NVIDIA · Local AI machine
NVIDIA Jetson AGX Thor (128 GB)
NVIDIA machine: NVIDIA Jetson T5000 (Blackwell), 128 GB memory, 273 GB/s memory speed, about 4.2 tok/s on a 70B model (estimate).
81.2Score
- Not ranked yet
Why this score
The score adds up these parts. Each part has a weight. A part counts as the average of its sources, pulled toward 80 when there are few sources.
AI power
50% weightcounts as 82.4Average of 2 sources: 83.6
Experts
30% weightcounts as 80.0No sources found yet. No data yet. This part counts as 80.
Users
20% weightcounts as 80.0No sources found yet. No data yet. This part counts as 80.
No reviews or buyer ratings yet: the score comes from the specs only.
80 is the field average. A part counts as 80 + (average − 80) × N ÷ (N + 1), where N is its number of sources. One source counts half of the gap from 80. No sources count as 80.
Price
€6856.34 at reichelt.at (via Geizhals.at) · Geizhals.at listing, not checked live
Developer kit 945-14070-0085-000, in stock, plus EUR 6.95 shipping. Geizhals lists one other offer at EUR 9,187.82 (EU region).
Prices checked 8 Oct 2026.
Specs
- Chip
- NVIDIA Jetson T5000 (Blackwell)
- CPU
- 14 cores / 14 threads (Arm Neoverse-V3AE)
- Memory
- 128 GB
- GPU memory
- 128 GBUnified 128 GB shared by CPU and GPU on Jetson Linux, per NVIDIA's Jetson Thor page.
- Memory speed
- 273 GB/s
- FP16 TFLOPS
- —
- NPU
- —
- OS
- Linux
- Software
- CUDA
- Ethernet
- 5 Gbit/s
- Thunderbolt / USB4
- —
- QSFP port
- Yes
- Idle power
- —
- Load power
- 130 W
- Noise
- —
- Volume
- 1.6 L
- Weight
- 1.9 kg
- SSD
- 1 TB
- Release
- 2025-08
Measured
- gpt-oss-120b, MXFP4
- 41.8 tok/s writing, 937.8 tok/s prompt · llama.cpp b6767 (source page does not name the backend)
- Llama 3.3 70B
- 4.7 tok/s writing · vLLM, concurrency 1 (NVIDIA; precision not listed)
- Qwen3 30B-A3B
- 61 tok/s writing · vLLM, concurrency 1 (NVIDIA; precision not listed)
- Qwen3 Coder 30B A3B, Q8_0
- 42.7 tok/s writing, 1533.7 tok/s prompt · llama.cpp b6767 (source page does not name the backend)
Notes
A robotics developer kit rather than a PC: 128 GB unified memory at the same 273 GB/s as GB10, so it runs gpt-oss-120b at about 42 tok/s but a 70B dense model at under 5 tok/s. Linux/JetPack only, 5 GbE, no Thunderbolt, and it costs about EUR 6,900 in Austria.
Llama 3.3 70B: NVIDIA's table says 4.7 tok/s at concurrency 1, ServeTheHome's author says in a comment 'low 10s' tok/s (not run through their standard process). NVIDIA's Jan 2026 data sheet text lists 'Gigabit Ethernet' while the Thor page and the 2025 data sheet say 5 GbE RJ45; 5 used. Weight is only from resellers (1.856 vs 2.00 kg). US list price is USD 5,499 now against USD 3,499 at launch; Austrian price is EUR 6,856.
More reading:servethehome.comjetsonhacks.comhothardware.com