Nvidia's DGX Spark gets a 64GB model at $4,999, and two can pool to 128GB - MIXED Reality News

Nvidia's 64GB desktop unit at $4,999 is not a product refresh—it is a reallocation of inference sovereignty. Pooling two units to 128GB shatters the local-inference threshold for 70B-parameter models, directly eroding the moat cloud providers built on "inference-as-a-service." The supply-chain ripple is already visible: HBM procurement logic shifts from datacenter exclusivity toward edge penetration, forcing SK Hynix and Micron to recalibrate capacity allocation. TSMC's (Taiwan, China) advanced-node yield pressure intensifies as desktop GPU volume dwarfs datacenter card shipments. On compliance, a single unit approaching node-level compute challenges the classification boundaries of current export controls. Washington draws lines by "cluster compute," yet desktop pooling sits squarely in the gray zone—expect this to anchor the next BIS rule revision cycle. Competitively, AMD's Instinct desktop path and Apple's M4 Ultra local-inference roadmap will be forced to respond within 12 months. But the structural variable is pricing: when inference cost shifts from "per-token billing" to "per-watt depreciation," the entire AI application-layer economics gets rewritten. Within 24 months, cloud inference becomes a training-only channel; edge inference becomes the default architecture. This is not a trend. It is a phase transition already in motion.