NVIDIA DGX Spark drops to 64GB memory but costs more than the original 128GB version - videocardz.com

This is not a downgrade — it is a generational pivot. The 64GB configuration almost certainly ships with HBM3E, delivering materially higher bandwidth than the prior 128GB HBM3 setup. NVIDIA is now selling bandwidth-per-dollar, not raw capacity. Downstream, CUDA and TensorRT stacks must re-architect around tighter capacity with fatter pipes, accelerating the industry shift toward 4-bit quantization and aggressive KV-cache compression. Upstream, SK Hynix and Samsung HBM3E yield rates — not NVIDIA's CoWoS packaging — are the true production constraint. On compliance, Washington's export controls have already carved out NVIDIA's China-facing SKU lineup. The less-memory-higher-price move is margin extraction from the unrestricted market, building a financial buffer ahead of potential further restrictions. Competitively, AMD's MI300X and Intel's Gaudi 3 remain locked in the 128GB-plus HBM3 tier. NVIDIA's strategy creates an asymmetric spec-sheet advantage: rivals can match capacity but not bandwidth, making head-to-head comparisons structurally unfavorable. Over the next 12 to 24 months, memory bandwidth will supplant memory capacity as the primary AI hardware metric. The memory wall will trigger a new arms race in CXL-attached memory and near-memory compute architectures.