As of October 19, 2026, the NVIDIA DGX Spark with 128 GB of memory and a 4 TB SSD is listed at $6,950 on NVIDIA Marketplace. The same GB10 chip is sold in partner boxes from ASUS, Acer, Dell, GIGABYTE, HP, Lenovo and MSI; the cheapest one with a listed price there is the ASUS Ascent GX10 with a 1 TB SSD at $5,999. A 64 GB version is announced but not on sale yet, and NVIDIA says it will come only from partners.
DGX Spark is a 1.2 kg desktop box for running and fine-tuning large AI models locally. It isn’t a robot computer. If you want a computer that goes inside a robot, you’re looking for Jetson Thor instead (covered below).
What you get for the money
These figures come from the specifications on NVIDIA’s DGX Spark product page:
- Chip: NVIDIA GB10 Grace Blackwell Superchip, with a 20-core Arm CPU (10 Cortex-X925 + 10 Cortex-A725) and a Blackwell GPU with 5th-generation Tensor Cores.
- AI performance: up to 1 petaFLOP at FP4 precision. FP4 is a 4-bit number format; the figure is lower at higher precisions.
- Memory: 128 GB LPDDR5x shared by the CPU and GPU (64 GB in partner-only models), on a 256-bit bus with 273 GB/s of bandwidth.
- Storage: up to 4 TB NVMe with self-encryption.
- Networking: a ConnectX-7 NIC at 200 Gbps for linking units, plus 10 GbE, Wi-Fi 7 and Bluetooth 5.4.
- Ports: 4 USB-C, 1 HDMI 2.1a, and up to 3 DisplayPort outputs over USB-C.
- Power: a 240 W power supply; the GB10 chip itself is rated at 140 W.
- Size and OS: 150 x 150 x 50.5 mm, running NVIDIA DGX OS (Linux).
The number that matters most is the memory. 128 GB of unified memory lets one box hold models that don’t fit on a 24 GB or 32 GB gaming GPU. The bandwidth, 273 GB/s, is much lower than a high-end desktop GPU’s, so a model that fits on both will usually generate text faster on the desktop GPU. DGX Spark wins on size of model, not on speed per token.
How big a model it can run
NVIDIA publishes these limits on the product page:
- One 64 GB unit: up to 100 billion parameters.
- One 128 GB unit: up to 200 billion parameters for inference, and fine-tuning up to 70 billion parameters.
- Two 128 GB units linked: up to 400 billion parameters.
- Four units linked: up to 700 billion parameters.
These are NVIDIA’s numbers and they assume FP4 or similarly compressed weights. A model stored at higher precision takes more memory, so the practical limit is lower.
Where to buy one
- NVIDIA Marketplace: the Founders-style DGX Spark (128 GB, 4 TB) at $6,950. Two-unit and four-unit bundles are listed with a “Contact Us” button instead of a price; the four-unit bundle doesn’t include the network switch you need to connect them.
- Partner systems: the same marketplace page links to GB10 boxes from ASUS, Acer, GIGABYTE, Lenovo, HP, MSI and others. Most send you to the maker’s own store, so the final price is set there. Partner boxes often ship with a 1 TB or 2 TB SSD instead of 4 TB, which is one reason they cost less.
- Retail and resellers: NVIDIA’s launch announcement named Micro Center stores in the U.S. and NVIDIA channel partners worldwide as sales channels. Prices there change, so check the seller’s page on the day you buy.
The 64 GB models (ASUS Ascent GX10 64GB, MSI EdgeXpert 64GB, GIGABYTE AI TOP ATOM 64GB and Acer Veriton GN100 64GB) were all marked “Coming Soon” on October 19, 2026, with no prices. NVIDIA’s product page has a “Notify Me” sign-up for that version.
Don’t confuse DGX Spark with RTX Spark, which NVIDIA uses for a separate line of Windows laptops and desktops. They’re listed in the same marketplace but run a different operating system and target different buyers.
Who it makes sense for
DGX Spark fits people who need to run or fine-tune a model larger than their GPU can hold, and who want to do it locally rather than rent cloud GPUs. Typical cases are testing a 70B-class language model, fine-tuning a vision-language model, or running a robotics foundation model for evaluation. NVIDIA ships its AI software stack preinstalled, including NIM microservices, and lists Isaac, Metropolis and Holoscan as supported frameworks for edge and robotics work.
It’s a weaker fit in three cases:
- You mainly need speed on models under about 30 GB. A desktop RTX GPU will usually be faster for the money.
- You want to train robot policies in simulation. Isaac Lab’s installation guide supports DGX Spark with some features missing, and RTX desktop GPUs remain the standard setup. See our Isaac Lab getting-started guide for the requirements.
- You need Windows or x86 software. DGX OS is Linux on Arm, so x86-only tools won’t run natively.
DGX Spark vs Jetson Thor
Both use Blackwell GPUs, but they do different jobs. DGX Spark sits on a desk and is plugged into the wall; you develop, fine-tune and test models on it. Jetson AGX Thor is a module built to run inside a robot, with power modes and camera and sensor interfaces for that job. A common setup is to fine-tune on a workstation or DGX Spark and deploy the trained model to Jetson Thor on the robot. Our Jetson AGX Thor price guide covers the robot side.
If you’re choosing models to run on either machine, our guides to NVIDIA Cosmos and Physical Intelligence’s openpi models list their GPU memory needs.