Nvidia introduced two new Jetson Thor modules today, the T3000 and T2000, aimed at pushing foundation-model-class AI compute into humanoid robots, autonomous machines, and visual AI systems at the edge. The Blackwell-based Jetson T3000 delivers 865 FP4 teraflops in a package roughly half the size and power of the existing T5000, while the T2000 hits 400 FP4 teraflops for lower-cost edge deployments. Both modules ship in Q1 2027, with emulation on the existing Jetson AGX Thor developer kit starting later this month via JetPack 7.2.1.
The T3000 pairs a Blackwell GPU with an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth, and 25 GbE networking. Nvidia says the module hits similar inference performance to the T5000 on multimodal workloads — large language models, vision-language models, vision-language-action models, and world foundation models — despite the smaller footprint. That parity matters given current memory pricing, where every gigabyte trimmed from a bill of materials shows up in unit economics.
A parallel IGX T3000 variant delivers the same performance with integrated functional safety and runs Nvidia's Halos for Robotics stack, targeting deployments where robots operate alongside humans on factory floors and warehouses. With the T2000 anchoring the lower end at 16GB of memory, Nvidia's edge AI lineup now spans 70 TOPS to 2,000 teraflops — a range wide enough to cover everything from smart cameras to full humanoid brains on a single software stack.
Key facts
- 01The Jetson T3000 delivers 865 FP4 teraflops of AI compute at roughly half the size and power of the T5000.
- 02The T2000 offers 400 FP4 teraflops and 16GB of memory, extending Thor into lower-cost edge AI systems.
- 03Nvidia's Jetson lineup now spans 70 TOPS to 2,000 teraflops across the edge AI portfolio.
- 04Cosmos 3 Edge is a 4-billion-parameter world foundation model that can be post-trained for a specific robot in about a day.
- 05Both T3000 and T2000 modules are scheduled to ship in Q1 2027, with T3000 emulation available later this month via JetPack 7.2.1.
Partner adoption is the tell. Nvidia named 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot as companies building on the Jetson Thor platform. That is a spread across humanoids, industrial arms, warehouse logistics, and factory automation — the four categories most likely to see real commercial deployment volume over the next 24 months.
Alongside the hardware, Nvidia released Jetson agent skills, an automation layer that handles memory optimization, system configuration, and deployment tuning across the Jetson portfolio, including older Jetson Orin modules. The company cited concrete savings from early users: UBTech, Agile Robots, and Connect Tech cut memory usage by up to 15GB, letting them move from the Jetson AGX Orin 64GB to the 32GB module. Retail vision company SandStar shaved 4GB, moving deployments from the Jetson Orin NX 16GB down to the 8GB SKU.
Companion-robot maker GROOVE X, which builds the LOVOT robot, used Jetson's heterogeneous accelerators to redistribute workloads onto lower-memory hardware. NoTraffic, which runs AI on intelligent traffic infrastructure, cut memory usage by 30% on the older Jetson TX2 NX, freeing headroom to add capabilities without a hardware refresh. The pattern is consistent: the agent skills let developers drop one memory tier within a product family, which directly reduces cost per unit at scale.
The software story extends to models. Nvidia expanded its Cosmos 3 world foundation model family with Cosmos 3 Edge, a 4-billion-parameter model built to run on Thor hardware. Cosmos 3 Edge handles on-device perception, real-time reasoning, and action prediction for embodied systems. Nvidia says developers can post-train the model for a specific robot embodiment and sensor suite in about a day using the open Cosmos framework — a tight enough loop to make per-customer or per-fleet fine-tuning practical.
The developer path is deliberately continuous. Because the T3000 and T2000 share chip architecture and software with the existing Jetson AGX Thor kit, engineers can begin work today in emulation mode and drop into production silicon when the modules ship in Q1 2027. Nvidia's physical AI stack — Isaac for simulation and perception, plus open models including Nemotron, Cosmos 3, and Isaac GR00T — carries across the transition. Ecosystem partners including ADLINK, Advantech, AAEON, Aetina, Connect Tech, and Seeed Studio are already shipping Thor-based carrier boards and systems.
The gap between announcement and availability is the friction. Q1 2027 is roughly two quarters out, and the humanoid and industrial-robotics categories Nvidia is targeting are moving fast enough that competitors — Qualcomm on the automotive side, Ambarella and Hailo on lower-power vision, and custom silicon efforts inside the largest robotics companies — will have time to counter-position. Emulation mode helps developers start, but hardware slips in this category are common, and the second-tier partner ecosystem depends on shipping dates holding.
Nvidia is treating robotics the way it treated data-center AI five years ago: build the compute platform, seed the frameworks, cultivate the ecosystem, and wait for the volume category to arrive. The T3000's cost-efficiency positioning — same inference throughput as the T5000, meaningfully less memory and power — is the more commercially interesting half of today's announcement, because humanoids and industrial robots ship in units of thousands, not millions of GPUs. If Cosmos 3 Edge and the Jetson agent skills genuinely compress deployment timelines from weeks to days, the economics of building a robot around a Jetson module start to look a lot more like building a phone around a Snapdragon, and that is the shift Nvidia needs the market to make.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
