NVIDIA Jetson T3000 Packs 865 TFLOPS Into a Tiny Module

Jetson Thor Grows Downward

NVIDIA expanded the Jetson Thor line on July 16, 2026 with two mid-range modules, the T3000 and T2000. The direction of travel is the interesting part: rather than pushing the ceiling higher, NVIDIA took the Blackwell edge platform and shrank it — smaller board, lower power, lower price point — for mainstream robotics and edge AI work. For anyone building a self-hosted vision or local LLM box, that matters more than another halo part.

• Jetson T3000: 1536-core Blackwell GPU with 5th-gen Tensor Cores, 8-core Arm Neoverse-V3AE CPU, 32GB LPDDR5X at 273GB/s, 25GbE, 865 FP4 TFLOPS
• Jetson T2000: 1024-core Blackwell GPU, 6-core Neoverse-V3AE, 16GB LPDDR5X at 137GB/s, dual 10GbE, 400 TFLOPS
• Both boards measure roughly 50 x 87 mm — about half the board area of the T4000 and T5000
• Availability is Q1 2027

What Do You Give Up Versus the T5000?

Less than the spec sheet suggests. NVIDIA says the T3000 draws roughly half the power of the T5000 while delivering comparable multimodal inference performance for LLMs and vision-language models. That is the claim to test independently when hardware lands, but the architecture makes it plausible: FP4 throughput on 5th-gen Tensor Cores is where most of the efficiency gain in this generation lives, and inference on quantized models is exactly the workload it was built for.

The memory bandwidth split is the clearer differentiator between the two new parts. 273GB/s on the T3000 versus 137GB/s on the T2000 is a 2x gap, and for local LLM inference bandwidth is usually the binding constraint long before compute is. If you are sizing a module for token generation rather than vision, the T3000's bandwidth is the number to plan around.

Why the 50 x 87 mm Footprint Is the Real Story

Halving board area changes what you can build. A module at this size fits inside a drone chassis, a mobile robot base, a compact camera head, or a fanless industrial enclosure — places a full-size Thor module simply does not go. Combined with the power reduction, it moves flagship-class edge inference from "specialist platform" toward the same design space as the NPU dev boards we surveyed recently.

Networking is generous for the class, too: 25GbE on the T3000 and dual 10GbE on the T2000. On a module aimed at multi-camera robotics and sensor fusion, that is a sensible allocation of pins — these systems tend to be starved for ingest bandwidth, not for ports.

Pairing With On-Device Models

The software side has been converging on this hardware for months. NVIDIA's own Cosmos 3 Edge world model is a 4-billion-parameter model designed to run on Jetson modules and do spatial reasoning locally — precisely the workload the T3000 is sized for. A smaller, cheaper module plus a small world model is a more approachable starting point for robotics teams than either piece was a year ago.

Q1 2027 is a long wait, and pricing has not been announced. But for anyone planning a build in our mini computer coverage territory — homelab inference boxes, robotics platforms, edge vision rigs — these two parts are worth pencilling into the roadmap now.

Sources: CNX Software — July 16, 2026; ServeTheHome — July 16, 2026.