NVIDIA DGX Spark 64 GB Reduces Memory to Stem RAM Price Increases - HotHardware

NVIDIA trimming DGX Spark to 64GB is not a product tweak—it is the first visible spec-for-margin concession in the AI compute chain. Rising DRAM spot prices have cracked the BOM structure of mid-tier inference cards, and NVIDIA's choice to cut capacity rather than raise price signals a clear read: edge-inference buyers are far more price-sensitive than spec-sensitive. The technical ripple runs deeper than the headline. 64GB forces 70B-parameter models into 4-bit quantization and aggressive KV-cache compression, compelling a rewrite of paging logic in vLLM and TensorRT-LLM. Upstream, HBM capacity is locked into training accelerators, leaving DDR5 supply for inference already tight. The cut is effectively "use fewer chips" to offset "each chip costs more"—a reallocation of supply-chain bargaining power. Competitively, AMD's MI300 and Intel's Gaudi 3, with different packaging paths, face less DRAM cost pressure, opening a short-term window in mid-range inference. Google's TPU and AWS Trainium, vertically integrated, sidestep the DRAM spot market entirely. NVIDIA's concession exposes the structural fragility of the fabless model when storage costs swing. Within 18 months, expect a memory-efficiency arms race: sparse computation and on-chip SRAM expansion will supplant brute-force HBM stacking as the core competitive axis. This cut draws a line: AI's bottleneck is shifting from raw FLOPS to the economics of bandwidth and capacity.