HP ZGX Fury Packs 748GB RAM And NVIDIA GB300 Starting At $132,000 - HotHardware

HP's decision to pair 748GB of memory with the GB300 at a $132K entry price is a strategic bet that large-model inference is migrating from the cloud to on-premises enterprise deployments. This is not incremental iteration—it is NVIDIA repositioning the GPU as a system-level memory controller, with HBM4 stacks directly challenging the traditional CPU+DRAM paradigm. CoWoS-L advanced packaging capacity becomes the new global chokepoint, elevating the strategic weight of Taiwan, China's foundry and packaging lines. On compliance, this pricing exposes the market bifurcation created by export controls: restricted markets are locked to prior-generation silicon while compliant markets absorb a "compliance premium." HBM4 supply—split among SK Hynix, Samsung, and Micron—means any single-node disruption cascades into end-delivery timelines. Competitively, AMD's MI400 series will likely match specs by Q3, but CUDA migration costs remain a near-term moat. The structural threat, however, comes from hyperscaler ASICs: when Trainium and TPU v6 compress inference unit costs by 30%+, a $132K local solution faces an existential "build vs. rent" economics test. Over the next 18 months, the AI workstation battleground shifts from peak FLOPS to "effective memory bandwidth per dollar." Once 70B-parameter models run in real time on a single box, data sovereignty and latency-sensitive workloads will fundamentally rewire enterprise IT procurement. That is where HP's real endgame lies.