Intel / post

GB300 NVL72 claimed at 7x perf-per-dollar over H200 for agentic inference

Source: x.com
Date: Published Aug 28, 2026

Context

SemiAnalysis's sevenfold performance-per-dollar claim concerns a particular comparison of complete model-serving systems. Its chart compares NVIDIA GB300 NVL72 with H200 serving DeepSeek V4 Pro on long- agentic requests, using different numerical precision settings and parallel configurations. The vertical axis counts total processed per dollar under estimated hyperscaler ownership costs; the horizontal axis measures per-user generation speed. The marked gap sits around 60 output tokens per second per user. The ratio therefore depends on the model, serving configuration, required responsiveness and cost assumptions shown in the comparison.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source x.com