NVIDIA's $4,999 DGX Spark 64GB ships October 23 with half the memory amid a RAM crunch

NVIDIA is adding a lower-priced 64GB configuration of its DGX Spark desktop AI system. It keeps the GB10 Grace Blackwell Superchip and 1 petaFLOP of compute, and can run models up to roughly 100B parameters locally. It ships October 23 through OEMs including Acer, Dell, ASUS and HP at $4,999. The 128GB model now costs $6,950.

Linking two units yields 1.7x the performance of one. A Sync Model Launcher for setting up clusters arrives in late October. NVIDIA pitches the box for local agents, fine-tuning and inference, and it offers Meta's Muse Glimmer 30B as a NIM microservice. Community setup guides show Glimmer running on 24GB of VRAM with GGUF quantizations. Full BF16 needs 55GB or more, which is pushing interest toward the 64GB Spark.

The pricing has become the story. The Register reports that the 64GB unit costs $1,000 more than the 128GB model did at launch, with half the RAM and storage. Users blame a memory supply crunch they call the 'RAMpocalypse.' Hardware tension is visible elsewhere too: Micro Center reportedly now requires no-export paperwork for RTX 5090 purchases.

For buyers, the competition is Apple's high-memory Macs and cloud GPU rental. Some practitioners admit that 'self-hosting AI does not save money' but do it anyway for control. Watch real-world throughput, OEM availability on the 23rd, and whether memory prices push NVIDIA to reprice again.