NVIDIA official newsroom (nvidianews.nvidia.com/bios/jensen-huang)
SemiAnalysis says Nvidia's FP4 entry remains 1.64 times worse in performance per dollar than Google's TPUv7 on agentic inference
Nvidia has spent years as the undisputed heavyweight of AI hardware. On one closely watched inference benchmark, it is now the one chasing.
According to SemiAnalysis, Nvidia’s GB300 FP4 submission remains 1.64 times worse in performance per dollar than Google’s TPUv7. That gap persisted even after Nvidia provided feedback and put engineering work into improving its results.
The numbers behind the gap #
The SemiAnalysis report, published on October 19, 2026, focused on agentic inference benchmarks used for external customers.
On the performance per dollar measure, Google’s TPUv7 submissions came in between 1.42 and 1.67 times better than Nvidia’s GB300 FP4. Nvidia’s first response to the benchmark did not close that gap, leaving it 1.64 times behind.
The format detail makes the result more striking. Nvidia’s entry used FP4, while Google’s TPUv7 ran on FP8. FP4 uses four bits and FP8 uses eight bits to represent each number during computation, so FP4 is the leaner, more aggressive format designed to squeeze extra throughput out of the hardware. Nvidia brought the lighter-weight tool and still came out behind on cost efficiency.
On paper, the GB300 NVL72 delivers up to 15 PFLOPS of dense FP4 compute per GPU and 1.1 exaFLOPS per rack.
The TPUv7 needed only four chips for its benchmarks, while Nvidia’s comparisons used a full configuration of 72 GPUs.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Where each chip wins #
SemiAnalysis found that the GB300 NVL72 holds an advantage in low interactivity scenarios. TPUv7 pulls ahead in medium-to-high interactivity tasks, which are more common in agentic inference, where an AI system holds long conversations, juggles multiple steps and processes lengthy context. The disparity is especially pronounced for workloads involving long-context processing and multi-turn interaction.
Why the agentic angle matters #
Agentic AI refers to systems that plan, call tools, check their own work and keep going across many turns. Every one of those turns is another inference call, and a single user request can trigger a long chain of model runs, each dragging along a growing pile of context.
When SemiAnalysis measures a gap of 1.64 times on agentic workloads specifically, the number speaks directly to what external customers would pay to run these systems.
Background: a rivalry heating up #
Nvidia has already taken one swing at closing the gap. SemiAnalysis anticipates more counter-submissions from Nvidia as it tries to improve its standing on performance efficiency.
What this means for the AI hardware market #
If the TPUv7 delivers between 1.42 and 1.67 times better performance per dollar on agentic inference, cost-conscious enterprises have a concrete reason to look beyond Nvidia for those workloads. The chip-count comparison adds another layer. Achieving results with four chips rather than a 72-GPU setup points to a different approach to scaling, one that could appeal to customers who want to start small or run many independent workloads.
For investors following AI infrastructure, the practical takeaway is to watch the resubmissions. The next round of results will show whether this is a temporary software gap that Nvidia can tune away or a structural cost advantage for Google’s architecture on agentic inference. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our