NVIDIA launches Jetson Thor T3000 and T2000 modules for mainstream robotics and edge AI

NVIDIA introduced two new Jetson Thor compute modules, the T3000 and T2000, extending its Blackwell GPU architecture down into compact, power-efficient systems built for robotics and edge AI. The modules ship alongside new Jetson software memory optimizations and prebuilt agent skills, with hardware scheduled to reach partners and customers in the first quarter of 2027.

What's new

NVIDIA's announcement centers on two tiers of the new module:

• Jetson T3000 combines an NVIDIA Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory and 273GB/s of memory bandwidth, delivering up to 865 FP4 teraflops.
• Jetson T2000 is a lower-power sibling in the same family, rated at 400 FP4 teraflops.
• New Jetson agent skills add memory optimization across the line, aimed at squeezing larger models onto constrained edge hardware.
• Cosmos 3 Edge, a lightweight version of NVIDIA's embodied-AI model family, is built specifically to run on these modules.
• An emulation mode ships this month in JetPack 7.2.1, letting developers start building against Jetson Thor's software stack before the physical modules arrive.
• More than a dozen manufacturing partners — including ADLINK, Advantech, AAEON, Connect Tech, NEXCOM and Seeed Studio — are already building carrier boards and systems around the new modules.

NVIDIA confirmed the launch timing directly: "The Jetson T3000 and T2000 modules are scheduled to become available in Q1 2027."

Context

Jetson is NVIDIA's line of edge compute modules for robots, drones, cameras, and other devices that need to run perception, planning, and control workloads locally rather than round-tripping to the cloud. Earlier Jetson generations were built on NVIDIA's Orin-class silicon; moving the line to Blackwell brings the same GPU architecture underpinning NVIDIA's current data-center lineup into a battery- and thermally-constrained form factor for the first time. NVIDIA has increasingly paired these hardware refreshes with dedicated software — Jetson agent skills, and now a purpose-built Cosmos model variant — rather than shipping silicon alone, treating Jetson as a full platform that customers build on rather than a one-time chip purchase.

Why it matters

Edge robotics has been constrained as much by memory bandwidth and power budget as by raw compute — running vision-language-action models on a mobile robot or drone means balancing all three simultaneously, something data-center GPUs were never designed for. By combining Blackwell-class compute with memory-optimization software and a Cosmos variant sized for edge deployment, NVIDIA is trying to make its top-tier AI stack usable outside the data center rather than simply scaling a smaller chip down. The long runway before general availability — an emulation path shipping now, hardware over a year out, but a dozen-plus manufacturing partners already committed — mirrors the playbook NVIDIA has used with DGX and prior Jetson launches: lock in developer mindshare well ahead of shipping hardware, so Jetson stays the default reference platform for robotics builders rather than ceding ground to Qualcomm, Intel, or in-house silicon efforts from robotics companies themselves.