Jetson Thor vs DGX Spark: Which to Buy

TL;DR

• Buy DGX Spark ($3,999) if you develop, fine-tune, and run large language models at your desk. It's a desktop AI workstation with the full NVIDIA CUDA/AI stack, 128GB unified memory, and inference for models up to ~200B parameters.
• Buy Jetson AGX Thor (dev kit $3,499) if you build robots or edge devices. It's an embedded module plus dev kit for physical AI — real-time sensor I/O, 40–130W configurable power, meant to ship inside a machine.
• Shared silicon, different jobs: both use NVIDIA Blackwell, both have 128GB LPDDR5X at 273 GB/s. Spark = develop locally. Thor = deploy at the edge.
• DGX Spark was first shown as "Project DIGITS" — same product, people still search the old name.

Jetson Thor vs DGX Spark: the short answer

For jetson thor vs dgx spark, pick by deployment, not by spec sheet. Buy the NVIDIA DGX Spark ($3,999) as a deskside workstation to build, fine-tune, and run large AI models locally. Buy the NVIDIA Jetson AGX Thor (developer kit $3,499) to put AI inside a robot or edge device. Same Blackwell generation, same 128GB memory — different jobs.

They get confused for good reason: both cost roughly $3.5k–$4k, both carry 128GB of unified LPDDR5X memory at 273 GB/s, and both run on NVIDIA's Blackwell architecture. But one sits on your desk and one ships inside a machine. That distinction is the whole decision.

What is the NVIDIA DGX Spark?

DGX Spark is a desktop AI workstation — a "personal AI supercomputer" built around the GB10 Grace Blackwell Superchip. It's made to sit on a desk and let you prototype, fine-tune, and run large models locally without renting cloud GPUs.

Per NVIDIA's DGX Spark product page, the GB10 pairs a 20-core Arm CPU (10 Cortex-X925 + 10 Cortex-A725) with a Blackwell GPU with 5th-gen Tensor Cores. It delivers up to 1 petaFLOP of FP4 AI performance (theoretical, using sparsity), carries 128GB LPDDR5X unified memory at 273 GB/s, and offers up to 4TB of NVMe storage.

NVIDIA rates it for inference of models up to 200 billion parameters and fine-tuning up to 70 billion parameters locally. It runs DGX OS (Ubuntu-based) and the full NVIDIA AI software stack — NeMo, RAPIDS, NIM microservices, PyTorch. It's the local-dev answer to the job you'd otherwise hand to Ollama or llama.cpp on a workstation GPU.

DGX Spark launched at $3,999 and went on sale October 15, 2025, according to Tom's Hardware (the price and date are not on NVIDIA's own product page). It was first previewed as Project DIGITS and renamed DGX Spark at GTC 2025, as HotHardware reported.

What is the NVIDIA Jetson AGX Thor?

Jetson AGX Thor is an embedded edge-AI module and developer kit for robotics and physical AI. It's built to be designed into a product — a humanoid, an autonomous machine, a smart camera — not to sit on a desk.

Per NVIDIA's Jetson Thor launch release, Jetson Thor delivers up to 2,070 FP4 teraflops of AI compute from a Blackwell GPU, pairs it with a 14-core Arm CPU, and carries 128GB of memory in a 130-watt power envelope. NVIDIA cites up to 7.5x higher AI compute and 3.5x greater energy efficiency than the previous-generation Jetson Orin.

The Jetson T5000 module datasheet lists a 2,560-core Blackwell GPU with 96 fifth-gen Tensor Cores, a 14-core Arm Neoverse-V3AE CPU, 128GB LPDDR5X at 273 GB/s, and a configurable 40W–130W power range — the knob that matters for battery- and thermal-limited machines.

The Jetson AGX Thor Developer Kit is available now, starting at $3,499. It runs the Jetson software stack and NVIDIA's robotics frameworks: Isaac, Isaac GR00T (humanoid foundation models), Metropolis (vision AI), and Holoscan (real-time sensor processing).

Jetson Thor vs DGX Spark: spec comparison

Both run NVIDIA Blackwell and both pack 128GB. Here is where they match and where they split.

Two numbers explain both the confusion and the difference. The 273 GB/s bandwidth and 128GB memory are identical — that's why both can run large models locally. But Thor's 2,070 FP4 TFLOPS peak is higher on paper than Spark's 1 PFLOP, because Thor is tuned for high-throughput real-time robotics inference. These are peak theoretical figures (sparsity-assisted), not LLM tokens-per-second — don't read them as a head-to-head LLM benchmark.

Jetson Thor vs DGX Spark: the differences that matter

The silicon generation is shared. The purpose is not. If your real choice is between data-center accelerators like the A100 vs H100, that's a different decision — this comparison is about desk-side and edge Blackwell parts.

• Form and deployment. Spark is a finished desktop appliance. Thor is a module meant to be designed into a product via a carrier board, plus a dev kit to prototype on.
• I/O. Spark leans into developer networking — ConnectX-7 at 200Gbps and dual QSFP to cluster two units or move data fast. Thor leans into real-time sensor and camera I/O for robots.
• Power. Spark draws a steady desktop load (140W chip, 240W supply). Thor is configurable from 40W to 130W to fit power-constrained machines.
• Software. Spark ships DGX OS and the full AI/CUDA stack for model work. Thor ships the Jetson stack and robotics frameworks (Isaac, GR00T, Metropolis, Holoscan).

What this means for you

• If you're an LLM developer or researcher: buy DGX Spark. You get a desk-side box that runs and fine-tunes large models locally on the full CUDA stack, with fast networking to pair two units. It replaces renting cloud GPUs for prototyping.
• If you're a robotics or embedded engineer: buy Jetson AGX Thor. It's the only one of the two you can actually deploy inside a robot or edge device, with configurable low power and real-time sensor handling.
• If you want to run big models locally and nothing else: DGX Spark. If a discrete workstation GPU fits your bench better than an all-in-one appliance, weigh the RTX PRO 6000 vs RTX 5090 instead.
• If you're deciding for a product that ships to customers: Thor is the deployable part; Spark is a developer workstation, not an embedded component.

The honest read: this is rarely a real either/or. If your question is "which do I deploy into a machine," it's Thor. If it's "which do I develop and run models on at my desk," it's Spark.

Frequently Asked Questions

What is the difference between Jetson Thor and DGX Spark?

DGX Spark is a desktop AI workstation for developing, fine-tuning, and running large models locally. Jetson AGX Thor is an embedded module and dev kit for building robots and edge devices. Both use NVIDIA Blackwell and 128GB of memory, but Spark sits on a desk and Thor ships inside a machine.

Should you buy a DGX Spark or Jetson Thor?

Buy DGX Spark ($3,999) if you're an AI/LLM developer who wants to prototype and run large models at your desk. Buy Jetson AGX Thor (dev kit from $3,499) if you're building robotics or edge-AI products that need to deploy the compute inside a physical machine. Match the tool to the job, not the price.

Is the DGX Spark the same as NVIDIA DIGITS?

Yes. DGX Spark was first previewed as "Project DIGITS" and renamed DGX Spark at GTC 2025. Many people still search the old name, but it is the same GB10 Grace Blackwell desktop system — there is no separate DIGITS product shipping today.

Is Jetson Thor faster than DGX Spark?

On NVIDIA's published peak FP4 figures, Jetson Thor lists up to 2,070 TFLOPS versus DGX Spark's up to 1 petaFLOP. But those are peak theoretical numbers tuned for different workloads — Thor for real-time robotics, Spark for local model development — and are not a direct LLM benchmark. Pick by use case, not the top-line number.

Can you run large language models on Jetson Thor?

Yes. Jetson Thor's 128GB of memory lets it run sizable LLMs locally, and it's used for on-device language and vision-language models in robots. But DGX Spark is the better-fit tool for LLM development and fine-tuning work, thanks to its desktop form, full NVIDIA AI stack, and faster developer networking.

Which is better for robotics, Jetson Thor or DGX Spark?

Jetson AGX Thor, clearly. It's purpose-built for physical AI — configurable 40–130W power, real-time sensor I/O, and NVIDIA's robotics frameworks (Isaac, GR00T, Holoscan), designed to be embedded in a robot. DGX Spark is a developer workstation and isn't meant to ship inside a machine.

Editor's note — sources

• NVIDIA DGX Spark product page (GB10 specs, memory, FP4, networking, model-size limits): https://www.nvidia.com/en-us/products/workstations/dgx-spark/
• NVIDIA press release, "NVIDIA Blackwell-Powered Jetson Thor Now Available" (2,070 FP4 TFLOPS, 128GB, 130W, $3,499, 7.5x/3.5x vs Orin): https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-Blackwell-Powered-Jetson-Thor-Now-Available-Accelerating-the-Age-of-General-Robotics/
• Tom's Hardware (DGX Spark $3,999, on sale October 15, 2025): https://www.tomshardware.com/desktops/mini-pcs/nvidias-dgx-spark-ai-mini-pc-goes-up-for-sale-october-15-1-petaflop-developer-platform-was-originally-slated-for-may
• HotHardware (Project DIGITS renamed DGX Spark at GTC 2025): https://hothardware.com/news/nvidia-project-digits-renamed-dgx-spark-and-dgx-station
• Jetson T5000 module specs (2,560-core Blackwell GPU, 96 Tensor cores, 14-core Neoverse-V3AE, 273 GB/s, 40–130W) verified against NVIDIA T5000 datasheet values as reproduced by CNX-Software and the RidgeRun developer wiki.