DGX Spark vs Jetson AGX Thor
Short answer: on paper the two are closer than they look. The Jetson AGX Thor developer kit has the same 128 GB of LPDDR5X at 273 GB/s as a DGX Spark, and at $5,499 it costs $1,451 less. But Thor is built to go inside a robot, and in the third-party LLM tests I can find it runs the same model slower than a Spark. I run GPUwerk, which rents DGX Sparks, so read this with that in mind.
Specs side by side
| Spec | DGX Spark | Jetson AGX Thor developer kit |
|---|---|---|
| Chip | GB10: Blackwell GPU and 20-core Arm CPU (10× X925, 10× A725) | Jetson T5000: Blackwell GPU with 2,560 cores and 96 Tensor Cores, 14-core Arm Neoverse-V3AE CPU |
| Memory | 128 GB LPDDR5x, unified | 128 GB LPDDR5X, 256-bit |
| Memory bandwidth | 273 GB/s | 273 GB/s |
| Peak AI compute (NVIDIA, sparse FP4) | 1 PFLOP | 2,070 TFLOPS |
| Power | 240 W maximum, whole system | 40 to 130 W, module, configurable |
| Networking | ConnectX-7, 200 Gb/s, two QSFP ports | One QSFP28 (4 × 25GbE) and one 5GbE port |
| Operating system | DGX OS (Ubuntu 24.04) | JetPack 7 (Ubuntu 24.04, Linux 6.8) |
| Robotics I/O | None | CAN, camera input through the Holoscan Sensor Bridge |
| US list price | $6,950 (since 2 October 2026) | $5,499 ($3,499 at launch in August 2025) |
Thor figures from NVIDIA's Jetson Thor page, the developer kit and module datasheets, and the Jetson price list, read on 9 November 2026; NVIDIA's price list carries no effective date. The smaller Jetson T4000 module has 64 GB at the same 273 GB/s. Spark figures are on GPUwerk's hardware page.
What that means for LLMs
On paper Thor has more AI compute, 2,070 sparse FP4 TFLOPS against the Spark's 1 PFLOP. That did not show up in third-party llama.cpp tests. Running the same build and model files on both machines, JetsonHacks measured prompt processing on gpt-oss-120b at 938 tok/s on Thor and 1,689 on a Spark, and generation at 41.82 against 52.87 tok/s. A separate llama.cpp community run on Thor got 42.00 tok/s on the same model, which agrees. I have not benchmarked Thor myself; GPUwerk's own Spark figures are on the benchmarks page.
With the same memory size and bandwidth, both hold the same models: gpt-oss-120b, a 70B model at 4-bit, or several smaller models at once. The difference is how fast they run, and what else is in the box.
Price
Thor is no longer the more expensive box. NVIDIA lists the developer kit at $5,499, up from $3,499 at launch, against $6,950 for a Spark since 2 October 2026. If an LLM is all you need to run, the cheaper Thor still gives you less speed per request in the tests I can find, plus a robotics carrier board, CAN headers and a camera bridge you will not use.
Two different jobs
The developer kit is a standalone box with HDMI and DisplayPort that runs Ubuntu, so it can sit on a desk. What sets it apart is the I/O around the chip: CAN for motor controllers, a sensor bridge for cameras, 25GbE for lidar and other sensors, and a module that drops into a robot or vehicle at 40 to 130 W. NVIDIA builds its robotics tools, Isaac and Holoscan, around it.
The Spark is built for LLM development at a desk: DGX OS with NVIDIA's AI software stack, and two ConnectX-7 ports at 200 Gb/s for linking two units, which NVIDIA documents for models too large for one. Thor's one QSFP28 port gives four 25GbE links, 100 Gb/s in total, and I know of no NVIDIA-documented two-unit LLM setup for it.
If your workload is LLM development or serving
The Spark is the box to evaluate, and it is the one GPUwerk has measured directly: single-user and concurrency tok/s figures across six models are on the benchmarks page, and the arithmetic for predicting a new model's speed before downloading it is there too. You can rent a dedicated Spark from $0.79/h, billed by the minute, to test your model; if you want to own one without running it yourself, GPUwerk's GB10 store sells machines from eight makers, and GPUwerk can host yours and rent it out for you.
If your workload is robotics or edge inference
That is Thor's design center, and GPUwerk does not sell or rent it. NVIDIA's Jetson documentation and developer resources are the right starting point. If you are unsure which category your project falls into, a first engagement can help you work out whether an LLM-serving box like the Spark is even the right category of hardware for what you're building.
FAQ
Is the Jetson AGX Thor a competitor to the DGX Spark?
Partly. The Thor developer kit has the same 128 GB of LPDDR5X at 273 GB/s as a Spark and costs less, $5,499 against $6,950 (US list prices as of October 2026). It is built for robots, with CAN, camera and 25GbE I/O, and in third-party llama.cpp tests it ran gpt-oss-120b slower than a Spark. For LLM work the Spark is the better fit; for a robot, Thor is.
Can the Jetson AGX Thor run vLLM or llama.cpp?
Yes. JetPack 7 runs on Ubuntu 24.04, and NVIDIA lists vLLM, SGLang, TensorRT-LLM and llama.cpp as supported. In a March 2026 NVIDIA forum post, one user reported 31 tok/s single-stream on gpt-oss-120b under vLLM. NVIDIA's own October 2025 figure for Llama 3.3 70B at one request is 6.27 tok/s, or 16.19 with EAGLE-3 speculative decoding. None of these numbers are mine.
Is Thor faster than a DGX Spark for LLMs?
Not in the tests I can find. JetsonHacks ran the same llama.cpp build and model files on both in October 2025: gpt-oss-120b generated 41.82 tok/s on Thor against 52.87 on a Spark, and processed prompts at 938 against 1,689 tok/s. Thor has more FP4 compute on paper, but that did not show up in these results.
Can I cluster two Thors like two Sparks?
Thor has one QSFP28 port split into four 25GbE links, 100 Gb/s in total. The Spark has ConnectX-7 at 200 Gb/s, and NVIDIA documents two-unit Spark clusters for larger models. I know of no NVIDIA-documented two-unit LLM setup for Thor.
Should I buy a Jetson AGX Thor instead of a DGX Spark for LLM serving?
Probably not. Thor is cheaper, but in the tests I can find it is slower on the same model, and you pay for robotics I/O you will not use. If LLM development or serving is the job, test your model on a Spark first; GPUwerk rents one by the minute.