The NVIDIA DGX Spark 64GB is now on sale, and it is the cheapest official way to get NVIDIA’s desktop AI supercomputer onto your desk. Starting at $4,999, the new configuration keeps the same GB10 Grace Blackwell Superchip as the original 128GB model but halves the unified memory, a move NVIDIA made as memory prices surge across the industry. It went on sale Friday, October 23, 2026, through six manufacturing partners — and it arrives with new clustering software that lets two units pool their memory. Here is everything you need to know: the price, the specs, what it can actually run, and whether 64GB is enough for your work.
NVIDIA DGX Spark 64GB Price and Release Date
The NVIDIA DGX Spark 64GB costs $4,999 to start and went on sale on October 23, 2026. Unlike the original model, there is no Founders Edition for this tier: NVIDIA is selling the 64GB configuration exclusively through six OEM partners — Acer, ASUS, Dell, Gigabyte, HP, and MSI. Pricing and exact configurations can vary slightly by partner, so the $4,999 figure is a starting price, not a single fixed SKU.
That price needs context. The original 128GB DGX Spark launched at $3,999, rose to $4,699 in February 2026, and now lists at around $6,950, according to reporting from The Register and Tom’s Hardware that attributes the climb to soaring memory costs. In other words, the new “cheaper” model costs about 25% more than the original model did at launch — but it is roughly $1,950 less than the 128GB model costs today. That is the trade the 64GB model is built around: less memory, a lower price than the current full-fat model, and the same silicon underneath.
NVIDIA DGX Spark 64GB Specs
The headline spec of the NVIDIA DGX Spark 64GB is what did not change. NVIDIA did not build a cut-down chip for the cheaper tier; it built the same machine with less memory attached:
- Chip: GB10 Grace Blackwell Superchip, identical to the 128GB model
- CPU: 20-core Arm processor, co-designed with MediaTek (reported as 10 Cortex-X925 plus 10 Cortex-A725 cores)
- Memory: 64GB unified LPDDR5X, shared between CPU and GPU
- Memory bandwidth: 273 GB/s — the same as the 128GB model, not halved with capacity
- Networking: built-in ConnectX-7, 200GbE
- Operating system: DGX OS (Ubuntu-based) with NVIDIA’s full AI software stack, including CUDA-X libraries, Nemotron models, and support for Ollama, vLLM, and PyTorch
- AI compute: up to 1 petaflop of AI performance per unit
The unchanged bandwidth matters more than it looks. Because the memory bus stays fully populated, the 64GB model keeps the same data throughput as its bigger sibling — it simply runs out of room sooner. One thing NVIDIA has not pinned down as a single figure is storage: partner configurations vary, and reporting differs on whether the 64GB model’s storage is halved versus the 128GB model’s 4TB. Check the specific partner listing before you buy rather than assuming a storage number.
What the DGX Spark 64GB Can Run
NVIDIA says a single NVIDIA DGX Spark 64GB can run AI models of up to roughly 100 billion parameters entirely on-device, with no cloud connection required. The models NVIDIA names for the machine include Qwen3.8 27B, Gemma 4 26B-A4B, and a 24GB version of Nemotron 3.5 Bolt 74B.
In practice, what fits depends on precision. A model stored at 4-bit quantization takes roughly half a gigabyte per billion parameters, so a 70B model at 4-bit needs around 40GB of weights plus room for context — comfortable on 64GB. The same model at full BF16 precision would need about 140GB and would not fit on one unit at all. That is the real dividing line for buyers: if your work lives in the 30B-and-under class of models, or quantized versions of larger ones, 64GB is plenty. If you routinely run big dense models at high precision, you will need to cluster or step up to 128GB.
The pitch is a private AI workstation. A coding or research agent can run around the clock on local data without sending prompts to a remote API — the kind of setup where privacy, latency, and freedom from per-token cloud bills matter more than raw scale.
DGX Spark 64GB Clustering: Two Units, 128GB
The most interesting part of the NVIDIA DGX Spark 64GB launch is what happens when you buy two of them. Connect two units with a single QSFP cable over their built-in ConnectX-7 networking and they pool their memory into 128GB, supporting models up to roughly 200 billion parameters. NVIDIA’s own test with Qwen3.8 27B showed up to 1.7 times the performance of a single system — a vendor test on one model, not a general speedup claim, but a sign of what clustering buys you.
NVIDIA’s Sync Cluster Assistant discovers connected units, validates the configuration, and sets up the network for you. Guided setup covers up to four systems today, while eight-unit clusters are possible in manual configuration right now — up to 512GB with 64GB units, or 1TB with 128GB units — with guided eight-system support in Sync expected later in October. NVIDIA also has a Sync Model Launcher coming at the end of October: it will download a model and its container matched to your configuration and connect it to OpenCode for browser-based coding. Note that the Launcher was not available on launch day, and NVIDIA’s Underdog desktop agent app announced alongside it is a tool for RTX Spark, not DGX Spark.
The arithmetic is worth doing before you dream of a cluster, though. Two 64GB units cost about $9,998 — more than one $6,950 128GB unit. What the pair buys you is double the compute and double the memory bandwidth, not a cheaper route to 128GB. Clustering makes sense for teams that want a shared backend or need the extra compute, not for a solo buyer trying to save money.
NVIDIA DGX Spark 64GB vs 128GB: Which Should You Buy?
| DGX Spark 64GB | DGX Spark 128GB | |
|---|---|---|
| Starting price | $4,999 | ~$6,950 (up from $3,999 at launch) |
| Availability | October 23, 2026 | Already shipping |
| Chip | GB10 Grace Blackwell | GB10 Grace Blackwell |
| Unified memory | 64GB LPDDR5X | 128GB LPDDR5X |
| Memory bandwidth | 273 GB/s | 273 GB/s |
| Networking | ConnectX-7, 200GbE | ConnectX-7, 200GbE |
| Single-unit model ceiling | ~100B parameters | Larger models with headroom |
| Sales channel | Acer, ASUS, Dell, Gigabyte, HP, MSI | NVIDIA direct and partners |
Buy the 64GB model if your models fit in 64GB and you want NVIDIA’s CUDA software stack — the tooling, model support, and developer ecosystem are the real product here, and no rival matches them yet. Buy the 128GB model if you run large models at high precision or want headroom for bigger models later. And if raw memory per dollar is your only metric, AMD-based rivals like the Lenovo ThinkCentre X Ultra, which pairs Ryzen AI Max+ chips with up to 128GB of RAM from $3,699, deserve a look — they just do not come with NVIDIA’s CUDA stack or ConnectX-7 clustering.
The Memory Crunch Behind the DGX Spark 64GB
The NVIDIA DGX Spark 64GB is the clearest consumer-facing symptom of the 2026 memory crunch. AI data centers are buying up memory supply, prices have surged, and NVIDIA’s answer was to halve the most expensive component in its desktop AI box to hold a sub-$5,000 entry price. The same crunch is behind price increases across the phone market this month, and it is why the 128GB model that launched at $3,999 now costs $6,950.
For developers, that makes the 64GB model a genuine opportunity: the cheapest official entry into NVIDIA’s local-AI stack, with a real software ecosystem and a clustering path upward. For everyone else, it is a reminder to check what you actually need before paying for memory you will never use — or buying a machine that cannot hold the models you want to run. Either way, the DGX Spark 64GB’s value hinges on one question: does your work fit in 64GB? If it does, $4,999 buys you a private AI workstation that no cloud subscription can match. If it does not, the 128GB model — or the cloud — is still the honest answer.
Internal links: NVIDIA’s RTX Spark platform launched earlier this month — see our NVIDIA RTX Spark AI PCs: specs, price, and release date coverage. For the open-weight model side of local AI, read Meta’s Muse Spark Open: 816B open-weight model specs and download. And for today’s other big tech story, see iPhone Duo display break reports from launch day.