# SemiAnalysis Benchmark of Real Coding-Agent Traffic Puts Nvidia About Five Times Ahead of AMD on Cost

The gap comes mostly from serving software, and the analysts say Nvidia would still be cheaper per token even if the competing hardware were free.

- Published: 2026-08-25T06:26:21.272Z
- Canonical: https://polylog.news/ai/2026-08-25/semianalysis-benchmark-of-real-coding-agent-traffic-puts-nvi
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [NVIDIA Blog (NVLink Fusion and XPUs)](https://blogs.nvidia.com/blog/nvlink-fusion-xpu-ai-factory/)

SemiAnalysis published AgentX, an open-source benchmark that replays recorded coding-agent sessions instead of synthetic prompts, and the result is unfavorable for AMD. At a target interactivity of 150 output tokens per second per user, the analysts report that Nvidia hardware reaches [up to five times better cost efficiency](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) than AMD when serving the GLM 5.3 model through the open-source SGLang runtime. [Forbes summarized](https://www.forbes.com/sites/janakirammsv/2026/08/24/one-agent-benchmark-puts-nvidia-5x-ahead-of-amd-on-cost/) the same finding, noting cost advantages as wide as 20 times on certain open-source software stacks.

The most significant finding in the analysis is not the multiple itself. SemiAnalysis argues that at 150 tokens per second per user, the gap is large enough that cost per token would still favor Nvidia even if a competitor gave its accelerators away for free, once hosting and power costs are counted. That is a claim about software quality, not about the chips themselves. The analysts attribute Nvidia's lead to cache management, request routing, and incremental tokenization, factors that matter far more under agent traffic than under single-turn chat.

The shape of the workload explains why. AgentX uses anonymized [Claude Code](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) sessions with a median input of roughly 142,000 tokens against a median output of 444 tokens, plus long idle gaps between turns. Under that profile, any stack that recomputes context aggressively or schedules requests poorly performs worse, while stacks that reuse prefixes and chunks perform better.

Two caveats limit how broadly this number applies. The comparison sets Nvidia's current Blackwell generation against AMD's current CDNA 4 chips, while Nvidia's upcoming Rubin generation, Google's tensor processing units, and AMD's upcoming MI455X chip are not yet included. AgentX also measures one category of work, coding agents, which happens to be the workload Nvidia's serving software has been tuned hardest against. A [second AI Post summary](https://t.me/aipost/7943) of the report repeated the five-times figure without those qualifications, which shows how a benchmark result can turn into an unqualified market narrative.

## What this means

The competitive question for AI accelerators has shifted from peak throughput to how well the serving software handles long, repetitive, bursty agent workloads. That favors Nvidia's CUDA ecosystem, where kernel and runtime engineering has accumulated advantages for years, and it puts pressure on AMD to close a software gap that new hardware alone will not close. Enterprises evaluating alternative accelerators now have a concrete public methodology to test claims themselves, which requires stronger evidence from vendors on both sides. If AMD's ROCm software improves measurably on AgentX before the MI455X ships, pricing pressure on Nvidia's inference capacity will return. If it does not, Nvidia keeps its pricing power in the fastest-growing segment of inference demand.

## What to watch

- Whether AMD publishes its own AgentX runs with updated software, which would show whether the deficit is fixable in the runtime or is structural to the memory system.
- Whether cloud providers start quoting agent-workload cost per token instead of hourly GPU rates, which would move the whole market onto SemiAnalysis-style measurement.
- How the benchmark behaves once Rubin and MI455X parts are both available, since a like-for-like generational comparison is the only way to separate silicon from software.