According to a report by SemiAnalysis published on October 19, 2026, Nvidia’s GB300 FP4 hardware trails Google’s TPUv7 in cost efficiency for agentic inference tasks. The analysis indicates that Google’s platform delivers between 1.42 and 1.67 times better performance per dollar compared to Nvidia’s solution.
The benchmark results remained consistent even after Nvidia attempted to improve its performance through engineering adjustments. In the current testing, Nvidia’s system lagged by a factor of 1.64 in performance per dollar. Notably, the Nvidia GB300 used the FP4 format, while the Google TPUv7 operated on the FP8 format.
The testing methodology highlighted significant differences in hardware requirements. Google achieved its results using a configuration of four chips, whereas Nvidia utilized a full setup of 72 GPUs. While Nvidia’s GB300 NVL72 maintains an advantage in low-interactivity scenarios, Google’s TPUv7 performs better in medium-to-high interactivity tasks.
Agentic AI workloads, which involve long-context processing and multi-turn interactions, favor the architecture of the TPUv7. Analysts are now monitoring future benchmark submissions from Nvidia to determine if the current performance gap is a temporary software issue or a structural advantage inherent to Google’s hardware design.
