NVIDIA pitches factory digital twins and Thor robot computers as the toolkit for US reindustrialization, enlisting Foxconn, Siemens, FANUC and four robot developers

NVIDIA used its GTC conference in Washington, D.C. on October 28, 2025 to pitch factory digital twins and robot brains as the toolkit for rebuilding US manufacturing, naming Foxconn, Caterpillar, Toyota, TSMC, Lucid Motors, Belden and Wistron as users of its Omniverse simulation software and Figure, Agility Robotics, Amazon Robotics and Skild AI as builders of what it called a collaborative robot workforce.

The company framed the push around one number. In 2025, NVIDIA said, announced investments to build out US production capacity reached $1.2 trillion, led by electronics makers, drugmakers and chip manufacturers. Founder and chief executive Jensen Huang said AI is turning the world's factories into "intelligent thinking machines". The same day NVIDIA introduced IGX Thor, an industrial and medical edge computer rated at 5,581 FP4 teraflops, with production systems and developer kits expected in December 2025.

The software centrepiece is an expanded version of the "Mega" NVIDIA Omniverse Blueprint, a reference design that now carries libraries for designing and simulating whole factories before steel is cut. NVIDIA said Siemens is the first to build digital twin software supporting it, currently in beta as part of Siemens Xcelerator. FANUC and Foxconn's industrial arm Fii are among the first robot makers to publish 3D digital twins of their robots in the OpenUSD format, so planners can drop machines into a virtual plant. Huang showed Foxconn using the tools to design its 242,287 square foot Houston facility, which will build NVIDIA AI infrastructure systems.

Other manufacturers described narrower uses. Caterpillar is building twins of its factories and supply chains and using NVIDIA NIM microservices to predict maintenance needs and cuOpt to tune its supply chain. Toyota is modelling automation scenarios for its Georgetown, Kentucky plant with idealworks' iw.sim tool. TSMC is using Omniverse for fab design at its Phoenix, Arizona site and the Isaac platform for robotics. Lucid Motors is twinning its factories for planning and robot training, Wistron is validating systems assembled in Fort Worth, Texas, and Belden has deployed Accenture's Physical AI Orchestrator for virtual safety fences and quality inspection.

On the robot side, NVIDIA said Figure and NVIDIA had agreed to collaborate on next-generation humanoids, with Figure using NVIDIA computing to build its Helix vision language action model and the Isaac platform for simulation and training. Agility Robotics refines whole-body control for its Digit humanoid through millions of reinforcement learning scenarios in Isaac Lab, and Digit runs on the Jetson AGX Thor module. Skild AI uses Isaac Lab and NVIDIA's Cosmos world models to train one foundation model across legged, wheeled and humanoid robots. NVIDIA also credited simulation training for helping Amazon's Blue Jay multi-arm manipulator move from concept to production in just over a year.

IGX Thor extends the Blackwell-based Thor family from robots into regulated settings such as hospitals and rail. According to NVIDIA's announcement blog, it combines an integrated and a discrete Blackwell GPU, offers 400 GbE connectivity, delivers eight times the integrated-GPU AI compute of IGX Orin and 2.5 times on the discrete GPU, and comes with 10 years of lifecycle support and NVIDIA's Halos safety stack. It ships as the IGX T5000 module and the IGX T7000 board kit. Early adopters named include Diligent Robotics, EndoQuest Robotics, Hitachi Rail, Joby Aviation and Maven Robotics, with CMR Surgical evaluating it. No price was disclosed.

The launch builds on Jetson AGX Thor, which NVIDIA made generally available on August 25, 2025 with a $3,499 developer kit, 2,070 FP4 teraflops within 130 watts and 128 GB of memory. At that launch NVIDIA listed Figure, Amazon Robotics, Boston Dynamics, Caterpillar and Agility among early adopters and said more than 2 million developers use its robotics stack.

NVIDIA's pitch rests on what it calls a three-computer architecture, one system to train models, one to simulate and test them, and one inside the robot. The approach puts it in competition on several fronts at once. Siemens, Dassault Systemes and PTC sell factory simulation, while robot makers from FANUC to ABB run their own controllers. NVIDIA's answer in Washington was to make partners such as Siemens and FANUC the doorway into its stack rather than rivals to it.

The technology is easier to describe than to deploy. A digital twin is a physics-accurate model of a plant, built in the OpenUSD scene format, where engineers test layouts, conveyor speeds and robot cells before building them. Robot policies are then trained in simulation, where Isaac Lab can run reinforcement learning across millions of scenarios, and transferred to real hardware. NVIDIA argues this shortens development cycles, the claim it attached to Amazon's Blue Jay.

The risks are plain in the release itself. The Siemens integration was still in beta, IGX Thor had no published price, and the collaboration with Figure carried no financial terms. Most manufacturers described pilots and planning tools rather than plants that now run on NVIDIA software. Editor's note. In an update dated February 25, 2026, Amazon said it no longer uses Blue Jay in operations, while the underlying technology continues to support employees across its network.