코인톡

NVIDIA DSX System: IREN’s Co-Development Role and Commercial Potential (I)

NVIDIA DSX (Data Center/Factory System eXperience) is a major platform officially launched by NVIDIA in early June 2026. It aims to provide infrastructure builders with a complete “AI factory hands-on playbook” from chip to power grid, and from design to operations.

The core objective is to achieve “Lowest Token Cost” through deep software-hardware co-design (Co-design), meaning maximizing intelligent output per megawatt under a fixed power budget.

1. Overview of the DSX system: content and components

DSX is an integrated framework that spans the entire technology stack, bringing together computing, networking, storage, software, power supply, cooling, and partner technologies.

Core components:

• DSX Reference Design: A validated AI factory architecture blueprint. It covers everything from rack-level compute clusters to building physical structure, power distribution, thermal design (such as 45°C liquid cooling), and civil engineering recommendations.

• DSX Sim (Simulation): A high-fidelity digital twin platform based on NVIDIA Omniverse. It allows users to simulate the entire factory lifecycle, performance validation, and bottleneck analysis before any physical assets are deployed.

• DSX OS: An open-source, modular infrastructure operating system designed for large-scale AI factory operations. It provides lifecycle management, intelligent scheduling, runtime consistency, automated health inspection, and resilience management.

• DSX MaxLPS (Maximum Logic Power System): A power optimization technology suite. It combines liquid cooling with rack-level energy efficiency management, dynamically adjusting GPU efficiency points to enable up to 40% additional GPU deployment without affecting performance.

• DSX Flex: A grid interaction layer. It connects the AI factory to external power grids and dynamically adjusts workloads and energy scheduling based on electricity prices and demand response signals.

• DSX Exchange: A communication hub between IT and OT (Operational Technology). It aggregates physical factory sensor data (temperature, power anomalies, etc.) via MQTT protocol for coordinated software scheduling.

• DSX Air: A cloud simulation environment used to simulate computing, networking (Spectrum-X), storage, and orchestration software, ensuring configuration and “Shift-Left” development before equipment arrives.

2. Standard setting and execution logic

DSX follows a Full-Stack Co-design principle:

• Standard setter: Through DSX Reference Design, NVIDIA is effectively defining the industry standard for modern AI factories.

• Execution logic:

– Simulation-first (Sim-first): Before physical construction, a full digital twin validation is completed in Omniverse.
– Programmable infrastructure: Infrastructure (power, cooling) is no longer static but dynamically adjustable through DSX MaxLPS based on GPU load demand.
– Automated operations: The traditional “fault alert” model shifts to DSX OS-based automated proactive remediation.
The above is a general introduction to the DSX system. This is a complex framework that takes considerable time to understand. Unlike the GPU era, where performance iteration cycles were easy to follow, the launch of DSX marks NVIDIA’s transition into system-level control and operations.
As Jensen said at the GTC conference in Taipei in early June:

“Previously NVIDIA was a GPU company, but now we are evolving into a systems company. NVIDIA has truly begun its transformation into an AI infrastructure company that helps customers build entire AI factories, not just sell servers.”

At the same event, Jensen also explained how AI factories are built, saying: “They must first be validated in a digital twin system before the first rack is deployed.”

On the same day, $IREN and BE Networks announced a DSX Air digital twin collaboration based on NVIDIA DSX Air architecture, building a large-scale liquid-cooled AI factory at the Childress site capable of hosting approximately 50,000 Blackwell GPUs, targeting next-generation high-density AI infrastructure.

I checked the roles of IREN, BE Networks, and NVIDIA in the design, testing, and construction of AI factories.

NVIDIA is the system designer; IREN is the engineering execution partner; BE Networks is the bridge between digital simulation and real-world engineering.

NVIDIA’s design is first simulated in DSX Sim, building a virtual AI factory. However, to achieve real value, the virtual factory must closely match reality.

This consistency is not only reflected in data center power systems, cooling systems, network topology, and rack layout, but also in GPU deployment sequencing, real-time sensor data, equipment failure modes, and daily operations.

In this process, BE Networks plays a key role. It digitizes, structures, and standardizes all physical infrastructure data in the data center, and synchronizes it in real time with Omniverse USD digital scenes (USD = the “world file” of digital twins, like an entire map in a game, but at industrial and physical-grade accuracy).

In other words, BE Networks acts as the bridge between IREN’s physical infrastructure and NVIDIA’s Omniverse digital twin platform, enabling the virtual world to accurately reflect real-world operations.

In one sentence:

NVIDIA builds the virtual AI factory and completes simulation via DSX Sim;
BE Networks converts simulation outputs into executable engineering plans;
IREN builds and operates the physical DSX system flagship AI factory.

Clearly, cooperation among the three parties has already entered the engineering execution stage.

Jensen’s recent trips to Taiwan and South Korea were aimed at supply chain coordination, ecosystem alignment, and market mobilization for large-scale deployment of the NVIDIA AI Factory system. He no longer positions NVIDIA as a pure chip company, but as an AI infrastructure company. DSX, DSX Air, Omniverse, Spectrum-X, NVLink, liquid cooling systems, and digital twin technologies all belong to this system.
What $IREN is doing is making DSX-based flagship AI factories become reality. A new era of AI computing infrastructure is unfolding.
I asked several models to estimate how long NVIDIA and IREN had been planning and collaborating before entering physical construction.

The estimates are broadly consistent: for DSX systems, from planning to execution likely takes 12–24 months, with the most likely range being 15–18 months. This aligns closely with the timing when Jensen first introduced the AI factory concept, and also with the timing of IREN’s Sweetwater site going quiet.

In Q4 2024, IREN engaged with multiple hyperscalers regarding large-scale data center partnerships and even hired Morgan Stanley as an advisor. Shortly after, the matter ended without result.

It is now highly likely that the use of the Sweetwater site had already been determined by late 2024. The design and planning of a DSX AI factory requires a fixed site, otherwise system design cannot proceed. It also requires a partner like IREN, with long-term experience in data center engineering and high-performance computing operations.

Many parts of the DSX system require close coordination with experienced engineering partners.

In data center construction, IREN acquired a Canadian data center startup before its Nasdaq listing. After engaging in Bitcoin mining, it invested heavily in operational data collection for HPC data center operations. Later, John Gross, a top expert in liquid cooling, joined IREN. He is Vice Chair of a key committee in ASHRAE (American Society of Heating, Refrigerating and Air-Conditioning Engineers). This ensures IREN can provide critical input on AI factory design and construction.

More importantly, IREN’s vertically integrated model allows it to provide the fastest real-world operational feedback for DSX design and planning. IREN also has multiple expanded sites in Canada whose intended use has not been disclosed, possibly related to specific application testing.

Over the past year, IREN has remained relatively low-profile, and even some execution consistency seen by investors has shown deviations. This may be due to coordination complexity in operating a larger system. This is speculation, but it is clear that IREN and NVIDIA did not begin collaboration only after last month’s DSX AI factory announcement.

Rather, collaboration likely began at the earliest design stage of the system. IREN’s Sweetwater site began foundational construction early. Once concrete foundations are poured, substation layout is already fixed. IREN must have understood NVIDIA’s DSX requirements for power density and liquid cooling interfaces in advance. NVIDIA’s DSX design must also have engaged early with real-world engineering operators.

BE Networks is an NVIDIA-recognized DSX execution partner. IREN’s cooperation with BE Networks marks the transition of the IREN–NVIDIA–DSX flagship AI factory from virtual design to real-world construction, with the digital twin system already mature enough to support physical deployment.

IREN is a co-builder of NVIDIA’s DSX system and will be responsible for physical construction and application, becoming both the builder and an implicit standard participant of the flagship AI factory.

The AI infrastructure industry may therefore be reset. What will its commercial prospects look like? I end with one of Jensen’s statements, which is also the opening for the next deep dive:

“AI factories are moving toward 1 gigawatt per site, with capital costs of $50–60 billion, soon reaching $80–100 billion per gigawatt. These facilities must start successfully in one go and must operate immediately—because any delay is extremely expensive idle capital.”

Jensen also provided an output estimate formula for such AI factories. Goldman Sachs and Morgan Stanley estimate that such facilities could generate $100–150 billion in annual output.

The commercial potential of AI factories is still under study and review. SWEETWATER is currently the largest single-grid renewable energy site in the U.S., and as a DSX flagship site, it may become a landmark symbol of the AI computing industry’s transition to a new phase. I will continue further deep analysis on this topic.