NVIDIA DSX
Last reviewed
Jun 2, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 2,354 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 2,354 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA DSX is a platform from NVIDIA for designing, simulating, building and operating large-scale data centers that the company calls AI factories. NVIDIA describes DSX as a "complete playbook" that pulls together reference designs, open-source software libraries, application programming interfaces, NVIDIA's accelerated computing platforms and a broad partner ecosystem into a single codesigned framework, so that operators can model an entire facility before construction and then run it with the kind of reliability that production AI workloads demand.[1][2]
NVIDIA formally introduced the DSX platform as a unified offering at GTC Taipei, held alongside COMPUTEX 2026 in Taiwan, where company founder and chief executive Jensen Huang delivered the keynote on June 1, 2026. The associated press release, titled "NVIDIA DSX Gives Infrastructure Builders the Playbook for AI Factories," carried a Taipei dateline of May 31, 2026.[1][3] DSX rounds out NVIDIA's three-letter naming scheme for its product tiers, sitting alongside RTX for consumer and workstation graphics and DGX for AI systems and supercomputers, with DSX positioned at the level of the data center itself.[4]
| Attribute | Detail |
|---|---|
| Developer | NVIDIA |
| Type | AI factory reference design and infrastructure software platform |
| Announced | GTC Taipei at COMPUTEX 2026; keynote June 1, 2026 (press release dated May 31, 2026)[1] |
| Purpose | Design, simulation, build-out and operation of gigawatt-scale AI data centers ("AI factories")[2] |
| Core software components | DSX OS, DSX MaxLPS, DSX Sim, DSX Flex, DSX Exchange[1][2] |
| Reference design | Vera Rubin DSX AI Factory reference design (built on the NVIDIA Vera Rubin platform)[5] |
| Digital twin | Omniverse DSX Blueprint, built on NVIDIA Omniverse[6] |
| Scale targeted | 100 megawatts to multiple gigawatts per facility[6] |
| Key efficiency claim | Up to 40% more GPUs within the same megawatt budget via DSX MaxLPS[1] |
| Stated availability | Many components offered on a "when-and-if-available" basis[1] |
NVIDIA frames DSX around its broader idea of the "AI factory," a term it uses to recast a data center as an industrial plant whose product is intelligence measured in tokens rather than a general-purpose computing facility. In NVIDIA's framing, a token is both the unit of AI output and the unit of revenue, so the central design goal of an AI factory becomes maximizing useful token inference throughput for every watt of power the site can draw.[1][7] Huang has summarized the economics bluntly, saying that with a gigawatt of power, "throughput per watt is revenues, because every token is profitable, every token is revenues."[7]
The scale NVIDIA is targeting is large. In his GTC Taipei remarks, Huang said the cost of a single gigawatt of AI infrastructure has climbed from $20 billion to $30 billion a few years ago toward $80 billion to $100 billion, and he repeated NVIDIA's expectation that roughly 100 gigawatts of AI factories will come online before the end of the decade.[7] DSX is NVIDIA's attempt to give the companies building that capacity, including cloud providers, neoclouds, enterprises and governments, a repeatable method rather than a one-off engineering project for each site. The company has gone so far as to claim it "is the only company that builds the full AI factory," spanning silicon, systems, networking, software and the facility design itself.[1]
NVIDIA structures DSX as a set of modular software libraries and reference materials that map onto three phases of an AI factory's life: design, deployment and operations.[2] Several pieces were introduced over the preceding GTC events, and the GTC Taipei announcement added two notable open-source elements, DSX MaxLPS and DSX OS.[3] The table below lists the named components NVIDIA has described.
| Component | What NVIDIA says it does |
|---|---|
| DSX Reference Design | Generation-specific, validated AI factory architectures covering compute, networking, storage, hardware cluster design and facilities infrastructure, including power, cooling and controls as well as civil, structural and architectural design.[1] |
| DSX MaxLPS | A suite of technologies to maximize token performance per megawatt within a fixed power budget. It combines 45 degrees Celsius liquid cooling with in-rack techniques so operators can run up to 40% more GPUs at their most energy-efficient operating point with minimal impact on workload performance.[1] |
| DSX OS | Open-source, modular software purpose-built for AI factory operations, providing lifecycle management, scheduling, runtime consistency, health automation, resiliency, multi-tenant operations and platform services.[1][8] |
| DSX Sim | A high-fidelity simulation layer for the AI factory lifecycle, used by NVIDIA, partners and customers to model, validate and optimize infrastructure decisions from planning and design through deployment and operations.[1] |
| DSX Flex | Connects AI factories to power-grid services, adapting workloads to grid signals such as load shedding, demand response and pricing events, and orchestrating renewable and hybrid power across utility, on-site renewables and storage.[1] |
| DSX Exchange | Enables scalable, secure integration of compute, network, energy, power and cooling plant signals between IT, operational technology and operations agents.[1] |
DSX MaxLPS is the component NVIDIA positions as the one that lowers the cost of each token. The name refers to maximizing what NVIDIA calls token performance per megawatt. Because an AI factory is constrained by the power it can secure rather than by floor space, NVIDIA's argument is that the operator who extracts the most performance from a fixed megawatt allocation wins on cost. According to NVIDIA, pairing 45 degrees Celsius liquid cooling with in-rack power management lets a site run as much as 40% more GPUs at their most energy-efficient operating point, with little measurable hit to workload performance.[1]
DSX MaxLPS is the successor concept to ideas NVIDIA discussed at earlier 2026 events under the names DSX Max-Q and DSX Boost. At GTC in March 2026 the company described DSX Max-Q as a library that maximizes computing output and token performance per watt, and in the Omniverse DSX materials it referenced DSX Boost as delivering up to 30% higher GPU throughput within the same power envelope.[5][6] The consolidation of these into DSX MaxLPS at GTC Taipei reflects how quickly the platform's branding has evolved.
DSX OS is the operations layer, announced as open-source and modular software for running multi-tenant AI factories at scale.[8] NVIDIA's technical description lists a set of named subsystems rather than a single monolithic product. These include the NVIDIA Infra Controller (NICo) for API-driven bare-metal lifecycle management and tenant isolation through NVIDIA BlueField DPUs; the NVIDIA AI Cluster Runtime (AICR) for version-locked software recipes that prevent configuration drift; NVSentinel for Kubernetes-native GPU fault detection and automated remediation; NVIDIA Fleet Intelligence for fleet-wide visibility and integrity verification; the KAI Scheduler together with NVIDIA Run:ai for GPU-aware workload placement and fractional allocation; and inference services such as NVIDIA Dynamo, Grove and Cloud Functions. DSX Exchange in this stack is described as an MQTT-based hub connecting IT and operational-technology systems.[8] NVIDIA's stated intent is that ecosystem partners build on these components to deliver AI services rather than rebuild operational tooling from scratch.
DSX Sim is the simulation layer, and a key part of it is DSX Air, a software-as-a-service platform NVIDIA introduced at GTC in March 2026 for logically simulating AI factories. NVIDIA says DSX Air produces high-fidelity digital simulations of its hardware, including GPUs, SuperNICs, DPUs and switches, and connects to partner tools for storage, routing, security and orchestration through open APIs. The company's pitch is that this kind of pre-build modeling can cut the time to a customer's first token from weeks or months down to days or hours.[9]
The first generation-specific blueprint under the platform is the Vera Rubin DSX AI Factory reference design, built around the NVIDIA Vera Rubin computing platform. NVIDIA describes it as a unified architecture that integrates compute, NVIDIA Spectrum-X Ethernet networking, storage, power, cooling and facility controls so partners can design, deploy and scale gigawatt AI factories with maximum token throughput per watt and improved uptime.[5] It is intended to give operators a tested starting point covering everything from the GPU racks to the building's electrical and mechanical systems.
A defining feature of the Vera Rubin generation, and therefore of this reference design, is a shift in power architecture. NVIDIA and more than twenty partners are moving toward 800-volt direct current (VDC) distribution for the gigawatt era, replacing the traditional 415-volt or 480-volt alternating-current three-phase systems used in most data centers. NVIDIA says 800 VDC improves scalability and energy efficiency, reduces copper and other materials, and supports the high power density of its Kyber rack architecture for Rubin-class GPUs.[10] This power work connects DSX to the broader Vera Rubin rack and system roadmap, including the NVL72 rack-scale designs in NVIDIA's MGX open architecture.
The simulation and design half of DSX is anchored by the Omniverse DSX Blueprint, built on NVIDIA Omniverse. NVIDIA first introduced this blueprint at GTC Washington in October 2025 and describes it as an open, comprehensive framework for designing and operating gigawatt-scale AI factories.[6] Using OpenUSD assets and Omniverse libraries, the blueprint lets engineering teams build a physically accurate digital twin of a planned facility, simulate its operation in real time and tune power, cooling, networking and operations before any concrete is poured. NVIDIA has said the blueprint is generally available through build.nvidia.com and is compatible with the Vera Rubin DSX reference design.[5][6] The blueprint targets facilities ranging from about 100 megawatts up to multiple gigawatts and supports both NVIDIA Grace Blackwell and Vera Rubin platforms.[6]
DSX is explicitly an ecosystem play, and NVIDIA's announcements name a long roster of partners across the hardware, cloud, software, facilities and power layers. The table below groups the partners NVIDIA cited as adopting or contributing to the platform.
| Category | Partners named by NVIDIA |
|---|---|
| Cloud and neocloud providers deploying DSX components | CoreWeave, Crusoe, Firmus, IREN, Lambda, Nscale, Nebius, Yotta Data Services[1] |
| System manufacturers building DSX-ready systems | Dell Technologies, HPE, Lenovo, Supermicro, ASUS, Foxconn, GIGABYTE, Pegatron, Quanta Cloud Technology (QCT), Wistron, Wiwynn[1] |
| DSX OS ecosystem adopters | Aible, BeyondAI, Bhashini, DCAI, Mirantis, OpenNebula Systems, Rafay, Red Hat, Sarvam, Simplismart, Spectro Cloud, vCluster, Vultr[1] |
| Omniverse DSX / digital twin software | Cadence, Dassault Systèmes, PTC, Siemens[1][6] |
| Power, grid and cooling | Emerald AI, Silicon Valley Power, Schneider Electric, Vertiv, Eaton, Trane Technologies, GE Vernova, Hitachi, Siemens Energy[1][6] |
| Engineering and AI operations | Jacobs, Bechtel, Phaidra[6] |
NVIDIA has highlighted a few partner outcomes to make the case concrete. It says Phaidra's AI agents, integrated through the platform, can recover roughly 10% more compute by smoothing cooling spikes, and it has pointed to a project by Nscale and Caterpillar that is bringing the Vera Rubin DSX reference design to life on a multi-gigawatt site in West Virginia, which NVIDIA calls one of the largest AI factories in the world.[5][6]
Among the system builders, ASUS used COMPUTEX 2026 to announce that it had adopted the DSX platform for its AI factory offerings, pairing it with rack-scale systems such as its AI POD built on the Vera Rubin NVL72 design. ASUS framed the value as the ability to evaluate power delivery, cooling, networking topology, storage and facility readiness in simulation before physical build-out begins, which it tied to faster time to first token and revenue.[11]
Trade and industry press generally read the DSX announcement as a strategic shift rather than a single new product. Several outlets characterized it as NVIDIA moving up the value chain from selling chips, and then racks, to selling a methodology and software stack for the entire facility, with DSX serving as the data center counterpart to its DGX systems and RTX graphics lines.[4][12] Coverage also placed DSX within NVIDIA's wider gigawatt-scale narrative at GTC Taipei, where the company put the Vera Rubin platform into full production and leaned heavily on the AI factory framing.[7][12]
It is worth noting what DSX is not, as of its announcement. NVIDIA cautioned that many of the products and features it described "remain in various stages and will be offered on a when-and-if-available basis," so component availability and final specifications may change.[1] The platform is also a moving target in naming and scope, having evolved across the October 2025, March 2026 and June 2026 GTC events, with several component names (DSX Boost and DSX Max-Q folding into DSX MaxLPS) consolidating along the way.[5][6]