Nvidia is giving local AI developers a cheaper door into its desktop hardware, at a cost to memory. The company announced a 64GB configuration of the DGX Spark local-AI desktop starting at $4,999, shipping October 23, while the existing 128GB model’s price has climbed from $4,699 to $6,950.
The 64GB units come from partner manufacturers rather than Nvidia itself: Acer, ASUS, Dell, Gigabyte, HP and MSI will each sell versions of the same GB10 Grace Blackwell-based machine, and Nvidia is not offering a branded 64GB unit. All run DGX OS with the full Nvidia AI software stack preinstalled. The 128GB model, sold through Nvidia’s own channels, retains the higher unified memory that matters most for running larger models locally, and it has become a noticeably more expensive option since it first hit the market.
The memory squeeze behind the pricing
The price split tells the story of the current DRAM market. Nvidia cited tight memory supply expected to persist through 2028 as the reason the 128GB model jumped roughly 75 percent above its original $3,999 launch price. DRAM makers across the industry have prioritized memory for hyperscale data-center contracts, which pay far better than consumer or workstation demand, and desktop memory prices have risen all year as a result. Nvidia’s choices here are not unique to this product. They reflect what the entire PC hardware ecosystem is currently facing across every category that uses system memory.
| Configuration | Price | Change |
|---|---|---|
| DGX Spark 64GB | $4,999 | New configuration, Oct 23 launch |
| DGX Spark 128GB | $6,950 | Up from $4,699, roughly 75 percent higher |
| Original Founders Edition | $3,999 | Raised to $4,699 in Feb 2026 |
For developers, the tradeoff is clear enough. The GB10 chip is unchanged across both versions, which means compute performance per task is identical. What differs is how large a model fits entirely in memory, and that limit matters for local inference of anything above mid-sized open-weight models. A 64GB ceiling covers models up to roughly 100 billion parameters at lower precision, which still covers most practical experimentation and small-team workloads. Push further, past 100 billion parameters at higher fidelity, and the machine spills to slower storage. Performance drops sharply enough that the difference between the two configurations starts to feel real rather than theoretical.
Who this machine is actually for
The DGX Spark works out of the box with common inference tools including llama.cpp, Ollama, vLLM and LM Studio. Nvidia’s pitch reads straightforwardly: pick a supported framework, load a recommended model for the task, and start working the same day, with no cloud account and no GPU cluster management anywhere in the setup. For two-unit deployments, a ConnectX-7 port links the machines so Nvidia’s Sync Cluster Assistant routes workloads across both automatically. That doubles available memory without requiring a larger cluster commitment upfront, which is the practical middle ground between buying one desktop and jumping straight into infrastructure territory.
The intended audience is narrow and Nvidia’s own pricing history makes that plain. The original Project DIGITS concept promised a $3,000 starting point in 2025. What shipped was a $3,999 product, raised to $4,699 within months on memory cost grounds, and now a $6,950 high-memory path for developers who need to run genuinely large models locally. The 64GB option at $4,999 sits between those points and is priced like a workstation, not a consumer impulse purchase. Buyers who want maximum local-model flexibility at the lowest cost per dollar of memory will likely wait, while those who need a machine this month now have a cheaper entry point than they had last week.
What else is moving inside Nvidia right now
The launch lands in a busy stretch for the company. Nvidia also released its Open Agent Safety Platform last week, an open-source system pairing BlueField Sentry monitoring on BlueField-4 DPUs with an OpenShell sandbox runtime. More than 100 organizations signed on, including Anthropic, Microsoft, Oracle, CrowdStrike and Palantir, with Claude Managed Agents integrating directly. OpenAI is absent from the partner list, which has drawn attention given recent friction between the firms over public AI risk warnings. Nvidia CEO Jensen Huang framed the platform as building actual safeguards infrastructure rather than issuing public warnings about hypothetical outcomes, a distinction he has spent recent weeks making in interviews and on stage.
That release came days after Huang questioned Anthropic and OpenAI’s executive warnings about AI risk in a New York Times interview, calling some of the framing odd, and after reporting from the White House frontier AI summit described Huang pressing Anthropic CEO Dario Amodei on the accuracy of his public claims. The safety platform and this hardware refresh together show Nvidia positioning itself as the practical engineering layer between research labs that talk about risk and enterprises that need AI agents deployed with controls they can actually audit and enforce.
Where the memory market goes from here
Nvidia says memory supply constraints will persist through 2028, which frames the current pricing as a floor rather than a peak. Every major consumer electronics category, from phones to laptops to gaming PCs, is contending with the same shift of DRAM capacity toward AI data-center customers. The value proposition of local AI hardware narrows when memory prices climb faster than model efficiency improves. For local AI developers, the 64GB model is the cheapest way to still get real inference-capable Nvidia hardware this year without a data-center contract or a partner sales cycle.
Whether that lifeline holds depends on memory pricing over the coming quarters. If DRAM costs keep climbing, the $4,999 price may not survive the launch year intact. If supply loosens, the gap between the two configurations should narrow again. The broader signal to the market is the same one Nvidia sends about its data-center business: demand for memory and compute keeps outpacing what the industry can build, and buyers shopping for local AI are now absorbing that cost directly alongside the hyperscalers signing multi-billion dollar memory deals.