News · 2026-10-03

NVIDIA's cheaper DGX Spark halves the memory and starts at $4,999

NVIDIA will sell a 64GB version of its DGX Spark desktop AI computer through six PC makers starting at $4,999 on October 23, the company announced on October 2. It keeps the same Grace Blackwell chip and software as the 128GB model, which NVIDIA's own store currently lists at $6,950.

Key facts

• Price: from $4,999 for 64GB, against $6,950 for the 128GB model on NVIDIA's store, which was out of stock when checked.
• When: on sale Friday, October 23, 2026.
• Who sells it: only Acer, ASUS, Dell, Gigabyte, HP and MSI; NVIDIA will not sell the 64GB version directly.
• Primary source: NVIDIA's announcement.

The DGX Spark is a small box that sits on a desk and runs AI models locally instead of in a cloud data center. Its central selling point is unified memory: the processor and the graphics engine share one pool, so a big model does not have to squeeze into a separate, smaller graphics-card memory. Memory size, more than raw speed, decides which models a machine like this can hold at all. That is why the 64GB configuration matters: it lowers the entry price by cutting the one thing that most limits what you can run.

What NVIDIA announced

NVIDIA says the 64GB model "supports up to 100-billion-parameter models and the agentic applications built on them, fully on device." It retains the GB10 Grace Blackwell superchip, the DGX OS operating system and the full NVIDIA software stack, and supports common tools such as Ollama, vLLM, llama.cpp and PyTorch out of the box.

Two units can be cabled together through their built-in ConnectX-7 network ports. NVIDIA says a pair pools memory to 128GB, supports models up to 200 billion parameters, and doubles memory bandwidth. "Two 64GB units clustered together don't just double the memory," the company wrote, citing a test in which two clustered systems ran its Qwen 3.8 27B workload up to 1.7 times faster than one. A new NVIDIA Sync Cluster Assistant app is meant to detect the second unit and configure the link automatically, and a Model Launcher arriving at the end of October will download and start Qwen3.8 27B and connect it to the OpenCode coding agent.

The price story

The 64GB model arrives against a backdrop of rising prices. NVIDIA's February price notice raised the Founders Edition from $3,999 to $4,699, citing worldwide memory supply constraints. Its live store listing now shows the 128GB model at $6,950 and out of stock. NVIDIA has not explained the move from $4,699 to $6,950. At list prices the new model costs $1,951 less for half the memory; two 64GB units cost $9,998 before cables and tax, more than one 128GB unit.

The memory squeeze is the same one Ground Truth reported this week, when Micron said memory and storage supply would stay tight into 2027 and 2028. A cut-down configuration at a lower price is one way to keep a product reachable when the expensive part is the memory itself.

Local-hardware friction also showed up elsewhere this week. TechSpot reported a buyer's account of being asked at a Micro Center store in California to show ID and sign a declaration that an RTX 5090 prebuilt PC would not leave the United States. That is one buyer's report about one store; Micro Center has not confirmed a chain-wide policy, and no new federal rule requiring the form has been found.

Why it matters

For someone who wants to run capable open models at home or in a small office, the question is usually "what is the cheapest machine with enough memory?" The 64GB Spark gives a lower starting point inside NVIDIA's ecosystem, and the clustering story lets a buyer add a second unit later rather than paying for 128GB up front. Our lesson on why LLM inference is memory-bound explains why memory, not compute, usually sets the ceiling; quantization is how people fit bigger models into less of it.

The caveat

"Up to 100 billion parameters" depends heavily on how much the model is compressed and how long a context you need, since the working memory for a long conversation comes on top of the model's weights. The 1.7 times clustering speedup is NVIDIA's result on one test, not a general guarantee. Partner pricing may differ from NVIDIA's starting figure, and the systems are not on sale until October 23, so there are no independent reviews yet.

Primary source, verified: read the paper →

Ground Truth. (2026, October 3). NVIDIA's cheaper DGX Spark halves the memory and starts at $4,999. Ground Truth. https://groundtruth.day/news/nvidia-dgx-spark-64gb-starts-at-4999.html

@misc{groundtruth:nvidia-dgx-spark-64gb-starts-at-4999,
  title  = {NVIDIA's cheaper DGX Spark halves the memory and starts at $4,999},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/nvidia-dgx-spark-64gb-starts-at-4999.html}
}

Topics: hardware · local-ai · nvidia · industry · inference