Analysis
Nvidia is launching a 64GB version of its DGX Spark personal AI supercomputer at $4,999, half the memory of the original 128GB unit, through OEM partners including Acer, Asus, Dell, Gigabyte, HP and MSI starting October 23, The Register reported. Both configurations use the same GB10 Grace Blackwell Superchip, software stack and ConnectX-7 networking; Nvidia says the 64GB model can still run AI models with up to 100 billion parameters locally, depending on the workload.
A Price Cut That's Actually a Memory Story
The original DGX Spark launched in October 2025 at a $3,999 MSRP for its 128GB configuration. Skyrocketing memory prices pushed that same configuration to $4,699 in February 2026 and then to $6,950 now -- nearly 75% higher than a year ago, Tom's Hardware reported. Rather than absorb that cost increase across its whole lineup, Nvidia introduced the 64GB variant as a new, lower-memory entry point priced close to where the 128GB unit used to sit.
“## A Price Cut That's Actually a Memory Story The original DGX Spark launched in October 2025 at a $3,999 MSRP for its 128GB configuration.”
Why This Matters Beyond One Product
DGX Spark targets researchers, startups and developers who want to prototype and run inference on large models locally rather than renting cloud GPU capacity -- the same buyers increasingly squeezed by GPU-cloud pricing and allocation constraints that have driven neocloud providers like Lambda to raise billions in GPU-backed debt this year. A cheaper local option, even with less memory, lowers the barrier for smaller teams to own rather than rent compute for development work, even as the broader memory shortage raises costs everywhere else in the stack.
The Numbers in Context
The nearly 75% price increase on the 128GB Spark in twelve months is a sharper move than typical hardware inflation, and it tracks with a broader DRAM and HBM shortage that's been pushing up costs across consumer and data-center hardware alike -- a dynamic separate from, but related to, the GPU scarcity driving neocloud debt financing. Nvidia launching a cheaper, lower-spec configuration rather than cutting the original price is itself a signal that memory costs, not demand for the product, are now the harder constraint to engineer around.
What the Headline Misses
A 64GB unit running 100-billion-parameter models "depending on the model and workload" is a meaningfully different product than the 128GB version for anyone doing serious local fine-tuning or running larger open-weight models, and Nvidia's October 23 launch date means the actual street price and availability through six OEM partners simultaneously still isn't tested in the market yet.