NVIDIA is transforming the AI infrastructure market by dramatically increasing agentic AI inference performance and cost efficiency with its next-generation 'NVIDIA Blackwell Ultra' platform.

NVIDIA announced on the 23rd that the Blackwell Ultra-based GB300 NVL72 system improves throughput per megawatt by up to 50x and reduces the cost per million tokens by up to 35x compared to the existing Hopper platform.

Agentic AI is a next-generation AI application that requires multi-step workflows and real-time interactions, where low latency and long-term context processing capabilities are key.

With the recent surge in demand for AI coding agents and conversational coding assistants, the importance of infrastructure that can process large codebases quickly and efficiently is growing.

According to Semianalysis InferenceX, NVIDIA's hardware and software co-design strategy offers a clear advantage in terms of inference performance and total cost of ownership (TCO).

The software stack, consisting of Blackwell Ultra GPUs, NVIDIA Dynamo, and TensorRT-LLM, maximizes token processing efficiency in low-latency environments and is suitable for agentic AI workloads. Provides optimized performance.

The GB300 NVL72, in particular, demonstrates strength in long-context workloads. In an AI coding assistant environment processing 128,000 token inputs and 8,000 token outputs, it can reduce cost per token by up to 1.5x compared to the previous generation. This is a result of improved NVFP4 compute performance and improved attention processing speed.

Major cloud service providers, including Microsoft, CoreWeave, and OCI, are already deploying the GB300 NVL72 at scale to power agentic coding and low-latency AI services. NVIDIA plans to continue its infrastructure innovation across AI inference and training, promising another performance leap with its upcoming Rubin platform.

Blackwell Ultra is positioned as a core infrastructure for the era of agentic AI, simultaneously enhancing the scalability and economic viability of AI services.