New hardware built around NVIDIA’s Blackwell architecture took center stage at the SIGGRAPH 2025 conference, where Dell Technologies introduced the PowerEdge R7725 server equipped with RTX PRO 6000 Blackwell GPUs for enterprise-scale artificial-intelligence workloads. Dell also displayed a 70-pound ProMax desktop workstation that pairs a Grace-Blackwell 300 GPU-CPU module and carries a list price of roughly $65,000, targeting researchers training models with as many as one trillion parameters. NVIDIA, meanwhile, highlighted its GB200 NVL72 system, which binds 72 Blackwell GPUs to 36 Grace CPUs via NVLink fabric. The platform, already in use by OpenAI, serves GPT-5 after the large language model was trained on earlier-generation H100 and H200 processors. Cumulative downloads of the company’s CUDA software toolkit have passed 450 million, underscoring the ecosystem’s scale. Cloud providers are moving quickly to accommodate the new chips. Amazon Web Services said its SageMaker HyperPod service now supports P6e-GB200 UltraServers, letting customers train and deploy trillion-parameter models on clusters of 72 Blackwell GPUs per node. The shift toward larger, more autonomous models is also reshaping back-end infrastructure. At Supermicro’s Open Storage Summit, executives from Supermicro, AMD, DDN and Western Digital outlined multi-site, low-latency storage designs intended to keep high-cost GPUs fully fed while controlling power and data-movement overheads.