NVIDIA DGX Spark gets 64GB version; Maintains cluster connect capability - Tech Critter

NVIDIA's 64GB DGX Spark with persistent cluster interconnect is not a memory bump — it is a category redefinition. Four to eight units form a 256-to-512GB compute pool at roughly one-tenth the power draw of an equivalent A100 node, collapsing the TCO curve for AI labs and startups. Upstream, high-bandwidth LPDDR5X and interconnect IP suppliers see accelerated demand. The regulatory minefield is the cluster link, not the silicon. US BIS rules target aggregate compute thresholds. A single Spark stays under the radar; eight linked units enter a gray zone. NVIDIA's "prosumer workstation" classification is a deliberate export-control positioning strategy. Beijing's parallel GPU import restrictions make this a potential circumvention vector — expect closer scrutiny from both Washington and Beijing. Competitively, AMD's MI300X and Intel's Gaudi 3 lack the "cluster-in-a-box" narrative. Apple's M3 Ultra unified memory is the closest analog but cannot overcome CUDA's gravitational pull. NVIDIA's moat is NVLink plus the software stack that makes eight boxes behave as one machine. Within 18 months, expect a "Spark Cluster" reference architecture targeting the $50K-to-$200K band currently occupied by mini-rack A100/H100 systems. Cloud providers will see mid-tier GPU instance demand erode. The structural shift: AI development migrates from "rent a cluster" to "own a cluster," fundamentally rebalancing the SaaS-versus-on-prem ratio in enterprise AI.