Source: NVIDIA Blog — 2026-07-16
Summary
NVIDIA introduced two smaller Jetson Thor modules, the T3000 and T2000, aimed at pulling Blackwell-generation edge AI compute down into mainstream humanoid, robotics, and edge-AI price and power envelopes rather than only high-end developer kits. The T3000 packs 865 FP4 teraflops into roughly half the size and power of the existing flagship T5000 module while NVIDIA claims comparable inference performance on multimodal workloads; the T2000 goes further down-market with 400 FP4 teraflops and 16GB of memory as an entry point for smaller autonomous machines. Both are scheduled to ship to partners in Q1 2027, with pricing still unannounced.
Key Takeaways
- T3000 specs: 865 FP4 teraflops, a Blackwell GPU paired with an 8-core Neoverse Arm CPU, 32GB of LPDDR5X memory at 273GB/s bandwidth, and 25GbE networking — in a form factor about half the size and power draw of the flagship T5000.
- T2000 specs: up to 400 FP4 teraflops, a 1024-core Blackwell GPU, 16GB of LPDDR5 memory at 137GB/s bandwidth, and 10GbE networking — positioned as the entry point for visual AI agents, autonomous mobile robots, and industrial manipulators.
- For reference, the existing flagship T5000 delivers 2,070 FP4 teraflops with a 14-core Arm Neoverse CPU and 128GB of LPDDR5X memory — so T3000 trades roughly 58% of the compute for roughly half the size/power, while T2000 goes further down to about a fifth of T5000's throughput.
- NVIDIA's specific performance claim is that T3000 matches T5000's inference performance on multimodal workloads specifically — LLMs, vision-language models, vision-language-action models, and world foundation models — meaning the compute cut is aimed at workloads with headroom, not a blanket claim across all model types.
- Both modules are scheduled to become available to embedded partners in Q1 2027; pricing hasn't been announced, though the existing T5000 developer kit sells for $3,499 and standalone T5000 modules run $2,999–$1,999 through embedded partners, giving a rough sense of where T3000/T2000 pricing is likely to land.
Reel Script
Hook (~16s, 36 words): NVIDIA just launched a robot brain with a fifth of the compute of its flagship chip — and the pitch is that for most robots, that's plenty. Here's the actual math behind "smaller can be enough."
Core Concept (~70s, 155 words): Jetson Thor is NVIDIA's line of compute modules that go inside robots and edge devices — the chip that lets a robot understand what a camera sees and decide what to do about it, without phoning home to a data center. Until now the flagship was the T5000: powerful, but expensive and power-hungry, which prices it out of a lot of real robots that need to be cheap and battery-friendly. The T3000 and T2000 are NVIDIA cutting that same Blackwell architecture down to size — literally: the T3000 is about half the size and power draw of the T5000. The bet is that most robotics and edge-AI workloads don't actually need the full T5000's headroom, so you can strip compute down significantly and still hit the performance that matters for the model types robots actually run — as long as you're specific about which workloads you're comparing.
Hands-On (~120s, 270 words): Here's the actual spec comparison worth putting on screen as a table. T5000, the existing flagship: 2,070 FP4 teraflops, 14-core Arm Neoverse CPU, 128GB of LPDDR5X memory. T3000, the new mid-tier: 865 FP4 teraflops — about 42% of T5000's raw number — paired with an 8-core CPU and 32GB of memory, in roughly half the size and power envelope. T2000, the new entry-level: 400 FP4 teraflops, a 1024-core Blackwell GPU, and just 16GB of memory. So on paper, T3000 looks like it's giving up more than half its compute. But NVIDIA's specific claim is narrower than "T3000 is slower" — they say it hits similar inference performance to T5000 specifically on multimodal workloads: large language models, vision-language models, vision-language-action models, and world foundation models running on the robot. That's a claim worth being skeptical of on camera, because "similar inference performance" on a curated workload set is doing a lot of work — it likely means these particular model architectures are memory-bandwidth-bound or batch-size-bound in a way that doesn't scale linearly with raw teraflops, not that the chips are secretly equivalent. Both ship to partners in Q1 2027; no pricing yet, but the T5000 dev kit at $3,499 and standalone modules at $2,999/$1,999 give a rough anchor for where these will likely land.
Takeaway (~22s, 48 words): This is NVIDIA trying to make Blackwell-class edge AI affordable enough for actual mass-market robots, not just research labs — worth tracking if you build anything battery-powered. Don't take "matches the flagship" at face value until independent benchmarks land; ask which workloads that claim covers.