NVIDIA Blackwell Ultra Achieves Up to 50x Performance Boost & 35x Cost Reduction for Agentic AI
Summarize this article with:
NVIDIA has announced that its Blackwell Ultra platform delivers up to 50 times better performance and a 35 times reduction in costs for agentic AI. Cloud providers like Microsoft, CoreWeave, and Oracle Cloud Infrastructure are already deploying NVIDIA GB300 NVL72 systems to support demanding applications requiring low latency and extensive context, such as agentic coding and AI coding assistants. This leap in capability comes as AI agents drive explosive growth in software-programming queries, jumping from 11% to approximately 50% last year. According to data from SemiAnalysis InferenceX, the combination of NVIDIA’s software and the Blackwell Ultra platform has achieved breakthrough advances, with GB300 NVL72 systems delivering “up to 50x higher throughput per megawatt, resulting in 35x lower cost per token compared with the NVIDIA Hopper platform.” This innovation promises to significantly scale real-time interactive AI experiences for a wider range of users. NVIDIA Blackwell Ultra Achieves 50x Performance Gains A staggering 50-fold increase in throughput per megawatt is now achievable with the NVIDIA Blackwell Ultra platform, according to recent SemiAnalysis InferenceX data, dramatically reshaping the economics of AI inference. This leap forward isn’t simply about speed; it directly translates to a 35-fold reduction in cost per token compared to the preceding NVIDIA Hopper platform, a critical factor as AI agents become increasingly prevalent. The improvements stem from a holistic approach to innovation, encompassing advancements in chip design, system architecture, and crucially, software optimization. Continuous refinements to NVIDIA’s TensorRT-LLM library, Dynamo, Mooncake, and SGLang teams have yielded substantial throughput boosts for mixture-of-experts (MoE) inference, with TensorRT-LLM alone delivering up to 5x better performance on GB200 for low-latency workloads in just four months. “As inference moves to the center of AI production, long-context performance and token efficiency become critical,” said Chen Goldberg, senior vice president of engineering at CoreWeave. The GB300 NVL72 also demonstrates a 1.5x lower cost per token for long-context workloads—such as AI coding assistants reasoning across extensive codebases—compared to the GB200 NVL72. This efficiency is powered by 1.5x higher NVFP4 compute performance and 2x faster attention processing within Blackwell Ultra. GB300 NVL72 Lowers Costs for Agentic AI The demand for increasingly sophisticated AI agents is reshaping the landscape of cloud computing, with a particular emphasis on reducing both latency and cost. This leap in efficiency directly translates to economic benefits, achieving 35 times lower cost per token. This improvement isn’t solely attributable to hardware; NVIDIA’s software optimizations, including TensorRT-LLM, Dynamo, Mooncake, and SGLang, are demonstrably boosting throughput for mixture-of-experts (MoE) inference. Signal65 analysis indicates that the GB200 NVL72, with its combined hardware and software codesign, delivers over ten times more tokens per watt than the Hopper platform. CoreWeave’s AI cloud is designed to translate the gains of GB300 systems into predictable performance and cost efficiency. NVIDIA GB300 NVL72 systems now deliver up to 50x higher throughput per megawatt, resulting in 35x lower cost per token compared with the NVIDIA Hopper platform.
Software Optimizations Boost Blackwell Throughput The GB300 NVL72’s advantages are particularly pronounced in long-context applications, such as AI coding assistants analyzing extensive codebases. The combination of software optimization and next-generation hardware is enabling AI platforms to scale real-time interactive experiences to a significantly larger user base.
Rubin Platform Promises 10x Higher Inference Throughput Beyond the current gains seen with Blackwell Ultra, the company anticipates a further leap in performance, promising up to 10 times higher throughput per megawatt for mixture-of-experts (MoE) inference. This translates directly to economic benefits, potentially reducing the cost per million tokens to one-tenth of what Blackwell currently achieves. The Rubin platform achieves this through a radically integrated design, combining six new chips into a single AI supercomputer. This holistic approach isn’t just about raw power; it’s about efficiency, enabling the training of large MoE models with just one-fourth the number of GPUs required by the Blackwell architecture. CoreWeave is specifically designing its AI cloud, including CKS and SUNK, to leverage the gains offered by the GB300 systems, building on the success of the GB200.
The Vera Rubin NVL72 system, built on the Rubin platform, is poised to deliver the next wave of performance improvements through continuous software optimizations. Source: https://blogs.nvidia.com/blog/data-blackwell-ultra-performance-lower-cost-agentic-ai/ Tags: The Neuron With a keen intuition for emerging technologies, The Neuron brings over 5 years of deep expertise to the AI conversation. Coming from roots in software engineering, they've witnessed firsthand the transformation from traditional computing paradigms to today's ML-powered landscape. Their hands-on experience implementing neural networks and deep learning systems for Fortune 500 companies has provided unique insights that few tech writers possess. From developing recommendation engines that drive billions in revenue to optimizing computer vision systems for manufacturing giants, The Neuron doesn't just write about machine learning—they've shaped its real-world applications across industries. Having built real systems that are used across the globe by millions of users, that deep technological bases helps me write about the technologies of the future and current. Whether that is AI or Quantum Computing. Latest Posts by The Neuron: Ant Group’s Ring-1T-2.5 1 Trillion Parameter Model Achieves Gold-Tier Performance on IMO 2025 & CMO 2025 Benchmarks February 13, 2026 Experian Launches ChatGPT App for Auto Insurance, Potential Savings Over $1,000 Annually February 12, 2026 Quantum Entanglement Boosts Teamwork Between Artificial Intelligence Agents February 11, 2026