DRAM has done what GPU yields never managed: it inverted the price ladder of the personal-AI tier. Nvidia this morning introduced a 64GB version of the DGX Spark — half the memory of the original GB10 box — starting at $4,999 from Acer, ASUS, Dell, Gigabyte, HP and MSI on October 23. The 128GB system launched at $3,999 and had its MSRP raised to $4,699 in February. The new “budget” SKU undercuts nothing: less memory now lists for more than the full-fat machine ever did.
DRAM reset the floor under local inference
What happened. Nvidia’s blog runs the 64GB config today, dated October 2: same GB10 Grace Blackwell Superchip, DGX OS and full NVIDIA AI software stack as the 128GB model, up to 100-billion-parameter models on device, from $4,999 starting Friday October 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI. Every unit keeps the built-in ConnectX-7 NIC; two can pool to 128GB via the new Sync Cluster Assistant, and Nvidia’s Qwen 3.8 27B test shows 1.7x throughput over a single unit. Street pricing on what 128GB inventory remains is running $7,000–$9,000, well past even the raised Founders MSRP.
Why it matters. This is the price ladder collapsing under memory supply, not a product refresh. Half the RAM, yet the SKU lists $300 above today’s 128GB Founders MSRP ($4,699) and $1,000 above the $3,999 launch price — the cheaper local box now costs more than the flagship did. For an operator the “thin local tier” assumption is gone: buying less memory barely bends the bill, so the rational move is one box clustered against a second rather than sizing up, and anything that won’t fit 100B resident leans back to the cloud. Local AI is not getting cheaper; its cost floor just moved up on DRAM’s account.
Source: blogs.nvidia.com
What I’m watching
Watch whether the 64GB SKU holds at $4,999 through October 23 or drifts up between announcement and availability the way the rest of the line just did — memory quotes, not Nvidia, have been writing these price tags for two quarters. And Sync Cluster Assistant is the quieter tell: Nvidia is selling the cluster as the upgrade path instead of more RAM, which is how a hardware vendor signals it can’t source the LPDDR5X either.