Nvidia Rubin with vLLM: 3.2x More Profit per Gigawatt
SemiAnalysis models Nvidia Rubin running vLLM with 10x performance per dollar and 3.2x profit per gigawatt over GB300 in agentic inference.
Preliminary modeling from SemiAnalysis shows big gains in agentic inference and FP4 precision, though metrics still await production silicon.
Preliminary benchmark data released by research firm SemiAnalysis indicates that Nvidia Rubin racks running the open-source vLLM inference engine can achieve up to 10x better performance per dollar and 3.2x higher profit per gigawatt than the Blackwell GB300 NVL72. The figures, published on October 9, 2026, come from an early preview submission within the firm’s InferenceX framework.
Economic efficiency driven by data center power caps
Profit per gigawatt has become the defining financial metric across large artificial intelligence facilities. Power grid interconnect delays and substation limits now constrain data center expansion, making revenue per megawatt a stricter bottleneck than server hardware pricing. In the SemiAnalysis model, the computational density of 72 Rubin chips connected via NVLink 6 offsets the higher upfront capital expenditure of the rack.
That trade-off appears clearly in total cost of ownership (TCO) comparisons. Running a Rubin GPU costs roughly $3.57 per hour for neocloud operators, compared to $2.36 on a GB300 and $1.84 on a GB200. Despite higher hourly operating expenses, throughput gains from native FP4 precision and expanded memory bandwidth let the cluster process substantially more tokens per cycle. During continuous agentic reasoning tasks, that extra throughput outpaces the higher hourly compute cost.
Open-source software optimization with vLLM
The benchmark gains relied on aggressive software integration upstream. According to SemiAnalysis, Nvidia engineers worked alongside vLLM maintainers, startup Inferact, and former CentML engineers now at Nvidia to merge dedicated kernels and support for the SM_107 compute architecture directly into the engine’s main branch.
The test used MiniMax M3, an open-weights mixture-of-experts (MoE) model with 428 billion total parameters designed for agentic workflows and multi-step reasoning. SemiAnalysis noted that a fully audited public benchmark covering both vLLM and SGLang on Rubin hardware will follow in the coming months.
Nvidia Rubin projections and early bringup caveats
Using its Agentic Inference Profit Calculator, SemiAnalysis modeled that a hypothetical 1-gigawatt cluster of Vera Rubin NVL72 racks could generate $39 billion in annualized operating profit serving MiniMax M3. That estimate assumes steady server utilization and consistent market pricing per million tokens.
SemiAnalysis explicitly noted the caveats in these numbers. Current figures stem from early silicon bringup tests and economic extrapolations rather than production workloads running on final shipping hardware. Real-world thermal constraints in dense facilities and competitive pressures in cloud pricing will test these theoretical margins once Rubin racks enter commercial deployment.