Blackwell resets inference bar

NVIDIA’s RTX PRO 6000 Blackwell Server Edition ships with 96 GB of GDDR7 ECC memory, a reported memory bandwidth around 1.79 TB/s, 24,064 CUDA cores and a configurable board power up to roughly 600 W. (provantage.com) Reseller and retail listings have shown street prices clustering between roughly $8,000 and $11,000 per card versus earlier advertised figures near $10,999, with multiple authorized resellers listing sub‑$9k offers. (techradar.com) CoreWeave was an early cloud adopter and made RTX PRO 6000 instances generally available, listing an eight‑GPU RTX PRO 6000 instance at about $20 per hour on its on‑demand pricing page. (coreweave.com) Independent single‑GPU LLM inference benchmarks published by a third‑party tester reported the PRO 6000 delivering roughly 28% lower cost‑per‑token on single‑GPU workloads versus H100 PCIe, while noting NVLink/SXM datacenter GPUs still maintain a multi‑GPU throughput advantage for very large models. (cloudrift.ai) OEMs and integrators are shipping RTX PRO 6000 systems at scale: HP’s new Z8 Fury and ZBook configurations support up to four RTX PRO 6000s, Cisco lists UCS RTX PRO server SKUs with both dense PCIe and NVLink‑enabled deployment paths, and Supermicro/ Pegatron are offering NVL/HGX Blackwell rack solutions. (seekingalpha.com) High‑density server tests and reference builds show eight RTX PRO 6000 cards can provide 768 GB of aggregate GPU RAM in a single 4U/5U enclosure, and running eight 600 W cards pushes platform electrical draw toward the 10 kW range that datacenter operators must plan for. (servethehome.com) Suppliers of thermal systems are shipping complementary solutions for that density profile today, including ZutaCore’s waterless two‑phase direct‑to‑chip cooling announced for Blackwell PCIe GPU servers to raise slot density without facility water loops. (comparethecloud.net)