--- title: Why Performance per Watt Is the Ultimate Metric for AI... description: Power constraints are the defining limit for AI infrastructure, making performance per watt the key metric for AI factories. NVIDIA's Blackwell NVL72 platform... canonical: https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency | daily.dev og:description: Power constraints are the defining limit for AI infrastructure, making performance per watt the key metric for AI factories. NVIDIA's Blackwell NVL72 platform... og:url: https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ og:image: https://api.daily.dev/og/posts/c9tzt9typ.png og:image:alt: Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency **[NVIDIA](https://daily.dev/sources/nvidia)** · 4 min read · 1 upvotes · 0 comments ## Summary Power constraints are the defining limit for AI infrastructure, making performance per watt the key metric for AI factories. NVIDIA's Blackwell NVL72 platform claims up to 25x performance per watt improvement over the Hopper generation through rack-scale codesign spanning silicon, NVLink switching, and inference software (Dynamo, TensorRT-LLM, SGLang, vLLM). Techniques like NVFP4 quantization, disaggregated serving, KV-aware routing, and the DSX MaxLPS power management software allow operators to run up to 40% more GPUs within the same power budget. Real-world deployments at Anthropic, OpenAI, Perplexity, CoreWeave, and Fireworks AI validate the platform's production reliability. The upcoming Vera Rubin platform is positioned to build further on this foundation. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much better is NVIDIA GB300 NVL72 in performance per watt compared to the Hopper generation? GB300 NVL72 delivers up to 25x higher performance per watt than the Hopper generation across the newest generation of leading open models. This figure reflects NVIDIA's current Blackwell benchmarking and is described as a starting point that continues to improve through ongoing software optimizations like NVFP4 quantization and disaggregated serving. _Engineers weighing GPU generation upgrades can track efficiency benchmarks like this on daily.dev._ ### How much more electricity actually turns into useful AI compute in a typical AI factory versus being lost to cooling and rack inefficiencies? Only about 60% of electricity pulled from the grid in AI factories converts into useful AI work, with the rest lost to cooling and rack-level inefficiencies. NVIDIA's DSX MaxLPS software addresses this by shifting power between GPUs and racks in real time and using power-steering techniques, letting operators run up to 40% more GPUs within the same power budget. _Teams sizing AI data center power budgets can follow efficiency-tooling news like DSX MaxLPS on daily.dev._ ### Which AI companies are running production inference on NVIDIA Blackwell NVL72 systems? Anthropic and OpenAI run inference on NVIDIA Blackwell NVL72 systems. CoreWeave deploys Kimi K2.6 on GB300 NVL72 with NVFP4 quantization and EAGLE3 speculative decoding, Perplexity serves Qwen3 235B and a post-trained Qwen3.5-397B-A17B on GB200 NVL72, and Fireworks AI deploys GLM 5.2 for customers including Cursor and Factory AI. _Developers choosing an inference platform can compare real production deployments like these on daily.dev._ ## Similar posts on daily.dev - [Scaling Token Factory Revenue and AI Efficiency by Maximizing Performance per Watt](https://daily.dev/posts/scaling-token-factory-revenue-and-ai-efficiency-by-maximizing-performance-per-watt-2qzgn7uql) · NVIDIA Developer · 0 upvotes · 0 comments - [Understanding the economics of AI factories](https://daily.dev/posts/understanding-the-economics-of-ai-factories-5c6ecwvmj) · InfoWorld · 0 upvotes · 0 comments - [Maximize AI Factory Energy Efficiency Through Full-Stack Inference and Training Optimizations](https://daily.dev/posts/maximize-ai-factory-energy-efficiency-through-full-stack-inference-and-training-optimizations-lwusm4jtd) · NVIDIA Developer · 0 upvotes · 0 comments - [The infrastructure lock-in costing AI companies hundreds of millions](https://daily.dev/posts/the-infrastructure-lock-in-costing-ai-companies-hundreds-of-millions-wuyswiobz) · The New Stack · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#nvidia](https://daily.dev/tags/nvidia), [#ai-infrastructure](https://daily.dev/tags/ai-infrastructure), [#ai-inference](https://daily.dev/tags/ai-inference) [View this post on daily.dev](https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency","url":"https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ"},"datePublished":"2026-07-14T16:46:24.280Z","dateModified":"2026-09-13T19:39:32.269Z","description":"Power constraints are the defining limit for AI infrastructure, making performance per watt the key metric for AI factories. NVIDIA's Blackwell NVL72 platform...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/467ef464df09131c7c4ddc03afda6fb1?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/467ef464df09131c7c4ddc03afda6fb1?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"NVIDIA","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"NVIDIA","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/236a5c1f410e450b8d696cf3a4bbe374","url":"https://daily.dev/sources/nvidia"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,nvidia,ai-infrastructure,ai-inference","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"NVIDIA","item":"https://daily.dev/sources/nvidia"},{"@type":"ListItem","position":3,"name":"Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/why-performance-per-watt-is-the-ultimate-metric-for-ai-infrastructure-efficiency-c9tzt9typ#faq","mainEntity":[{"@type":"Question","name":"How much better is NVIDIA GB300 NVL72 in performance per watt compared to the Hopper generation?","acceptedAnswer":{"@type":"Answer","text":"GB300 NVL72 delivers up to 25x higher performance per watt than the Hopper generation across the newest generation of leading open models. This figure reflects NVIDIA's current Blackwell benchmarking and is described as a starting point that continues to improve through ongoing software optimizations like NVFP4 quantization and disaggregated serving. Engineers weighing GPU generation upgrades can track efficiency benchmarks like this on daily.dev."}},{"@type":"Question","name":"How much more electricity actually turns into useful AI compute in a typical AI factory versus being lost to cooling and rack inefficiencies?","acceptedAnswer":{"@type":"Answer","text":"Only about 60% of electricity pulled from the grid in AI factories converts into useful AI work, with the rest lost to cooling and rack-level inefficiencies. NVIDIA's DSX MaxLPS software addresses this by shifting power between GPUs and racks in real time and using power-steering techniques, letting operators run up to 40% more GPUs within the same power budget. Teams sizing AI data center power budgets can follow efficiency-tooling news like DSX MaxLPS on daily.dev."}},{"@type":"Question","name":"Which AI companies are running production inference on NVIDIA Blackwell NVL72 systems?","acceptedAnswer":{"@type":"Answer","text":"Anthropic and OpenAI run inference on NVIDIA Blackwell NVL72 systems. CoreWeave deploys Kimi K2.6 on GB300 NVL72 with NVFP4 quantization and EAGLE3 speculative decoding, Perplexity serves Qwen3 235B and a post-trained Qwen3.5-397B-A17B on GB200 NVL72, and Fireworks AI deploys GLM 5.2 for customers including Cursor and Factory AI. Developers choosing an inference platform can compare real production deployments like these on daily.dev."}}]} ```