The Pulse

Bentaus says AI cluster cuts energy per token by 30%

Bentaus and DriveNets say their AI cluster reduced energy use per token by 30%, with a 3% drop in token throughput. The system combines AMD GPUs, DriveNets networking and Bentaus power-management software.

AI.info Team · · fact-checked · against 2 sources ↓ · Published 13 Nov 2026

Bentaus and DriveNets say their AI cluster cut energy use per token by 30% while giving up 3% of token throughput. The companies also say Bentaus reduced GPU power consumption by 30% while retaining 97% of throughput on the cluster, which runs AMD Instinct MI325X GPUs, DriveNets AI Fabric and Bentaus Ziani software.

Networking and power controls

Bentaus has deployed DriveNets AI Fabric as the back-end network for a new cluster using AMD Instinct MI325X GPUs. The cluster also works with GPU-FLX and CPU-FLX, products in the Bentaus Ziani Systems platform, which manages compute power in real time for enterprises, data centers and NeoClouds.

The Ethernet-based fabric is built on a reference architecture jointly engineered by DriveNets and AMD. It includes networking optimization from the switch and network interface card through RCCL and the kernel.

In published benchmarks, the architecture delivered approximately 5% higher throughput and 12% better time to first token than publicly available results, according to the companies. Time to first token measures how long it takes for the first token to appear.

Responding to grid demand

Bentaus develops, deploys, owns and operates AI infrastructure, hardware and software. Its Ziani platform is designed to adjust compute power consumption in response to grid conditions.

The companies say the software can dynamically reduce GPU and CPU power consumption while keeping AI workloads running. They describe that as a way for data centers to curtail power use in real time.

Upcoming presentation

DriveNets and Bentaus will present the joint solution at SC26 in Chicago, Illinois, at the AMD booth on November 19 at 11:30 a.m. local time. The presentation will be held at booth 3028.