NVIDIA CEO Jensen Huang announced on May 31 at GTC Taipei 2026 the launch of the DSX platform, positioning it as an end-to-end reference design and operations platform for AI factories. It is NVIDIA’s third-largest product line after RTX and DGX. The cost of building a single 1GW-class AI factory has risen from $20 billion to $30 billion to $50 billion to $60 billion.
Confirmed Technical Specifications for DSX’s Four Core Modules
DSX consists of the following four confirmed components, covering the full AI factory lifecycle from planning to operations:
DSX SIM: Based on Omniverse’s high-fidelity digital twin, it completes the layout planning of the entire factory, power and thermal simulation, and network validation before deployment of the first rack
DSX OS: Open-source modular software responsible for lifecycle management, intelligent scheduling, runtime consistency, health automation, flexibility, and multi-tenant operations
DSX MAX LPS: Combines 45°C liquid-cooling technology and per-watt optimization within the rack, enabling up to 40% more GPU deployments under a fixed power budget to solve the widespread 40% power over-provisioning problem in AI factories
DSX Flex: Real-time reads of grid signals, dynamically adjusts factory power usage based on load reductions, demand response, and electricity price events, making the AI factory a flexible energy asset for the power grid
Confirmed Partner Ecosystem
Hardware: Dell Technologies, HPE, Lenovo, Supermicro, as well as Taiwan-based ASUS, Foxconn, GIGABYTE, Pegatron, QCT (Quanta Cloud Technology), Wistron, and Wiwynn have all built systems supporting NVIDIA DSX.
Cloud partner deployments of DSX core components include CoreWeave, Crusoe, Firmus, IREN, Lambda, Nebius, Nscale, and Yotta Data Services. DSX OS ecosystem partners include Red Hat, Mirantis, Rafay, Spectro Cloud, Supermicro, and Vultr, among others.
Confirmed Pilot Projects for DSX Flex Grid Response
DSX Flex has partnered with Emerald AI and Silicon Valley Power to carry out commercial multi-megawatt pilot projects, demonstrating AI factories with grid-response capabilities. These factories can dynamically adjust their power usage according to signals from the power utility while protecting the performance of AI workloads, helping to unlock additional power capacity to support AI development.
FAQ
Which core issue in AI factories does DSX MAX LPS solve?
In his speech, Jensen Huang pointed out that AI factories commonly have a 40% power over-provisioning problem. DSX MAX LPS solves this by dynamically allocating power between racks and using 45°C liquid-cooling technology. Under the same power budget, it can deploy up to 40% more GPUs. At the same time, NVIDIA’s high-temperature liquid-cooling technology eliminates the need for traditional chiller units, reducing water and energy consumption.
What scale does “100GW by the end of this decade” mean?
In a speech at GTC Taipei 2026, Jensen Huang said that by the end of this decade, 100GW of AI factories will be online worldwide, calling it “the largest infrastructure buildout in human history.” The current cost of a single 1GW-class AI factory is $50 billion to $60 billion, and Huang said this figure will soon reach $80 billion to $100 billion.
How does DSX differ from NVIDIA’s earlier DGX positioning?
Jensen Huang said NVIDIA has completed three transitions: from a GPU company to a system company, and then to an AI infrastructure company. DSX is positioned as the third-largest product line after RTX (consumer graphics cards) and DGX (AI supercomputers). It targets end-to-end design, deployment, and operations at the entire AI-factory level, rather than a single compute system.
