboratory. ORNL's pellet injector enables world record performance in W7-X [EB/OL]. 2025-08-28. https://www.ornl.gov.
[4] Commonwealth Fusion Systems. CFS delivers its first fusion magnet — a stronger, smaller design [EB/OL]. 2026-01-06. https://blog.cfs.energy.
[5] Southwestern Institute of Physics. Series of announcements on Huanliu-3 experimental progress [EB/OL]. 2025-03/2025-06.
[6] EUROfusion. JET sets new fusion energy record of 69 MJ [EB/OL]. 2024-02-08. https://www.euro-fusion.org.
[7] ITER Newsline. New records on Wendelstein 7-X [EB/OL]. 2025-06-10. https://www.iter.org.
[8] Ding X Y, Li G Q, Wang J, et al. Integrated physics design of China Fusion Engineering Demonstration Reactor (CFEDR) [J]. Plasma Science and Technology, 2025, 27(10): 104004.
[9] EAST Team. Accessing the density-free regime with ECRH-assisted ohmic start-up on EAST [J]. Science Advances, 2026, 12(1).
[10] Fusion Industry Association. The Global Fusion Industry in 2025 [R]. 2025.
[11] ZAP Energy. Zap Energy exceeds gigapascal fusion plasma pressures on new fusion device, FuZE-3 [EB/OL]. 2025-11-18. https://www.zapenergy.com.
[12] Xinzhu Shidai secures 60 million RMB funding, focused on AI for Fusion operating system R&D [EB/OL]. DoNews, 2026-04-20. https://www.donews.com.
[13] Installing a brain in the "artificial sun" — what kind of business is this? [EB/OL]. Huxiu APP, 2026-04-16. https://www.sohu.com.
[14] Zhao Ling. Inertia Enterprises obtains investment through false promotion based on NIF [EB/OL]. Zhihu Column, 2026-02-18. https://zhuanlan.zhihu.com.
[15] Breaking: Trump Media and TAE Technologies merger and allegations of unpaid bills [EB/OL]. Wallstreetcn, 2025-12-20. https://wallstreetcn.com.
[16] Nuclear fusion power countdown: the national team's timeline reaches 2045 — how dare private enterprises claim 2033? [EB/OL]. OFweek, 2026-03-15. https://power.m.ofweek.com.
[17] Seo J, Kim S, Jalalvand A, et al. Avoiding fusion plasma tearing instability with deep reinforcement learning [J]. Nature, 2024, 626(8000): 746-751. DOI: 10.1038/s41586-024-07024-9.
[18] Gopakumar V, Pamela S, Zanisi L, et al. Plasma surrogate modelling using Fourier neural operators [J]. Nuclear Fusion, 2024, 64(5): 056025.
[19] Citrin J, Ho A, et al. Predictive JET current ramp-up modelling using QuaLiKiz-neural-network [J]. Nuclear Fusion, 2023, 63: 066014.
[20] Ma R H, Lin Z, Cai H S, et al. Suppression of ion temperature gradient turbulence near magnetic axis via geometric curvature [J]. Nuclear Fusion, 2026. (online advance publication)
[21] Princeton Team. Regulation compliant AI for fusion: explainable image-based feedback control of divertor detachment in DIII-D tokamak [J]. Physics of Plasmas, 2026.
[22] IAEA. AI for Fusion Coordinated Research Project (CRP) [EB/OL]. 2022-2027. https://conferences.iaea.org.
[23] U.S. Nuclear Regulatory Commission. Regulatory Framework for Fusion Machines (Proposed Rule) [EB/OL]. Federal Register, 2026-02-26.
[24] Troyon F, Gruber R, Saurenmann H, et al. MHD-limits to plasma confinement [J]. Plasma Physics and Controlled Fusion, 1984, 26(1A): 209-215.
[25] Greenwald M, Terry J L, Wolfe S M, et al. A new look at density limits in tokamaks [J]. Nuclear Fusion, 1988, 28(12): 2199-2207.
[26] Shanghai Superconductor Technology Co., Ltd. IPO prospectus and capacity expansion announcement [EB/OL]. 2025-06.
[27] IFMIF-DONES. EU to invest €202m to IFMIF-DONES; multi-party agreement signed [EB/OL]. 2025.
[28] China Atomic Energy Authority et al. China Fusion Energy Technology Development Roadmap [R]. Beijing, 2026.
[29] Pulsar Fusion. First plasma achieved in Sunbird fusion propulsion exhaust system [EB/OL]. 2026-03-26. https://www.militaryaerospace.com.
[30] Pulsar Fusion partners with UKAEA on Sunbird fusion propulsion modeling [EB/OL]. Military Aerospace, 2026-03-06. https://www.militaryaerospace.com.
[31] Bussard R W. Galactic Matter and Interstellar Flight [J]. Astronautica Acta, 1960, 6(4): 179-192.
[32] Nuclear Fusion from First Principles — Vol.9: Fusion Propulsion [EB/OL]. Dev.to, 2026-02-27. https://dev.to.