Jetson T3000 Brings Blackwell Edge AI to Small Robots

The Jetson T3000 is NVIDIA's answer to a specific complaint: Jetson Thor had the compute, but the module was too big and too thirsty for the robots most people are actually building. NVIDIA has expanded the Jetson Thor line downward with two mainstream modules at roughly half the physical size and power of the existing T4000 and T5000, and the specs on the larger of the two are genuinely close to its bigger sibling.

• Jetson T3000: 1536-core Blackwell GPU, 8-core Arm Neoverse-V3AE CPU, 32GB LPDDR5X at 273GB/s, 25GbE, up to 865 FP4 TFLOPS (sparse)
• Jetson T2000: 1024-core Blackwell GPU, 16GB LPDDR5X at 137GB/s, 400 FP4 TFLOPS
• Half the size and power of the existing T4000 and T5000 modules
• Ships Q1 2027; T3000 emulation support arrives in JetPack 7.2.1 in late July 2026

What Do the T3000 Specs Actually Buy You?

Start with memory bandwidth, because on edge inference that is usually the binding constraint rather than raw FLOPS. The T3000's 32GB of LPDDR5X at 273GB/s is the number to watch — it determines how large a model you can hold resident and how fast you can stream weights through the GPU. Paired with a 1536-core Blackwell GPU and an 8-core Neoverse-V3AE CPU, NVIDIA claims near-T5000 inference performance on multimodal workloads: LLMs, vision-language models, vision-language-action models, and world foundation models.

The 865 FP4 TFLOPS figure is a sparse number, which is the standard way these get quoted — read it as a ceiling, not a floor. Onboard 25GbE networking is a meaningful inclusion for multi-node robotics rigs where sensor data has to move off-module fast.

The T2000 steps down to a 1024-core GPU, 16GB of LPDDR5X at 137GB/s, and 400 FP4 TFLOPS. Roughly half the bandwidth, roughly half the throughput — a clean tier below rather than a crippled part.

Where These Sit in the Edge AI Ladder

This widens NVIDIA's edge range considerably, from the ~70 TOPS parts at the bottom up to the 2,000 TFLOPS flagship. The gap the T3000 fills is the one that has frustrated builders: plenty of projects need more than an entry-level module can deliver but cannot accommodate a T5000's size and power envelope. Readers working through our mini PC and edge coverage — or weighing local inference options in our mini PC buyer's guide for local LLMs — will recognize that as the same tradeoff, one form factor down.

CNX Software estimates T3000 power draw in the 20-65W range, and flags it as an estimate rather than an official figure. Treat it that way.

What NVIDIA Hasn't Said

No pricing has been announced for either module. Availability is Q1 2027, with T3000 emulation support landing in JetPack 7.2.1 in late July 2026 — so developers can start porting well before silicon arrives, which is the sensible part of this rollout.

A few secondary specs come from CNX Software alone rather than NVIDIA's own materials: the T2000's 6-core Neoverse-V3AE CPU, its networking configuration, and the 1024-core GPU count. NVIDIA's blog is less specific on the T2000 than on the T3000, which is normal for a lower-tier part this far from launch. Expect the details to firm up before Q1.

Sources: NVIDIA Blog — July 15, 2026; CNX Software — July 16, 2026.