NVIDIA's memory crunch has finally forced its hand. The company is rolling out a 64GB DGX Spark at $4,999 this month while the original 128GB model climbs past $6,000—nearly 75% above its $3,999 launch price from October 2025. The half-capacity unit ships October 23 through partners like Acer, ASUS, Dell, Gigabyte, HP, and MSI, keeping the same GB10 Grace Blackwell Superchip and clustering support but cutting unified LPDDR5X memory in half.

The move comes straight from skyrocketing DRAM costs that have already pushed street prices on the 128GB version to $4,999+ at retailers like Micro Center and Amazon, with some partner configs hitting $6,000–$8,000. NVIDIA explicitly blamed global memory supply constraints when it raised the 128GB MSRP to $4,699 earlier this year and is now hiking it further to $6,950. Clustering still works—the new 64GB unit pairs with others via the built-in ConnectX-7 200GbE NIC—but two units are now required to approach the original 128GB capacity.

Community reaction on X and hardware forums has been predictably sour. Users are calling the de-spec'd model a blatant tiering play that hands memory makers like Micron even more pricing power, especially as Tesla reportedly trims AI5 memory demands and supply forecasts stay tight into 2028. One trader noted real-world server DDR5 128GB kits already cost more than a full Spark, meaning buyers are effectively paying for RAM and getting the superchip for free. AMD's competing Ryzen AI Halo systems at $3,500–$4,500 with higher memory configs are looking like the smarter buy for many developers.

The 64GB model still supports up to 100B-parameter local inference out of the box with DGX OS and the full CUDA stack, but the writing is on the wall: unified memory has become the new bottleneck, and NVIDIA is adapting by giving enthusiasts less for more.