NVIDIA DGX Spark 64GB Goes on Sale Today at $4,999: What the Price Actually Buys

Today is on-sale day for NVIDIA's new entry-level local-AI box, with a twist baked into the price tag. The 64GB DGX Spark, announced October 2, reaches Acer, ASUS, Dell, Gigabyte, HP and MSI shelves on October 23 at a starting price of $4,999. It carries the same GB10 Grace Blackwell Superchip as the 128GB model, with the unified memory cut in half. And here is the number NVIDIA's announcement does not headline: this "budget" configuration costs $1,000 more than the 128GB model did when it launched at $3,999 a year ago — because the memory shortage has since pushed the 128GB Founders Edition to roughly $6,950. I read NVIDIA's announcement and the launch coverage so you can see exactly what $4,999 buys.

Quick answer

• What it is: a 64GB unified-memory configuration of NVIDIA's DGX Spark desktop AI workstation, announced October 2, 2026.
• Price and availability: starts at $4,999, on sale October 23, 2026 — sold exclusively through six hardware partners, with no NVIDIA-direct Founders Edition.
• What you get: the same GB10 Grace Blackwell Superchip, 20-core Arm CPU, 273 GB/s of unified memory bandwidth, ConnectX-7 networking, DGX OS and the NVIDIA AI software stack as the 128GB model. NVIDIA says one unit runs models with up to 100 billion parameters.
• The catch: the 128GB model now lists at roughly $6,950, up about 74% from its $3,999 debut, and clustering two 64GB units to reach 128GB of pooled memory costs about $9,998.
• Who it is for: developers and small teams who want to run AI agents and inference locally, privately, without cloud bills — and who don't need the largest models.

What changed and what didn't

Almost nothing about the compute platform changed. Per NVIDIA's announcement, the 64GB Spark keeps the GB10 Grace Blackwell Superchip with its 20-core Arm CPU, the same 273 GB/s of LPDDR5x unified-memory bandwidth, the same rated 1 petaFLOP of sparse FP4 compute, the built-in ConnectX-7 NIC, DGX OS, and the full NVIDIA AI software stack — the Agent Toolkit, CUDA-X libraries, Nemotron open models, and runtimes like Ollama, vLLM and PyTorch.

Two things did change. First, unified memory is halved: 64GB instead of 128GB. NVIDIA advertises support for models with up to 100 billion parameters on a single unit, compared with roughly 200 billion on the 128GB configuration. Those are vendor-advertised ceiling figures, and as launch coverage has pointed out, they describe capacity under suitable configurations — a 100-billion-parameter model only fits in 64GB with heavy quantization, so treat the ceiling as an upper bound, not a promise for every model.

Second, there is no NVIDIA Founders Edition this time. The 64GB SKU is sold only through Acer, ASUS, Dell, Gigabyte, HP and MSI. That matters because pricing and configuration now vary by partner. The "starting at" in $4,999 is doing real work: partners will differentiate on extras like SSD capacity, and reporting indicates NVIDIA has not published a standard storage capacity for the 64GB SKU. The 128GB Founders Edition shipped a 4TB self-encrypting NVMe drive; on the 64GB model, one outlet reports the storage is halved, while others note partners set their own SSD — so confirm the drive size on the partner listing before you buy.

The price history, in one table

The DGX Spark's price story is the memory shortage's price story. NVIDIA has not added features between these hikes; both increases were tied to memory supply.

A couple of honest caveats on this table. NVIDIA's own announcement did not state a new 128GB price — the ~$6,950 figure comes from ServeTheHome and Tom's Hardware, and one German outlet reported around $6,000, so treat the exact number as vendor-reported-to-press rather than an official MSRP. And the 64GB's $4,999 starting price is, per NVIDIA's statement, exactly that: a starting price, with partner configurations priced above it.

The irony that several outlets landed on is worth stating plainly: the 64GB Spark — a box built by cutting memory — costs about $1,000 more than the fully loaded 128GB model did at launch. "Providing less memory" has become NVIDIA's way of keeping an entry price under $5,000, and that says more about the DRAM market than about the chip inside.

The cluster math: why two boxes cost more than one

NVIDIA's growth path for the 64GB is to buy a second one. Two units connect over their built-in ConnectX-7 ports with a QSFP cable (no switch needed for a pair), and the Sync Cluster Assistant software discovers the units and configures the network, pooling memory to 128GB. NVIDIA says a two-system cluster supports models up to 200 billion parameters.

The arithmetic is awkward, as AI Weekly put it: two 64GB boxes at $4,999 each run to about $9,998 — roughly $3,000 more than a single $6,950 128GB unit for the same 128GB memory pool. The upside is that you also get double the compute and memory bandwidth, which matters for compute-bound work rather than purely memory-bound workloads. NVIDIA's own Qwen 3.8 27B test showed two clustered 64GB systems delivering up to 1.7x the performance of a single unit — but that is NVIDIA's own test on one model, not a general performance doubling.

One more honest note: splitting a single large model across two linked boxes does not make it run faster. Clustered memory buys capacity — the ability to fit a model that one box cannot hold — not a speed boost across every workload.

The memory shortage behind all of this

This is not an NVIDIA-only problem. AI data centers are consuming an outsized share of global DRAM and HBM production — one frequently cited estimate is that a single AI server needs 8 to 10 times more memory than a regular server. Samsung, SK Hynix and Micron have been redirecting wafer capacity toward higher-margin HBM for AI data centers, which tightens supply and raises prices for the conventional DRAM that desktops and laptops use. The Register reported DRAM prices rising roughly 70% on this demand, and Gartner has projected a 130% year-over-year DRAM price increase for 2026.

The shortage has a long tail. Micron has warned that DRAM supply will stay tight through 2028, and the pricing creep is showing up at the consumer end too: per Tom's Hardware's RAM index, a 64GB DDR5 kit now runs $680–$1,070 — roughly four times its 2025 price. When memory is the most expensive part of the box, halving it is the only lever left for holding an entry price steady. That is exactly what the 64GB Spark is.

How the 64GB Spark compares to its alternatives

If your goal is running large local models, the Spark is not the only option — and on raw price per gigabyte of memory, it is not even the cheapest in its own class. Here is how the 64GB Spark stacks up:

The Framework Desktop line is the most direct price comparison: per Pandromeda's tracking of Framework's own configurator this week, the 128GB Strix Halo configuration ships at $3,449 — half the price of the 128GB Spark for the same memory capacity. (Framework's prices are themselves up from 2025, which tells you the DRAM shortage is hitting AMD's partners too.) The RTX Spark line is a different proposition: Arm-based Windows 11 machines aimed at general Windows users rather than AI developers. Apple's Mac Studio with M3 Ultra, configurable up to 512GB of unified memory with over 800 GB/s of bandwidth, remains the high-memory option for Mac-based developers.

The Spark's differentiator is not memory-per-dollar. It is NVIDIA's software stack and the GB10's architecture: a developer box preconfigured with CUDA-X, the Agent Toolkit and model runtimes, built for prototyping, fine-tuning and local agent inference with everything kept on-device. If you want a general-purpose machine, the alternatives above give you more memory for less. If you want the NVIDIA developer appliance, the 64GB is now the cheapest way in.

Pros and cons

Pros: the lowest entry price into the GB10 platform at current list prices; identical compute, bandwidth and software stack to the 128GB model; same-day clustering path to 128GB when workloads outgrow one box; familiar OEM partners with their own support channels; purpose-built for local agents and fine-tuning without cloud bills.

Cons: half the memory ceiling — one box tops out at NVIDIA-advertised 100B-parameter models, which assumes heavy quantization; no NVIDIA-direct Founders Edition, so final prices and SSD capacities vary by partner; clustering two units costs ~44% more than one 128GB unit for the same memory pool; the whole lineup's pricing remains exposed to the DRAM shortage, so $4,999 may not hold.

What to watch

NVIDIA's Sync Model Launcher is scheduled for release by the end of October, and it is aimed squarely at the 64GB's audience: one-click setup of Qwen 3.8 27B on a single Spark or a cluster, accessible from a laptop. There are also playbooks for agent tools like OpenClaw, NemoClaw and Hermes Agent. If the software onboarding is as smooth as NVIDIA promises, the 64GB Spark's value proposition shifts from "a compromised Spark" to "a local-agent appliance with an upgrade path" — which is clearly how NVIDIA wants developers to see it.

For developers who prefer software-only local AI, the trend behind this hardware is bigger than NVIDIA: open models small enough to run on consumer hardware are the reason a 64GB box is viable at all. Perplexity's fully local Mac client, which runs frontier models entirely on-device behind a 36GB hardware bar, is the software side of the same bet — we covered its setup and credit catch here. And for coding workloads specifically, laptop-scale open models like Recursive's Flint are now competitive on a machine you already own — see our Flint breakdown.

FAQ

How much does the 64GB DGX Spark cost, and where can I buy it?

$4,999 to start, on sale October 23, 2026. It is sold exclusively through six hardware partners — Acer, ASUS, Dell, Gigabyte, HP and MSI — not directly from NVIDIA. Partner configurations above the base price will exist, so check the listing.

Is it cheaper to buy two 64GB units instead of one 128GB Spark?

No. Two 64GB units at $4,999 each cost about $9,998 for a pooled 128GB of memory, versus roughly $6,950 for one 128GB unit. The pair gives you double the compute and memory bandwidth, which can be worth it for compute-bound workloads — but not cheaper total memory.

What models can the 64GB Spark actually run?

NVIDIA's advertised ceiling is up to 100 billion parameters on a single unit. That is a vendor claim, and it assumes heavy quantization (NVFP4 compression) to fit a model that large into 64GB. Practical launch coverage puts the comfortable range around 80–100 billion parameters depending on your precision settings. For context, a 27-billion-parameter model like Qwen 3.8 27B — the model in NVIDIA's own clustering benchmark — is the kind of workload this box handles comfortably.

Why did the 128GB DGX Spark's price go up so much?

Memory supply, per NVIDIA itself. The Founders Edition rose from $3,999 at launch to $4,699 in February 2026 — NVIDIA's developer forum attributed that to "worldwide constraints in memory supply" — and then to roughly $6,950 in October 2026. The broader cause is AI data centers consuming DRAM and HBM capacity faster than fabs can add it, a dynamic Micron warns could stay tight through 2028.

Should I buy the 64GB Spark or wait?

Buy it if you want to run AI agents, local inference or fine-tuning on NVIDIA's stack today and your workloads fit in NVIDIA's 100B-parameter ceiling. Skip it if you need the largest local models — one 128GB unit remains the cheapest way to 128GB. And watch partner listings: with no Founders Edition and no published standard SSD, the final price and configuration of "the" 64GB Spark will vary.

Key takeaways

• The 64GB DGX Spark goes on sale today at $4,999 starting, exclusively through Acer, ASUS, Dell, Gigabyte, HP and MSI — there is no NVIDIA-direct Founders Edition this time.
• It is the same GB10 compute platform with half the memory; NVIDIA's advertised model ceiling drops from roughly 200B to 100B parameters on one unit, and all ceiling figures are vendor-advertised.
• The 128GB model's climb from $3,999 to roughly $6,950 is about memory supply, not new features — NVIDIA said so itself in February.
• The clustering math does not favor memory-per-dollar: two 64GB units cost about $9,998 for pooled 128GB versus roughly $6,950 for one 128GB unit; the upside is double compute and bandwidth.
• Check the SSD and final price on the partner listing: NVIDIA publishes no standard storage capacity for the 64GB SKU, so configurations will differ.

Published October 23, 2026 — the 64GB SKU's on-sale date. Local AI hardware is moving fast and pricing is hostage to the DRAM market, so verify the partner price before you buy.