Nvidia will start selling a 64GB version of its DGX Spark desktop AI computer on Friday, October 23, at a starting price of $4,999. It's pitched as the cheaper way into the Spark lineup. The odd part is that it costs $1,000 more than the original 128GB Spark did when it went on sale a year ago.
The timing explains the price. The memory chips inside the Spark are getting scarcer and more expensive, and Nvidia has now raised the price of the 128GB model twice. It reportedly lists at $6,950, almost 75% above its $3,999 launch price. So the "entry" Spark mostly tells you what the global memory squeeze costs developers who want to run AI models on their own desk instead of renting cloud GPUs.
What Nvidia is actually selling
The 64GB Spark uses the same GB10 Grace Blackwell Superchip, DGX OS operating system and Nvidia AI software stack as the 128GB model, Nvidia said in a blog post on October 2. It also keeps the built-in ConnectX-7 networking, which matters later for linking two machines.
The pitch leans on how fast open models are shrinking. "Local AI is becoming more useful by the token," Nvidia wrote. It argues that as AI agents move into everyday development, capable open models now fit on smaller hardware.
Nvidia says one 64GB unit can run models with up to 100 billion parameters (parameters are the internal values a model learns, and a rough measure of its size). Back2Gaming put the practical range at 80 to 100 billion. A model that size only fits in 64GB if it's heavily compressed, so expect that ceiling to depend on the quantization (compression) settings you choose.
The box ships with Nvidia's Agent Toolkit, CUDA-X AI libraries, Nemotron open models and common runtimes such as Ollama, vLLM and PyTorch. Nvidia also points to setup guides, which it calls playbooks, for agent tools including OpenClaw, NemoClaw and Hermes Agent.
Some things do get cut. The 64GB version has half the onboard storage of the 128GB model, and Nvidia won't sell its own Founders Edition. It will only come from partners. Those partners are Acer, ASUS, Dell, Gigabyte, HP and MSI. When the original Spark launched in October 2025, Nvidia also listed Lenovo as a partner. Lenovo isn't named for the 64GB model.
Memory bandwidth stays at 273 GB/s, the same as the larger model, according to Hardware Busters. Bandwidth is how quickly the chip can read from memory, and for AI inference it often limits speed more than raw compute does.
How a $3,999 machine became a $6,950 one
The Spark's price history explains why a cut-down model costs more than the full one once did.
Nvidia first showed the machine as Project Digits at CES in January 2025. It went on sale October 15, 2025, for $3,999, rated at one petaflop of AI performance with 128GB of unified memory, meaning the CPU and GPU share one pool.
In late February 2026, Nvidia raised the Founders Edition price by $700, or about 18%, to $4,699. It blamed memory. "The price adjustment reflects industry wide memory supply constraints," the company said at the time.
It didn't stop there. By September, partner versions of the 128GB system were selling for $5,999 to $9,000. Then, on the same Friday the 64GB model was announced, the 128GB Spark went to $6,950, as first reported by The Register. Nvidia's own blog post about the 64GB model doesn't mention the 128GB price.
Laid end to end, the climb is hard to miss. The 128GB Spark launched at $3,999 in October 2025. By February 2026 it had risen to $4,699. This October it sits at $6,950, while the new 64GB model starts at $4,999.
The root cause is in the memory market. Chipmakers are moving production toward the high bandwidth memory used in data center AI accelerators, which leaves less capacity for the LPDDR5X memory that machines like the Spark use. Micron has warned that tight supply could last through 2028. So the AI boom is raising the price of the hardware meant to let developers do AI work locally.
The two-box option, and what it really costs
Nvidia's answer for people who need more memory is to buy a second Spark. A feature in its NVIDIA Sync app, the Cluster Assistant, links two 64GB units so they pool their memory to 128GB. Nvidia says that setup can handle models up to 200 billion parameters, with twice the memory bandwidth.
On speed, Nvidia says two linked 64GB systems ran up to 1.7 times faster than a single one in its own test with a model it calls Qwen 3.8 27B. That's a vendor benchmark on one model, so treat it as a best case, not a guarantee for your workload. Back2Gaming reports that up to four units can be linked over the 200Gb Ethernet connection.
The price comparison isn't flattering, though. Two 64GB units at $4,999 each cost $9,998, about $3,000 more than one 128GB Spark at its reported $6,950 price. Unless you specifically want the extra bandwidth and compute of two chips, pairing two small boxes is the expensive way to get 128GB.
The Cluster Assistant ships with the hardware, but a companion tool, the NVIDIA Sync Model Launcher, is listed as coming at the end of the month. Nvidia also says Blender support is "coming soon". Those are promises for now, not shipped features.
The 64GB Spark makes the most sense for a developer who wants Nvidia's CUDA software environment on their desk and mostly works with models in the tens of billions of parameters, the size class Nvidia says now handles real agent work. For them, keeping the full software stack at $4,999, about $2,000 below the reported $6,950 for the 128GB model, is a real option.
It makes less sense for anyone who remembers the Spark at $3,999. Nvidia originally talked about the Project Digits idea as a way to give individual developers a small AI supercomputer for the price of a high-end PC. A year later, the cheapest way into that system costs more, with half the memory and half the storage. Several outlets led their coverage with exactly that point.
The partners are in an awkward spot too. With no Nvidia-branded 64GB model, Acer, ASUS, Dell, Gigabyte, HP and MSI set the shelf prices and storage options. The $4,999 figure is a starting price, and the 128GB partner machines have already ranged far above Nvidia's list price.
Nvidia frames the Spark as a place to build and test AI locally, then scale the work up to bigger Nvidia systems, using the same software stack throughout. That makes it less a profit center than an on-ramp that keeps developers on Nvidia's tools from day one.
Nvidia (NVDA) stock closed Friday at $233.95, up 1.34%. Nothing in the coverage ties that move to the Spark news. The memory shortage behind the price increases is a broader worry for Nvidia's customers, because higher component costs can squeeze how much they spend on AI hardware.
What to watch next
The first date is October 23, when the 64GB systems go on sale through the six partners. Watch for real partner prices and storage options, and for whether the cheapest configurations actually come in at $4,999.
By the end of October, Nvidia says the Sync Model Launcher should arrive. Independent tests of the two-unit setup will show whether the 1.7x speedup holds outside Nvidia's own benchmark.
Longer term, the price of every memory-heavy device, from AI desktops to laptops, depends on how quickly memory makers add capacity, and Micron's outlook runs to 2028. If you build with local models, is a $4,999 machine with 64GB of memory still worth it, or does the math now favor renting cloud GPUs by the hour?
Sources
- 1.NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI · NVIDIA
- 2.Nvidia's New DGX Spark Costs More Despite Having Half the Memory · TipRanks
- 3.Nvidia DGX Spark 64GB Arrives at $4,999 as the 128GB Model Jumps to $6,950 · Hardware Busters
- 4.NVIDIA DGX Spark Sees $700 Price Hike Due To Ongoing Memory Shortages, $4699 New Pricing · Wccftech
- 5.NVIDIA DGX Spark 64GB Launches at US$4,999, Built to Be Clustered · Back2Gaming
- 6.Nvidia DGX Spark now available for $3,999, but real impact will be AI at the edge · Constellation Research
- 7.Nvidia DGX Spark gets $700 price hike as memory shortages bite · Tom's Hardware
Reported by the WattsUpNext desk from the sources linked below. Spot an error? Tell us at corrections@wattsupnext.com.
The WattsUpNext Brief
Get stories like this in your inbox.
One email each weekday, only the topics you choose. Real news, sourced, no fluff. Unsubscribe in one click.
