Direct Schrödinger–Poisson simulations generate 1,000 three-dimensional fuzzy-dark-matter halos and demonstrate how high-performance computing can turn an exotic particle hypothesis into an observational test

Determining the nature of dark matter, specifically whether it consists of conventional cold, massive particles or ultralight quantum waves, remains one of the most formidable and computationally intensive challenges in modern cosmology. 

A recent study published in The Astrophysical Journal Letters (https://iopscience.iop.org/article/10.3847/2041-8213/ae9a9e) demonstrates how high-performance numerical simulations can transition this inquiry from theoretical speculation to an observationally testable framework. Jiajun Zhou and his collaborators conducted the first calculations of gravitationally lensed images derived directly from three-dimensional fuzzy-dark-matter (FDM) density fields, evolved via the Schrödinger–Poisson equations. By computationally evolving the dark-matter wave field rather than relying on statistical approximations, the researchers were able to predict how these quantum structures perturb the images of distant, gravitationally lensed quasars. 

This work marks a significant computational milestone, as it effectively tests whether the intricate structures generated by quantum wave evolution persist through numerical processing to produce observable consequences that align with astronomical measurements. While these findings provide encouraging support for the fuzzy-dark-matter hypothesis, the authors emphasize that further research is essential to fully validate these results.

From particles to waves

Fuzzy dark matter, also called wave dark matter, proposes that dark matter is composed of extremely light particles whose quantum-mechanical de Broglie wavelengths can become comparable to astrophysical scales.

For a particle mass of 10⁻²² electronvolts, the characteristic de Broglie wavelength in the simulated galaxy-scale system is roughly 100 parsecs. That is an extraordinary scale for a quantum effect: roughly hundreds of light-years.

At these scales, the dark-matter halo cannot be treated simply as a collection of classical particles.

It must be treated as a coherent wave field.

That changes the computational problem fundamentally.

The researchers solve the coupled Schrödinger–Poisson equations, in which the complex wave function describes the FDM field while the gravitational potential is obtained from the density generated by that field.

The density is proportional to the squared magnitude of the wave function:

[
\rho = M|\psi|^2.
]

The gravitational field generated by that density then feeds back into the evolution of the wave itself.

This creates a nonlinear, self-gravitating wave problem.

It is precisely the kind of problem for which numerical resolution and algorithmic efficiency become inseparable from the scientific result.

A 512³ computational grid

The researchers employ a global Fourier pseudospectral method.

The choice is important from an HPC perspective.

Pseudospectral methods represent the field in Fourier space and can achieve high spectral accuracy for smooth wave fields while reducing numerical diffusion. The paper states that this approach is particularly suitable for the galaxy-scale lensing problem being investigated.

The production calculations use a 512³ grid, equivalent to more than 134 million spatial cells.

The simulation volume is a cube approximately 40 kiloparsecs on a side, and each realization is evolved for approximately 3.3 billion years of physical evolution time. The simulations use a total dark-matter mass of approximately 4 × 10¹¹ solar masses and investigate particle masses of 10⁻²² and 10⁻²³ eV.

The grid resolution is not arbitrary.

The researchers require the computational cell size to be smaller than the de Broglie wavelength, with several cells needed across the wavelength to resolve the interference pattern.

The grid must also resolve the physical scale corresponding to the observed tens-of-milliarcsecond positional anomalies in the gravitationally lensed system.

This is a classic HPC constraint: the physics dictates the resolution, and the resolution dictates the computational cost.

Reducing the cell size increases the number of grid points in three dimensions rapidly. A modest increase in linear resolution therefore produces a much larger increase in memory requirements and computational work.

1,000 universes inside the computer

Perhaps the most revealing computational figure in the study is not 512³.

It is 1,000.

The researchers generated 1,000 independent initial conditions, each constructed from five randomly distributed three-dimensional Gaussian wave packets.

Every realization was then evolved through the full Schrödinger–Poisson calculation to produce an independent three-dimensional fuzzy-dark-matter halo.

This transforms the project from a single numerical experiment into a statistical computational campaign.

The objective is not merely to produce one halo that happens to resemble the observations.

Instead, the researchers ask how often the structures naturally generated by the underlying equations produce lensing configurations compatible with the observed system.

That distinction is important.

A single simulation can demonstrate possibility.

A large ensemble begins to address probability.

The computer must preserve the wave physics

Numerical integration becomes particularly important because the researchers are not simply tracking the motion of individual particles.

They are evolving a wave field whose phase and interference structure matter.

The simulations therefore use a split-step pseudospectral method. During each time step, the kinetic and gravitational-potential operators are applied separately. The time step is constrained by the fastest phase oscillations associated with the kinetic and gravitational terms, with a safety factor imposed to avoid phase aliasing.

That is an HPC issue as much as a physics issue.

A simulation can run faster by taking larger time steps or using lower spatial resolution.

But if those shortcuts erase physically relevant wave structure, the resulting gravitational lensing prediction can become numerically precise but physically wrong.

The researchers instead make the numerical resolution part of the physical model.

From a three-dimensional supercomputer field to a two-dimensional sky

The computational workflow does not end when the dark-matter halo has been evolved.

The researchers then have to turn the three-dimensional simulation into an observable lens.

For each simulated halo, they determine its principal axis and rotate the three-dimensional density field through representative viewing directions.

The density is projected along the line of sight to generate a two-dimensional convergence map, after which the gravitational lens equation is solved to generate simulated image positions. The open-source lenstronomy package performs the lensing calculations.

This creates a computational pipeline that can be summarized as: wave equation → gravitational potential → three-dimensional density field → viewing geometry → projected mass → lens equation → multiple images → statistical comparison with observations.

The researchers sample 103 representative viewing directions and 10⁴ source positions during the forward-modeling process.

At this point, the project begins to resemble a modern scientific computing workflow more than a conventional analytic astronomy calculation.

The supercomputer is effectively generating synthetic observations from first-principles numerical evolution.

Matching the geometry without fitting away the physics

The study introduces another computationally interesting element.

The researchers compare the simulated and observed four-image configurations using pairwise-distance invariants and Procrustes alignment.

This allows translations, rotations, and reflections that do not represent physical differences to be removed from the comparison.

For four images, the six pairwise distances provide a complete set of geometric invariants for the relative configuration. The researchers use these distances to identify the source position that best reproduces the observed geometry and then apply Procrustes alignment to quantify the remaining image-position anomaly.

That is an important numerical safeguard.

Without it, the calculation could incorrectly interpret a simple coordinate-frame difference as evidence that the dark-matter model is wrong.

The computational machinery therefore has to be careful not only about solving the equations, but also about comparing the output with observational data in a statistically meaningful way.

The result: wave simulations reproduce the observed lens

The target is HS 0810+2554, a quadruply lensed quasar system containing two compact radio sources.

High-resolution radio observations have measured eight lensed radio images with sufficient astrometric precision to expose discrepancies between the observations and smooth conventional lens models.

For fuzzy dark matter with a particle mass of 10⁻²² eV, the wave-evolved halos produce median image-position anomalies of approximately 12 and 6 milliarcseconds for the two radio components.

Some realizations produce anomalies below approximately 3 milliarcseconds, within the roughly 3σ observational uncertainty level used in the analysis.

The comparison is particularly interesting because the simulations are not tuned to force the halos into the observed configuration.

The halos evolve from randomly generated initial conditions.

The researchers report that the wave simulations can reproduce the observed image positions to within approximately 3σ without fitting the internal state of the simulated halo to the observations.

By comparison, the Gaussian-random-field approximation generally produces larger positional fluctuations, while the best-fit smooth NFW model produces substantially larger discrepancies for most of the observed images.

Particle mass becomes a computationally observable quantity

One of the most important results is the sensitivity to the assumed particle mass.

When the researchers reduce the FDM particle mass from 10⁻²² to 10⁻²³ eV, the de Broglie wavelength increases and the resulting density fluctuations occur on larger physical scales.

The simulated lensing position anomaly rises to a median of approximately 50 milliarcseconds, roughly four times the value produced in the 10⁻²²-eV case.

This is precisely where HPC becomes scientifically powerful.

The computer is not merely illustrating a theory.

It is establishing a mapping: particle mass → wave scale → density structure → gravitational potential → image displacement.

That mapping gives astronomers a route toward constraining the mass of a hypothetical dark-matter particle through observations.

The paper concludes that future high-angular-resolution lensing observations could narrow the allowed FDM mass range.

Why Gaussian approximations are not enough

Previous FDM lensing calculations have often relied on Gaussian random fields because they are computationally efficient.

The approach can reproduce broad statistical characteristics of the fluctuations.

But it does not actually evolve the underlying three-dimensional wave system.

The distinction becomes important at higher precision.

The full Schrödinger–Poisson calculation naturally retains spatial correlations, mode coupling, and non-Gaussian higher-order structure generated during the evolution.

The researchers find that Gaussian random fields remain useful as efficient statistical approximations for moderate-precision calculations.

But for precision gravitational-lensing predictions, the direct wave calculation becomes increasingly important.

This is a familiar pattern in computational science.

Reduced-order models can provide enormous computational savings.

But as observational precision improves, the approximations that were once adequate can become the dominant source of error.

The HPC challenge is about to become larger

The authors explicitly acknowledge that full three-dimensional wave simulations are computationally expensive.

Future work will investigate larger simulation boxes and more efficient numerical approaches while preserving sufficient accuracy in the strong-lensing region.

That points directly toward the next generation of HPC requirements.

The current calculation uses a 512³ grid.

Moving toward larger physical volumes while maintaining comparable spatial resolution would increase the number of grid cells dramatically.

Increasing the resolution from 512³ to 1024³, for example, increases the number of spatial cells by a factor of eight.

Moving to 2048³ would increase it by another factor of eight.

And the problem is not simply memory.

Every time step requires large-scale numerical operations, including Fourier transforms and repeated evaluation of the gravitational potential. The long physical integration time compounds the workload.

An ensemble of thousands of realizations would turn the problem into a substantial distributed-computing campaign.

This is precisely where modern HPC architectures, large memory systems, high-bandwidth interconnects, accelerators, distributed FFT libraries and efficient parallel I/O, become critical.

China’s expanding computing ambitions

The scientific work is also part of a broader Chinese computational environment that is placing increasing emphasis on large-scale intelligent and scientific computing.

The research itself includes authors from Beijing Normal University and Tsinghua University, while the paper acknowledges support from China’s National Key Research and Development Program, the National Natural Science Foundation of China and the Strategic Priority Research Program of the Chinese Academy of Sciences.

That institutional investment exists alongside a much broader national effort to expand computing infrastructure.

In June 2026, China’s State Council called for accelerating breakthroughs in key AI technologies and specifically urged construction of ultra-large-scale intelligent computing clusters. 

In September, a separate State Council meeting emphasized that computing networks provide fundamental support for artificial intelligence and called for improved computing infrastructure, coordination between computing capacity and electricity supply, and integration of computing and communications networks. 

China’s information and communications development plan released this month sets a 2030 target of 9,800 EFLOPS of intelligent computing capacity and calls for continued development of a nationwide integrated computing-power network. 

Those targets concern AI and national computing infrastructure rather than the specific astrophysical simulations described in the paper. Nevertheless, they illustrate the scale of the computing environment China is attempting to develop.

For scientific HPC, that matters.

The same fundamental infrastructure required for enormous AI workloads, high-bandwidth memory, accelerators, high-speed networking, storage, and large-scale parallel computing, can also expand the computational envelope available to astronomy, cosmology and fundamental physics.

Supercomputing as an instrument for dark-matter physics

The significance of this work extends beyond the study of fuzzy dark matter, representing a pivotal shift in the field of computational astrophysics. As modern instrumentation provides observations of unprecedented precision, capable of distinguishing between physical models previously obscured by measurement uncertainty, simulations must evolve to achieve commensurate realism. In the context of fuzzy dark matter, this necessitates moving beyond statistical approximations in favor of the direct evolution of the underlying wave field.

In this framework, the supercomputer functions as a laboratory where candidate universes are constructed, simulated, and observed. The study by Zhou et al. illustrates a rigorous numerical pipeline: evolving three-dimensional dark-matter halos, applying varied viewing geometries, projecting these into gravitational lenses, and benchmarking the results against milliarcsecond-scale astronomical data. This approach underscores a fundamental reality of contemporary high-performance computing: scientific advancement increasingly relies not merely on scaling computational capacity, but on resolving governing equations with sufficient fidelity to generate observationally testable predictions. 

Because the universe does not permit direct experimental manipulation of dark-matter particles, researchers must instead construct numerical proxies to evaluate theoretical consequences. By demonstrating that wave-driven gravitational structures produce measurable shifts in quasar images, this research establishes a vital mapping between particle properties and observable phenomena. Future progress will require simulations that are not only larger in scale but also more robust, statistically comprehensive, and deeply integrated with observational data. Ultimately, the trajectory of dark-matter research may depend on the extent to which the current computational frontier can be expanded.

As Washington embraces the language of “superintelligence,” Alibaba unveils a full-stack computing strategy to scale machine reasoning, from trillion-parameter models and recursive self-improvement to 20 GW of data-center capacity.

The term superintelligence gained diplomatic prominence on Tuesday, as artificial intelligence emerged as a central theme alongside international security and global governance at the United Nations. During the 81st session of the UN General Assembly in New York on September 22, 2026, President Donald Trump announced that the United States would formally adopt the term superintelligence in official documentation, asserting that the technology carries implications far more significant than conventional terminology implies.

Meanwhile, in Hangzhou, China, Alibaba articulated a more granular vision for the hardware needed to support increasingly advanced machine intelligence. Alibaba’s comprehensive AI roadmap encompasses a vertically integrated strategy, ranging from custom processors and high-speed networking to massive-scale model training, storage solutions, autonomous agents, and recursive self-improvement. Key strategic objectives include scaling future Qwen models to 5–10 trillion parameters, developing proprietary AI accelerators, deploying supernode architectures capable of supporting clusters of up to 500,000 accelerator cards, and achieving a global data-center capacity of 20 gigawatts by 2032. 

For the supercomputing industry, these developments signify a fundamental shift: the emerging global rivalry in artificial intelligence is evolving into a competitive race for foundational compute infrastructure. Furthermore, China is signaling a clear, strategic commitment to expanding its domestic capacity to supply this critical infrastructure.

The supercomputer behind “superintelligence”

Alibaba’s announcement at its Apsara Conference is notable because it does not treat AI as merely a software problem.

The company is attempting to vertically integrate the stack.

At the accelerator level, Alibaba’s T-Head semiconductor division introduced the Zhenwu V900, an AI processor designed for both training and inference. Alibaba says the processor delivers three times the performance of its Zhenwu M890 predecessor and includes 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth.

The processor supports FP8 and FP4 data formats alongside higher-precision computation, allowing the same architecture to target both computationally expensive model training and lower-precision inference workloads. Mass production is scheduled for the first quarter of 2027, according to Alibaba. 

Those numbers matter because modern AI performance is increasingly constrained not simply by arithmetic throughput, but by how quickly enormous quantities of model state can move through the system.

A 10-trillion-parameter model is not simply a larger version of today’s language model.

It becomes a distributed-memory supercomputing problem.

The system must move weights, activations, gradients, optimizer states and training data across thousands, or potentially hundreds of thousands of processors while keeping the expensive accelerators busy.

Every byte that has to travel unnecessarily costs time, energy and money.

That makes memory capacity, memory bandwidth, network bandwidth, collective communication and storage throughput just as important to the AI system as raw floating-point performance.

Alibaba’s roadmap reflects that reality.

From AI chips to AI supernodes

Alibaba’s new supernode architecture combines the Zhenwu V900 processor with its ICN Switch, Panmai SmartNIC and Zhenyue SSD controller.

The goal is system-level integration.

Alibaba says the architecture can support a supernode cluster containing as many as 500,000 cards. 

That is an extraordinary scale.

At that point, the question is no longer whether an individual accelerator is fast.

The question becomes whether the entire machine can behave like one coherent computational system.

Large-scale AI training requires synchronization among thousands of processors. Matrix operations must be distributed, intermediate results exchanged, parameters synchronized and datasets continuously supplied. Network congestion, memory stalls, storage latency and failed components can all reduce effective utilization.

This is classic supercomputing territory.

The AI industry is therefore rediscovering many of the problems HPC engineers have worked on for decades: parallelism, locality, interconnect topology, distributed memory, collective communication, checkpointing, fault tolerance, storage bandwidth and energy efficiency.

The difference is scale and workload.

A 100-petabit network

Alibaba’s proposed AI infrastructure includes HPN 8.0 Pro, its proprietary networking architecture.

The company says the system provides 100 petabits per second of aggregate bandwidth, while a single cluster can support more than 130,000 network ports operating at 800 Gb/s.

Alibaba also says the architecture incorporates redundancy designed to prevent optical-transceiver and link failures from interrupting service. 

That is not networking as an accessory to the supercomputer.

It is part of the computer.

At massive AI scale, the network becomes the fabric through which the computational workload itself is executed.

The same principle has driven the evolution of classical supercomputers from relatively independent nodes toward tightly coupled systems with increasingly sophisticated interconnects.

AI is pushing that architecture into another regime.

Storage becomes part of the intelligence engine

Alibaba is also targeting one of the least glamorous, and most important, parts of the AI stack: storage.

Its Cloud Parallel File Storage system, or CPFS, is designed for AI training and is advertised as capable of delivering hundreds of terabytes per second of throughput and hundreds of millions of I/O operations per second.

Alibaba says the architecture can reduce enterprise AI storage costs by as much as 69 percent. Those are company-reported figures and should be understood as such. 

The importance of this is straightforward.

A giant AI model does not train in isolation.

Training pipelines continuously consume enormous datasets, generate checkpoints, write intermediate information and feed data to distributed accelerators.

If storage cannot keep up, processors wait.

And when a machine containing tens of thousands of expensive accelerators is waiting for data, the economics become ugly very quickly.

Supercomputing has long understood this principle.

The fastest processor in the world is not particularly useful if the rest of the machine cannot feed it.

Qwen moves toward trillion-parameter territory

The hardware roadmap exists to support an equally aggressive model roadmap.

Alibaba says Qwen 4 is currently in training, while subsequent Qwen 4.5 and Qwen 5 generations are projected to scale toward 5 trillion to 10 trillion parameters. 

Parameter count alone does not establish intelligence.

More parameters do not automatically mean a more capable system, and model quality depends on architecture, training data, optimization, inference techniques and evaluation methodology.

But enormous models dramatically increase the computational resources required for training and serving them.

The important development is therefore not simply the number of parameters.

It is the attempt to build an infrastructure ecosystem capable of sustaining models at that scale.

The more consequential development: machines improving machines

Perhaps the most intriguing, and concerning, from a supercomputing perspective is Alibaba’s emphasis on recursive self-improvement, or RSI.

Alibaba says Qwen3.8-Max completed 33 automated improvement cycles over more than a month, covering pipeline design, data validation, experimentation, and error diagnosis. The company reports that its Artificial Analysis score increased from 40 to 45 following autonomous training optimization and post-training techniques. 

Alibaba also describes an experiment in which a model worked through an entire chip-design lifecycle for more than 60 hours, making more than 10,000 electronic-design-automation tool calls.

The resulting chip design, according to Alibaba, reduced chip area by 42 percent without compromising performance. 

This is where the phrase superintelligence begins to acquire a distinctly HPC meaning.

The important transition may not be from one large model to an even larger model.

It may be from human-directed computation to increasingly autonomous computational experimentation.

Instead of engineers designing every experiment, an AI system can propose an experiment, execute it, evaluate the result, identify an error, modify its approach and run another experiment.

Then another.

And another.

The computational infrastructure becomes the laboratory.

China is building for the long game

Alibaba’s announcement should not be interpreted as evidence that China has already achieved artificial superintelligence.

It has not established that.

What it does demonstrate is an increasingly explicit Chinese strategy to expand AI capabilities by attacking the problem at multiple layers simultaneously.

China’s 2026–2030 Five-Year Plan calls for stronger AI research, improved model architectures and algorithms, large-scale intelligent-computing infrastructure, high-performance AI resources and consideration of ultra-large-scale intelligent computing clusters. It also calls for advances in AI agents, multimodal systems, embodied intelligence and exploration of artificial general intelligence. 

In June, China’s State Council called for accelerating breakthroughs in key AI technologies and expanding construction of ultra-large-scale intelligent-computing clusters. 

And in September, China’s Ministry of Industry and Information Technology announced an AI-focused software-industry action plan targeting broader deployment of AI development tools and agent-based software applications. 

Alibaba’s roadmap fits into that larger technological environment, although Alibaba remains a commercial company and its roadmap should not automatically be treated as a statement of Chinese government capability.

The distinction matters.

But the direction is difficult to miss.

China is simultaneously pursuing models, accelerators, CPUs, networking, storage, data centers, AI agents and applications.

The 20-gigawatt problem

Perhaps the most revealing number in Alibaba’s announcement is not 10 trillion parameters.

It is 20 gigawatts.

Alibaba CEO Eddie Wu said the company aims to exceed 20 GW of global data-center capacity operated by Alibaba Cloud by 2032 to support growing AI demand. 

Twenty gigawatts is a statement about physical infrastructure.

It means electricity generation.

It means substations.

It means cooling.

It means high-voltage distribution.

It means land, fiber, networking, storage and thousands upon thousands of servers.

It means that the race toward more capable AI is simultaneously becoming an industrial race over energy and infrastructure.

The computational revolution is becoming an electrical-engineering problem.

America and China are converging on the same computational reality

That is what makes today’s developments at the United Nations and Alibaba’s Apsara Conference particularly significant.

The political vocabulary may be changing.

The engineering vocabulary is not.

Whether policymakers call it artificial intelligence, machine intelligence, advanced AI or “superintelligence,” the underlying technology still requires processors, memory, networks, storage, power, and cooling.

And increasingly, it requires enormous amounts of all of them.

The United States remains deeply invested in frontier AI and AI infrastructure, while China is pursuing its own path toward large-scale intelligent computing. International discussions are simultaneously turning toward questions of AI safety, governance and control. UN Secretary-General António Guterres warned Tuesday that AI represents one of four major global “tests of power” and called for international cooperation on AI governance. 

China’s President Xi Jinping similarly argued at the July 2026 World AI Conference that AI presents both opportunities and governance challenges, while calling for expanded AI innovation, computing infrastructure, international cooperation and systems intended to keep AI secure and controllable. 

That creates a difficult technological paradox.

The world is simultaneously trying to accelerate AI capability and control its consequences.

Those objectives can pull in opposite directions.

Supercomputing is becoming the strategic infrastructure underneath AI

The High-performance computing community is facing an increasingly clear reality: the next generation of artificial intelligence will not be achieved solely through algorithmic innovation, but rather through the construction of increasingly sophisticated supercomputing systems. The winning architectures will be those capable of coordinating processors at an unprecedented scale, managing massive data bandwidth, optimizing storage for enormous datasets, mitigating communication overhead, ensuring fault tolerance, and operating within stringent power constraints. 

Furthermore, if recursive self-improvement becomes a primary component of model development, machines may soon begin designing the very experiments that determine the architecture of future intelligence. This represents a profound shift. For decades, supercomputers have served as humanity's primary instruments for exploring complex scientific phenomena, from climate modeling to materials science. Now, the supercomputer itself is becoming an active participant in the research process. Alibaba’s roadmap, characterized by trillion-parameter models, massive accelerator clusters, high-speed networking, and multi-gigawatt power requirements, illustrates this trajectory. The core challenge is no longer merely the growth of AI, but the rapid evolution of computational infrastructure into a new class of global industrial system. As global discourse continues to define the terminology of this technology, the structural foundation is already being laid, forcing the world to determine whether it can build this computational capacity quickly enough to effectively understand and govern the intelligence it is creating.