JERA, Dell and RHAELM target $15 billion Chiba hyperscale AI facility as Japan experiments with a new model for building compute where electricity is already available

The next generation of AI supercomputing may not begin with a GPU.

It may begin with a power plant.

In a strategic collaboration, Japan’s primary power generator, JERA, has partnered with Dell Technologies and AI infrastructure developer RHAELM Holdings to pioneer a novel hyperscale computing model. This initiative integrates electricity generation, electrical infrastructure, advanced cooling, data center construction, and rack-scale AI computing into a single, cohesive deployment rather than treating them as fragmented projects.

The model’s inaugural implementation is a planned 400-megawatt AI data center in Chiba, situated adjacent to JERA’s existing thermal power station. This ambitious undertaking represents a total capital commitment exceeding $15 billion, covering land acquisition, power infrastructure, facility development, and AI compute capacity. According to JERA and Reuters, operations are projected to start in phases by 2028, with the facility achieving its full 400-MW capacity by 2029.

For the supercomputing industry, however, the financial scale and construction timeline are secondary to the underlying innovation. The partners are addressing one of the AI era's most critical challenges: rapidly converting raw electrical power into scalable, high-performance computing capacity.

The power bottleneck becomes the computing bottleneck

For decades, supercomputer deployments have generally been treated as technology projects. A facility is built, electrical capacity is secured, cooling systems are installed, networks are provisioned, and compute systems are eventually delivered and integrated.

AI is putting pressure on that sequence.

Modern AI clusters can require enormous amounts of power concentrated into relatively small physical footprints. Thousands of accelerators operating simultaneously create not only a power requirement but also a corresponding cooling, networking, storage and electrical-distribution problem.

As AI systems become larger, the availability of processors is therefore only one part of the equation.

There has to be somewhere to plug them in.

The Chiba project attacks that problem from the opposite direction. Rather than first developing a conventional grid-connected data center and then waiting for sufficient electrical capacity to become available, the partners plan to place the compute infrastructure alongside an existing generation asset.

JERA says the project will operate behind the meter, using power capacity from its operating generation asset. The company argues that this can allow AI computing capacity to come online years earlier than a conventional grid-connected development schedule.

That concept changes the starting point for hyperscale AI construction.

Instead of:

Data center → grid connection → wait for power → compute

the model becomes:

Generation → electrical infrastructure → cooling → compute

The distinction could become increasingly important as AI data-center developers encounter lengthy grid-interconnection queues and constrained transmission capacity.

400 MW is a supercomputing-scale number

The Chiba facility’s 400 MW power capacity provides a useful indication of the scale of infrastructure being contemplated.

It is important, however, to distinguish between facility power capacity and compute power.

A 400-MW data center does not mean 400 MW of GPUs or accelerators.

The site’s electrical envelope must support much more than processors. It encompasses power conversion and distribution, cooling infrastructure, networking, storage, CPUs, memory, accelerator systems, building systems, and other facility loads.

The actual IT load will depend on the final design and power-utilization characteristics of the facility.

Nevertheless, 400 MW represents an enormous potential computing envelope.

At high accelerator densities, the facility could support an extraordinary concentration of AI compute. The precise number of GPUs or other accelerators cannot be calculated from the 400-MW figure alone because it depends on the eventual rack architecture, accelerator selection, networking topology, storage requirements, and facility overhead.

That uncertainty is important.

400 MW is a statement about the infrastructure’s ability to deliver power, not a published specification for the number of processors inside it.

The rack becomes the building block

Dell’s role in the project points toward another important architectural change.

The company will provide standardized, rack-scale AI infrastructure, with JERA describing Dell’s AI Factory as the mechanism for standardizing the compute layer.

That is significant because the rack is becoming an increasingly important unit of AI infrastructure.

Traditional data-center architecture often treated servers as relatively interchangeable building blocks. AI clusters are different. Accelerator systems require carefully engineered relationships among:

• accelerators;
• host CPUs;
• high-bandwidth memory;
• local and parallel storage;
• high-speed fabric;
• power distribution;
• thermal management;
• software;
• orchestration; and
• cluster management.

The result is closer to a supercomputer node architecture multiplied across thousands of systems than to a conventional server farm.

At hyperscale, the physical rack itself becomes a systems-engineering object.

Power delivery, liquid cooling, network topology, and compute density must all be designed around the rack.

Dell’s standardized rack-scale approach is therefore intended to make the compute component repeatable.

RHAELM’s role is to integrate that compute infrastructure with the site, power and data-center construction.

JERA supplies the power and physical location.

The three pieces form the proposed deployment model.

Cooling becomes a first-class HPC problem

Power is only half of the physical equation.

Nearly every watt consumed by an accelerator eventually becomes heat that must be removed.

That means a 400-MW-class AI facility is also a massive thermal-management project.

As accelerator densities increase, traditional air cooling becomes increasingly difficult to apply economically at the highest rack densities. Liquid cooling, whether direct-to-chip, cold-plate or other advanced architectures, can provide substantially greater heat-transfer capability.

JERA’s announcement specifically identifies cooling infrastructure as one of the elements the partners intend to standardize alongside power generation, electrical infrastructure and AI computing.

That detail may ultimately prove more important than it first appears.

A future AI supercomputer is not simply a collection of processors connected by a network.

It is a coupled physical system:

electricity → voltage conversion → compute → heat → liquid cooling → heat rejection

Every stage affects the others.

A higher-density accelerator rack may deliver more AI performance per unit of floor space, but it simultaneously increases electrical and thermal engineering requirements.

The ability to standardize those systems could become one of the keys to shortening deployment times.

The network will determine whether 400 MW becomes useful compute

There is another critical layer between electricity and useful AI performance: interconnect.

Large AI models are increasingly distributed across thousands of accelerators. Training efficiency depends not simply on the raw compute capability of each accelerator but on how efficiently the cluster can exchange data.

That makes the network a component of the supercomputer.

High-bandwidth links connect accelerators within systems and racks, while larger-scale fabrics connect nodes across the cluster. Storage systems must simultaneously feed training pipelines with enormous quantities of data.

Consequently, a 400-MW facility cannot be evaluated simply by asking how many accelerators can fit inside it.

The more meaningful question is:

How much synchronized AI computation can the entire facility sustain?

That depends on the interaction among compute, memory, networking, storage, software and power.

The Chiba architecture is therefore better understood as a potential hyperscale AI supercomputing platform than simply a very large data center.

From one power station to a national AI infrastructure model

This is where the project becomes considerably more ambitious.

JERA, Dell and RHAELM are not presenting Chiba solely as a one-off campus.

The companies signed an MoU to develop a standardized framework for building AI infrastructure at national scale in Japan.

JERA and RHAELM intend to explore deploying the model at additional JERA sites, with an ambition of supporting multi-gigawatt-scale AI infrastructure across Japan during the 2030s.

That changes the significance of the 400-MW Chiba installation.

It becomes a prototype.

If the partners can establish a repeatable relationship between:

power station + site + electrical infrastructure + cooling + data center + rack-scale AI

then the next facility potentially does not have to begin as a blank-sheet engineering exercise.

The architecture can be repeated.

That is the industrialization opportunity.

Japan’s answer to the “speed to power” problem

The phrase emerging around the project is “speed to power.”

JERA’s CEO Yukio Kani has identified access to large-scale reliable energy as one of the critical constraints on AI infrastructure deployment. The company’s argument is straightforward: if the computing facility can be placed alongside existing generation, developers can potentially avoid some of the delays associated with conventional grid-connected development.

That is becoming a fundamental issue for the entire AI industry.

The global AI buildout is creating demand for data centers faster than traditional energy and transmission infrastructure can necessarily be developed.

The resulting problem is not simply a shortage of data-center buildings.

It is a synchronization problem.

Compute can be manufactured faster than power infrastructure can be connected.

AI developers therefore increasingly have to solve several schedules simultaneously:

• accelerator availability;
• data-center construction;
• grid interconnection;
• transmission capacity;
• power generation;
• cooling equipment;
• network equipment;
• storage;
• and capital deployment.

A delay in any one of those components can delay the entire AI cluster.

Chiba attempts to collapse several of those schedules into one integrated project.

The LNG connection

There is another unusual element to the Japanese approach.

JERA’s business spans much of the LNG value chain, including procurement, shipping, receiving and regasification infrastructure, and power generation.

The company’s announcement describes the initiative as an integrated model connecting LNG supply with AI computing.

That makes the project fundamentally different from a conventional hyperscaler simply purchasing electricity from the grid.

The proposed architecture links fuel supply, generation and compute infrastructure more tightly together.

The implication is significant for AI infrastructure planning: energy procurement itself can become part of the supercomputer architecture.

The processor may sit thousands of miles away from a natural-gas field, but the reliability of the resulting AI cluster ultimately depends on the physical energy chain that supplies its electricity.

$15 billion is the infrastructure, not a GPU purchase

The project’s financial headline deserves careful interpretation.

The announced investment exceeds $15 billion across all phases, but that amount encompasses the entire Chiba development, including land, power infrastructure, construction, and AI computing capacity. It is not a $15-billion Dell hardware order.

Apollo Global Management is expected to serve as a strategic investment and financing partner for RHAELM.

That financing structure reflects another reality of hyperscale AI:

The next generation of supercomputers is becoming an infrastructure-finance problem as much as a semiconductor problem.

A 400-MW facility requires enormous capital commitments before it can generate computing revenue.

Investors therefore have to underwrite not merely processors and servers but long-lived physical infrastructure, power contracts, cooling systems, buildings and operating capacity.

A potential template beyond Japan

The companies explicitly intend to explore applying the model beyond Japan over time.

That could make Chiba an interesting experiment for other electricity-constrained markets.

The underlying idea is not geographically complicated:

Find large, reliable generation.

Place compute beside it.

Standardize the electrical, cooling, and compute architecture.

Repeat.

That model could be attractive wherever conventional grid expansion is becoming the limiting factor for AI infrastructure.

It also potentially changes the geography of supercomputing.

Historically, compute clusters have often been located according to network connectivity, proximity to users, real-estate costs or access to established data-center markets.

The AI era may increasingly add another dominant variable:

Where is the power?

The supercomputer of the future may be built around the megawatt

For the HPC industry, perhaps the most important lesson from Chiba is that compute capacity is increasingly measured in two dimensions simultaneously.

One is familiar:

How much computation can the machine perform?

The other is becoming unavoidable:

How many megawatts can the site deliver continuously?

A cluster can have the world’s fastest accelerators and still fail to become a useful supercomputer if it cannot supply sufficient power, remove sufficient heat, move data fast enough, or maintain the system at high utilization.

That makes power, cooling, and interconnect first-class components of computational architecture.

The Chiba project puts that concept into physical form.

A power station becomes the foundation.

A data center becomes the computational shell.

Rack-scale AI systems become the building blocks.

High-speed networks connect them into a distributed machine.

And software turns the entire installation into a usable computational resource.

That is a supercomputer.

Just a very, very large one.

Chiba could be the beginning, not the destination

The proposed project is significant, representing a 400 MW capacity, a capital expenditure exceeding $15 billion, and a phased operational timeline commencing in 2028, with full functionality expected by 2029.

However, the broader implication of this initiative is the potential to establish the Chiba architecture as a repeatable model for infrastructure development. Should this effort succeed, Japan will not merely be constructing a hyperscale AI data center; it will have pioneered a standardized mechanism for repurposing existing power infrastructure into robust, large-scale sovereign AI computing capacity.

This represents a profound shift in the economics of supercomputing. While computation and semiconductor manufacturing capacity have historically served as the primary constraints, electricity has emerged as the critical scarce resource. Consequently, the competition to develop next-generation AI supercomputers will likely be determined not only within semiconductor fabrication plants or server assembly facilities but at the power stations themselves. In Chiba, Japan is asserting that the most efficient pathway to expanding AI compute capacity begins with a foundational asset: the power plant.

The emergence of the first stars marked a fundamental transition in the universe's evolution, as these stellar bodies forged the first heavy elements, altered the chemistry of primordial gas, and catalyzed the transformation of a nearly dark cosmos into the structured universe we observe today. Reconstructing this transition through numerical modeling remains a significant computational undertaking. By utilizing the MEGATRON suite of high-resolution cosmological radiation-hydrodynamics simulations, researchers are leveraging supercomputing to investigate how the chemical signatures of the universe’s first stars may be preserved within the smallest galaxies orbiting the Milky Way. 

This research underscores a critical advancement for the high-performance computing community: accurately modeling the earliest stages of galaxy formation requires simulations that simultaneously resolve gravity, hydrodynamics, adaptive mesh refinement, radiative transfer, stellar formation, stellar feedback, and non-equilibrium chemistry. MEGATRON integrates these complex physical processes at unprecedentedly small spatial scales. 

Consequently, this computational experiment tracks the evolution of a proto-Milky-Way environment from the early universe to a redshift of approximately z ≈ 8, resolving structures on a parsec scale. Ultimately, these simulations provide a robust numerical laboratory for addressing a foundational inquiry: how did the first stars dictate the chemical composition of the smallest galaxies present in the modern universe?

A universe inside the supercomputer

MEGATRON is not a conventional galaxy simulation that represents a large volume of the universe at relatively coarse resolution.

It uses a zoomed cosmological simulation focused on a region that will ultimately develop into a Milky Way-mass system.

The dark-matter particle mass is approximately 2.5 × 10⁴ solar masses, while the spatial resolution evolves from roughly 2.5 parsecs at z = 25 to approximately 5 parsecs at z = 8.

That resolution is critical.

The researchers are attempting to resolve the environments where Population III, or first-generation, stars form inside some of the earliest and smallest galaxies.

Instead of simply prescribing the effects of the first stars, the simulations can follow much of the physical environment in which those stars emerge.

The calculations use the RAMSES-RTZ adaptive-mesh-refinement code, allowing computational resolution to be concentrated where the underlying physical processes demand it.

That is an important HPC strategy.

A uniform grid capable of providing parsec-scale resolution across an entire cosmological volume would be computationally prohibitive. Adaptive mesh refinement instead allows the simulation to devote increasingly fine computational resources to the dense, dynamically important regions where stars and galaxies are forming.

The result is a multiscale problem: cosmological structure must be followed over enormous distances while the simulation simultaneously resolves processes occurring on scales measured in parsecs.

More than hydrodynamics

The computational difficulty becomes even clearer when the physical model is examined.

MEGATRON simultaneously incorporates radiative transfer and non-equilibrium chemistry involving more than 80 primordial species, molecules and metal ions.

That matters because radiation from the first stars changes the chemical and thermal state of surrounding gas.

The chemistry then affects cooling.

Cooling changes gas collapse.

Gas collapse changes star formation.

Stars generate radiation and mechanical feedback.

Supernovae distribute newly created elements through the surrounding medium.

Those metals subsequently influence later generations of stars.

In other words, the simulation is not executing one isolated physical calculation.

It is solving a tightly coupled chain of processes in which the output of one physical system becomes the input to another.

For HPC systems, that is precisely the kind of workload that makes scientific simulation difficult: the computational challenge is not simply the number of particles or grid cells, but the interaction of many physical models operating simultaneously across dramatically different spatial and temporal scales.

Four universes, one experiment

MEGATRON is actually a simulation suite rather than a single calculation.

The researchers use the same initial conditions for four simulations, varying the Population II star-formation and feedback models while keeping their Population III assumptions fixed.

The four models include different treatments of star formation and feedback, allowing the researchers to determine which features of the resulting dwarf-galaxy population remain robust when the uncertain physics of later stellar generations is changed.

That makes the project particularly interesting from an HPC perspective.

The supercomputer is being used not merely to produce one enormous dataset, but to perform a computational experiment across multiple physical models.

The researchers can then compare the resulting galaxy populations and determine which observable signatures arise repeatedly.

Across the simulations, the calculation produces a population of more than 500 simulated dwarf galaxies.

That statistical population is essential.

A single simulated galaxy could produce an interesting result. Hundreds of simulated galaxies allow researchers to ask whether a phenomenon is a systematic consequence of the underlying physics.

The iron mystery

The scientific target is an unusual feature observed in ultra-faint dwarf galaxies around the Milky Way.

At very low stellar masses, roughly below 10⁵ solar masses, observed dwarf galaxies show an approximately flat iron-metallicity relation centered around:

⟨[Fe/H]⟩ ≈ −2.5

The origin of that plateau has been debated.

Several possibilities have been proposed, including variations in Population II star formation, Population III physics, external enrichment, and the efficiency with which supernova products escape dwarf galaxies.

MEGATRON provides a computational route to test those possibilities.

The simulations naturally produce an iron-metallicity plateau at low stellar masses, broadly consistent with observations.

The researchers attribute the effect primarily to the chemistry created by massive Population III stars.

And that is where the computational experiment becomes particularly interesting.

One explosion can change an entire galaxy

The simulations indicate that most of the faint dwarf galaxies undergo approximately one Population III explosion before transitioning to Population II star formation.

The massive first-generation star can explode as a pair-instability supernova, producing substantial quantities of heavy elements.

But the supernova does not simply blow all of those elements into intergalactic space.

In sufficiently massive early halos, the gravitational potential can retain the enriched material.

The next generation of stars then forms from gas carrying the chemical signature of the first explosion.

MEGATRON finds that Population III stars tend to form in halos with masses of at least approximately 5 × 10⁶ solar masses.

At those masses, the halos can retain metals produced by high-energy pair-instability supernovae.

The simulations show that diluting the iron produced by a high-mass Population III pair-instability supernova into the gas of a roughly 10⁷-solar-mass halo naturally produces an iron abundance near the observed plateau.

The computer simulation therefore provides a physical connection between an event that occurred more than 13 billion years ago and chemical measurements made in tiny galaxies surrounding the Milky Way today.

Radiation changes the calculation

One of the most important aspects of MEGATRON is that radiation is not treated as an afterthought.

The simulation follows the buildup of radiation from early stars and its interaction with the surrounding gas.

The resulting radiation background can influence where Population III stars form and how subsequent star formation proceeds.

The four simulations consequently produce different reionization histories.

The researchers find that these differences are driven substantially by feedback-modulated escape fractions, the fraction of ionizing radiation that escapes dense star-forming environments and reaches the surrounding intergalactic medium.

This is another reason why a simple galaxy-formation model would not be sufficient for the scientific question.

The radiation field affects the chemistry.

The chemistry affects the gas.

The gas affects star formation.

Star formation determines the radiation field.

The simulation must therefore follow the feedback loop rather than calculate each component independently.

From the early universe to today’s galaxies

There is another computational trick behind the connection to present-day observations.

The researchers also evolve a matching dark-matter-only realization to z = 0.

They then use halo matching and particle-tagging techniques to follow the dynamical fate of simulated dwarf galaxies.

This allows the researchers to ask whether the tiny galaxies produced in the high redshift simulation could plausibly survive as recognizable structures in the present-day Milky Way environment.

The approach is important because the simulations themselves stop at approximately z ≈ 8.

The researchers therefore distinguish between what is directly simulated and what is inferred from subsequent dynamical modeling.

That distinction is important scientifically, and computationally.

The result is not simply a 13-billion-year movie of one galaxy.

It is a combination of high-resolution early-universe radiation-hydrodynamics and additional modeling used to connect those early structures to their potential descendants.

The supercomputer becomes a time machine

The most compelling aspect of MEGATRON may therefore be less about producing another simulated galaxy and more about creating a computational bridge between epochs that cannot be observed directly.

Astronomers cannot watch a Population III star explode inside a primordial dwarf galaxy.

Those events happened billions of years ago.

Instead, researchers can search today’s ultra-faint dwarf galaxies for the chemical fingerprints those ancient explosions may have left behind.

MEGATRON allows that hypothesis to be tested computationally.

The simulation starts with the physics of the early universe, follows the formation of the first stars, tracks radiation and chemical enrichment, generates later generations of stars, and then connects those early systems to dwarf galaxies that could survive around a Milky Way-like host.

The supercomputer effectively becomes a laboratory for experiments that nature performed only once.

Why this matters for HPC

MEGATRON illustrates a broader transition taking place in computational astrophysics.

The next generation of scientific simulations is increasingly defined not simply by larger particle counts or bigger grids, but by more complete physical models operating simultaneously at higher resolution.

Here, the challenge is the combination.

A cosmological calculation must cover a large enough region to capture the formation of a Milky Way like environment.

Within that volume, adaptive mesh refinement must resolve regions down to parsec scales.

Within those regions, the calculation must track gas dynamics, gravity, radiation, chemistry, stellar formation, and feedback.

And those calculations must be repeated under different physical assumptions to determine which predictions are robust.

That is precisely where HPC becomes a scientific instrument rather than merely a faster calculator.

The computer is enabling researchers to perform controlled experiments on physical conditions that cannot be recreated in a laboratory and cannot be directly observed in real time.

The beginning of computational archaeology

The researchers characterize MEGATRON as a vital framework for bridging high redshift simulations with the field of Galactic archaeology. This classification is highly appropriate, as the simulations suggest that specific chemical signatures observed in contemporary dwarf galaxies may function as fossilized remnants of the universe’s first stars. Furthermore, this research underscores the necessity of increasingly sophisticated supercomputing capabilities as astronomical data acquisition becomes more granular. 

Future observations from the Vera C. Rubin Observatory and next-generation extremely large telescopes will significantly expand the census of ultra-faint dwarf galaxies, providing critical datasets for validating models like MEGATRON. To maintain this scientific trajectory, computational models must evolve in tandem with observational precision. While the researchers acknowledge limitations, such as uncertainties regarding primordial star formation, nucleosynthetic yields, and the representativeness of the simulated environment, these factors do not diminish the value of the approach; rather, they define it. Each subsequent simulation serves as a refined computational experiment, narrowing the scope of plausible physical histories that could have produced the universe as we observe it today.

MEGATRON demonstrates what happens when supercomputing is pushed deep into the physics of the early universe: billions of years of cosmic history can be transformed into a numerical experiment, and the chemical fingerprints of the first stars can be tested against galaxies that still exist today.

The first stars are long gone.

Their computational fingerprints are just beginning to come into focus.