Blackwell Cloud Gains
CoreWeave posted benchmark gains showing NVIDIA GB200 and GB300 NVL72 systems leading DeepSeek‑R1 throughput in v6.0 tests—GB200 NVL72 topped server and offline modes measured in tokens/sec per GPU (investing.com). That matters for cloud AI inference and could also accelerate cloud‑streamed gaming and content workflows that rely on Blackwell tensor performance (investing.com).
CoreWeave published its MLPerf Inference v6.0 submissions on April 1, 2026, entering the Datacenter Closed division with NVIDIA GB200 NVL72 and GB300 NVL72 reference configurations. (coreweave.com) The company’s v6.0 runs explicitly covered two frontier reasoning models: DeepSeek‑R1 and the newly added open‑weight GPT‑OSS‑120B. (mlcommons.org) CoreWeave reported that its GB300 NVL72 portfolio produced roughly 2× the server throughput compared with its own MLPerf Inference v5.1 results on the same hardware footprint. (businesswire.com) MLCommons said MLPerf Inference v6.0 introduced five new or updated datacenter tests, including an expanded DeepSeek‑R1 interactive scenario and the GPT‑OSS‑120B benchmark used in these submissions. (mlcommons.org) NVIDIA’s technical blog credited co‑design and software optimizations across the Blackwell platform for up to 2.7× throughput improvements and more than a 60% reduction in cost per token in v6.0 submissions, and listed CoreWeave among the 14 partner submitters. (developer.nvidia.com) CoreWeave described the v6.0 results as “verified, production‑ready baselines,” noted that eight of the top 10 model providers run on CoreWeave Cloud, and identified the company as CoreWeave, Inc. of Livingston, N.J., trading on Nasdaq under CRWV. (coreweave.com)
