NVIDIA Expands Jetson Thor With T3000 and T2000 Modules

NVIDIA introduced Jetson Thor T3000 and T2000 modules as lower-power additions to its Blackwell-based platform for robotics and edge AI. The official announcement positions T3000 for humanoid robots, industrial systems and other physical AI workloads, while T2000 is aimed at visual AI agents, autonomous mobile robots and compact intelligent machines. Separately authored Wccftech coverage reports the same product introduction, specifications and planned rollout. The announcement broadens the Thor family below the existing flagship configurations rather than replacing them.

The T3000 combines a Blackwell GPU, an eight-core Arm Neoverse CPU, 32GB of LPDDR5X memory and 273GB/s of memory bandwidth. NVIDIA lists 865 FP4 teraflops of AI compute, 25 GbE connectivity and a 70-watt operating target. The company says it can deliver inference performance similar to T5000 for multimodal workloads such as large language, vision language, vision language action and world foundation models. The T2000 is the smaller entry point, with 400 FP4 teraflops, 16GB of memory and a 40-watt target for edge deployments.

The launch fills the space between Jetson Orin systems and the higher-end Jetson Thor lineup. That matters because many physical AI products need local perception and reasoning but operate under tight memory, thermal and power limits. NVIDIA also introduced Jetson agent skills intended to automate memory optimization and system configuration. The company presents those tools as a way to fit workloads onto lower-memory modules, but the retrieved evidence does not provide independent benchmark results for the new hardware or measured production savings.

Teams should treat the announcement as an architecture and capacity-planning option, not a proven deployment result. A sensible evaluation should reproduce the target sensor mix, model stack, latency budget and sustained power envelope on the existing Jetson AGX Thor development path before committing to a future module. Developers can begin with T3000 emulation later this month through JetPack 7.2.1; NVIDIA says T2000 emulation will follow. The emulation path can help identify memory pressure and software compatibility, but final hardware validation will still be necessary.

Moving more perception, language and control workloads onto edge devices can reduce dependence on cloud round trips, but it also concentrates models, credentials, sensor data and actuator access on the deployed system. Product teams should isolate service privileges, sign software updates, constrain network exposure and retain a safe recovery path when an agent or model behaves unexpectedly. For robots operating near people, model performance should remain separate from functional-safety controls and deterministic stop mechanisms.

NVIDIA says T3000 emulation will become available later this month with JetPack 7.2.1, while T2000 emulation is planned for a later release. Both modules are scheduled for Q1 2027 availability. Hardware partners named in the official announcement are preparing Thor-based systems, and software partners are working on emulation and migration support. Those commitments establish an ecosystem path, but they do not yet demonstrate shipping volume, final pricing, field reliability or comparative total cost of ownership.

The next useful evidence will be final module datasheets, pricing, carrier-board support, sustained performance tests and production availability. Buyers should also watch how closely emulation predicts actual memory behavior, latency and power draw across multimodal workloads. Independent benchmarks will be especially important for NVIDIA's claim that T3000 can approach T5000 inference performance in a smaller power and memory envelope. Until those results arrive, the new modules are best understood as a credible roadmap expansion with a development path, not a completed deployment outcome.