Nvidia is cutting the price of local AI by selling you less of it. The company announced a new 64GB version of its DGX Spark desktop AI computer, priced from $4,999 and shipping October 23 through six hardware partners. That’s $1,000 more than the 128GB model’s original $3,999 launch price, for half the memory.

Welcome to 2026, where memory costs have stopped being a footnote and started setting the price of everything.

The new configuration keeps the same GB10 Grace Blackwell Superchip as the original, plus DGX OS and Nvidia’s full AI software stack. Memory bandwidth is unchanged at 273 GB/s. What you’re giving up is capacity: the 64GB model is rated for models up to 100 billion parameters running on-device, versus up to 200 billion on the 128GB version. The actual ceiling will move around with quantization and context length, but the shape of the tradeoff is clear.

Why Nvidia is selling a smaller DGX Spark

This isn’t a kindness discount. The 128GB DGX Spark’s price has been climbing all year. It launched at $3,999, rose to $4,699 in February 2026 as memory supply tightened, and now sells for roughly $6,950 at retail — nearly 75 percent more than the original sticker, per Crypto Briefing’s reporting. Nvidia attributes the increases to the ongoing memory crunch.

So the 64GB model at $4,999 is Nvidia’s answer: a way to hold the platform at a psychologically acceptable price while the flagship drifts into workstation-budget territory. VideoCardz put it bluntly: the new model costs $1,000 more than the original 128GB launch price while carrying half the memory. The math only makes sense if you assume the memory shortage isn’t ending soon.

Nvidia, evidently, assumes that.

What you still get for $4,999

Despite the memory cut, the platform is otherwise intact. Buyers get the GB10 superchip, ConnectX-7 networking, DGX OS, and the preinstalled AI stack: the Nvidia Agent Toolkit, CUDA-X libraries, Nemotron open models, and runtimes like Ollama, vLLM and PyTorch.

Nvidia is also leaning hard into clustering. Two 64GB units can be linked directly with a QSFP cable to pool memory back up to 128GB, and the company claims up to 1.7x the performance of a single system in its internal Qwen 3.8 27B test. “The cluster assistant feature detects connected units, validates device configuration and configures the ConnectX-7 network,” Nvidia said. At month’s end, Nvidia plans to ship a Sync Model Launcher that starts Qwen3.8 27B on one system or a cluster in a few clicks. Blender is among the first creator applications with support, with a downloadable installer coming soon.

That positioning is revealing. Nvidia isn’t really marketing this as a cheaper box; it’s marketing it as a modular one. Buy one now, add a second later, and you’ve effectively built your way back to the 128GB model’s memory at roughly $10,000. Funny how that works.

The memory shortage isn’t done with us

Here’s the thing that matters beyond one product. Memory prices have become the silent tax on the entire AI buildout: not just on DGX Sparks, but on data center GPUs, gaming cards, and the mid-range cards that AlexTech noted already saw 30 percent price increases this summer. Nvidia can’t hold prices on the 128GB model because it can’t hold memory costs.

The DGX Spark was supposed to be the affordable on-ramp to local AI: a desk-side box for developers who wanted to run serious models without cloud bills or cloud dependencies. At $4,999 with 64GB, it’s still cheaper than cloud spend for heavy workloads, but the affordability story now has an asterisk the size of the RAM market.

Why you should care

If you’re a developer weighing local versus cloud inference, this launch resets the math. A $4,999 box that runs 100-billion-parameter models locally is genuinely useful for agentic coding, research agents, and offline inference, and the cluster option means the hardware can grow with your needs. But don’t read the price tag as a sign that hardware is getting cheaper. It’s Nvidia adapting to a world where memory is the scarce resource and every gigabyte costs more than it did last year.

The 64GB DGX Spark ships October 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI. Independent benchmarks haven’t landed yet. Until they do, treat the 100-billion-parameter figure the way you’d treat any vendor benchmark: as a starting point, not a promise.

Frequently asked questions

How much does the Nvidia DGX Spark 64GB cost?

It starts at $4,999 and goes on sale October 23, 2026, through Acer, ASUS, Dell, Gigabyte, HP and MSI. It’s sold exclusively through those partners, not directly by Nvidia.

How much memory does the 64GB DGX Spark have?

64GB of unified memory — half the original 128GB model. Nvidia rates it for local AI models with up to 100 billion parameters, compared with up to 200 billion on the 128GB version.

Why is the 128GB DGX Spark so expensive now?

Memory prices have surged. The 128GB model launched at a $3,999 MSRP, was raised to $4,699 in February 2026 amid worldwide memory supply constraints, and now sells for about $6,950 at retail, according to reports.

Yes. Two units connect over a ConnectX-7 QSFP cable to pool memory to 128GB, supporting models up to 200 billion parameters. Nvidia claims up to 1.7x the performance of a single system in its internal Qwen 3.8 27B test.

Sources: VideoCardz, Wccftech, TechnoBezz, ThePCEnthusiast, Particle News