NVIDIA’s 64GB DGX Spark starts at $4,999 on 23 October

What is shipping on 23 October?

NVIDIA’s 2 October Local AI blog says the new 64GB unified-memory DGX Spark configuration will be sold by Acer, ASUS, Dell, Gigabyte, HP, and MSI starting Friday, 23 October, from $4,999. The product page marks 64GB as available exclusively through participating OEM partners.

NVIDIA says the SKU retains the GB10 Grace Blackwell Superchip, DGX OS, and the full NVIDIA AI software stack used on the 128GB model. Stated memory bandwidth is 273 GB/s of coherent unified memory. Tensor performance is listed as up to 1 PFLOP FP4 on the product sheet.

This is a DGX Spark desktop SKU, not the Windows RTX Spark PCs NVIDIA and Microsoft have been teasing; for that hybrid-PC thread see Microsoft and NVIDIA’s RTX Spark Windows note.

Read the 23 October date as NVIDIA’s stated availability window for the 64GB configuration at the listed starting price, not as a guarantee that every region’s channel partners will have units on shelves that morning. The product story sits beside the larger DGX Spark / RTX Spark builder-class narrative from earlier in the week: a compact Grace-Blackwell-class desk-side box aimed at local inference and fine-tuning rather than a rack DGX cloud node. Keep those form-factor claims in NVIDIA’s column until you measure thermals and sustained tokens/sec on your models.

How do two 64GB units cluster?

Every DGX Spark includes a ConnectX-7 NIC. NVIDIA says two 64GB units can connect directly with a QSFP cable, pooling memory to 128GB and expanding claimed model support to about 200 billion parameters while doubling memory bandwidth.

NVIDIA Sync Cluster Assistant, part of the NVIDIA Sync app, is described as detecting connected units, validating configuration, and configuring the ConnectX-7 network so software does not need reconfiguration when scaling from one box to two.

In NVIDIA’s Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7× performance versus one system. That multiplier is NVIDIA’s internal comparison, not an independent lab result.

Dual-unit clustering is the capacity headline for teams that almost fit in 64GB but need headroom for KV cache or multi-adapter serving. Confirm which interconnect NVIDIA documents for the pair, what software versions pin the cluster path, and whether your framework’s tensor-parallel recipes are supported without custom patches. Two boxes are not automatically a single 128GB address space for every runtime.

What software stack does NVIDIA list?

The blog lists NVIDIA Agent Toolkit, CUDA-X libraries, Nemotron open models, and runtimes such as Ollama, vLLM, llama.cpp, LM Studio, and PyTorch with CUDA as supported paths. A Blender installer for the platform is described as coming soon.

NVIDIA Sync Model Launcher is scheduled for the end of October to download and launch Qwen3.8 27B on one unit or a cluster and expose it to a laptop, including an OpenCode setup path. That launcher was not claimed as shipping on 23 October.

For agent runtime controls on NVIDIA’s stack, see our separate note on OpenShell; it is a different product from DGX Spark hardware.

How should buyers read the price and capacity claims?

$4,999 is NVIDIA’s US starting price for partner systems. Partner street prices, memory market moves, and regional pricing can differ; NVIDIA’s post gives no UK or EU prices.

“Up to 100-billion-parameter models” on one 64GB unit and “up to 200 billion” on a pair are NVIDIA marketing capacity lines. Usable context, quantization, and workload mix will decide what actually fits.

The 128GB DGX Spark remains available for larger single-node work. The 64GB SKU is positioned as a lower entry point with a clustering path, not a replacement for every 128GB buyer.

Compare $4,999 against the total cost of a used workstation plus discrete GPUs only after matching memory, interconnect, and supported software—not just peak FLOPs slides. If your workload is latency-sensitive agent serving, measure batch-1 tokens/sec and cold-start load time; if it is LoRA fine-tuning, measure tokens/sec at your rank and sequence length. Marketing GB figures are necessary but not sufficient for either job.

What should a team try after 23 October?

If you buy a unit, NVIDIA’s get-started list is: install llama.cpp, Ollama, vLLM, or LM Studio; load a model that fits; for two boxes, cable ConnectX-7 and run Sync Cluster Assistant.

Demand partner-quoted power, storage, and warranty terms. The blog’s 240 W power-supply and 1.2 kg chassis figures are product-sheet claims to verify on the SKU you order.

Treat the 1.7× dual-unit Qwen figure as a vendor demo until you rerun your own serving stack.

On day one, flash the supported stack NVIDIA lists, load one dense 30–70B-class model you already score, and one MoE with an active-parameter count closer to your production agent. Record power draw and fan noise if the box will sit beside a human. Only then decide whether a second unit’s cluster premium beats renting cloud GPUs for the same experiments.

What did we not test?

We did not purchase, power on, or benchmark a DGX Spark 64GB system, nor run Sync Cluster Assistant. This article reports NVIDIA’s 2 October blog and product page only.