A new entry point, at a higher number
Nvidia announced a 64GB version of its DGX Spark desktop AI system on Friday, priced from $4,999 and going on sale on 23 October through OEM partners including Acer, Asus, Dell, Gigabyte, HP and MSI. It uses the same GB10 Grace Blackwell Superchip, the same software stack and the same ConnectX-7 networking as the existing 128GB machine.
The pricing is the part worth looking at twice. The 128GB DGX Spark launched in October 2025 at $3,999, rose to $4,699 in February 2026, and is now reported at $6,950, according to Back2Gaming. On those figures the new cut-down model costs 25 per cent more than the full-memory model did at launch a year ago.
Nvidia’s own product page lists no prices.
What 64 gigabytes gets you
About 8GB is reserved for the system, leaving roughly 56GB for model weights and KV cache. Nvidia says a single unit handles models up to 100 billion parameters, against 200 billion for the 128GB machine.
The clustering story is the interesting engineering. Each DGX Spark carries an onboard ConnectX-7 NIC, so two units connect directly over a QSFP cable on a 200GbE link with no switch in between, pooling to 128GB of unified memory and the same 200-billion-parameter ceiling. Nvidia says two clustered 64GB systems delivered up to 1.7 times the performance of a single 128GB system on Qwen 3.8 27B — a vendor-run comparison, published by the vendor, on the day it announced the product.
That figure is plausible on its face, since two machines bring two sets of compute rather than only more memory. It is still Nvidia’s number about Nvidia’s hardware.
Why memory price is the whole story
The DGX Spark exists because running a large model locally is bounded by memory capacity rather than by compute. The repricing of the 128GB model over twelve months tracks what has happened to high-bandwidth memory supply during the same period, and it is the clearest consumer-visible signal of that squeeze so far.
The machines ship with DGX OS, the CUDA-X AI libraries, Nvidia’s Agent Toolkit and Nemotron open models, and support llama.cpp, Ollama, vLLM and LM Studio. A Sync Model Launcher for deploying across single or clustered units is due at the end of October.
What to watch is whether the 64GB price holds through the launch window. The 128GB model’s history suggests the list price of a local AI box is now a function of the memory market rather than of the product.
