Intel / post

AA-AgentPerf Adds Market-Based GPU Pricing

Source: x.com
Date: Published Nov 2, 2026

Context

Artificial Analysis has added a cost measure to AA-AgentPerf, its of how much work datacenter hardware can serve. The new metric, Throughput per $/hr, takes the total throughput a system reaches with a given serving setup and divides it by the hourly cost of renting that system. The result is how much agent work a dollar of rented compute buys, which the firm says is what its users and partners want to know (announcement).

The rental prices come from Silicon Data, whose price indices track what cloud providers charge to rent GPUs. Artificial Analysis says the metric updates daily as those prices change, and it currently shows indices for NVIDIA's H100, H200, B200 and B300. Because the price side follows the market, a system's value for money can rise or fall without a new benchmark run. Earlier AgentPerf results ranked systems by throughput per megawatt of power, as in the DeepSeek V4.1 Flash results from October. That measures energy efficiency, while the new figure measures cost to someone renting the hardware.

In the first results, among systems with rental pricing available, NVIDIA's B300 generally comes out ahead, and the B200 matches it on some models when each user is served at faster output speeds. The accompanying chart also includes AMD's MI355X, but Artificial Analysis says it assumed that chip's rental rate from recent market quotes. Its place in the ranking therefore rests on an estimated price, not on a tracked Silicon Data index like the NVIDIA systems.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source x.com