The Cool Aisle by George Notter... 11% MFU and a Special Request

研报英文原文证据摘录

The Cool Aisle by George Notter... 11% MFU and a Special Request

May 24, 2026

how the models are designed. Nonetheless, we think it's a good view on how model performance is trending. Jensen

– at a presentation at Stanford last week – called the MFU metric "just simply wrong," and he would "rather be at low

MFU." The point he made was simply that he would rather overprovision hardware than not having enough compute.

Also, Jensen noted that any given time, constraints in compute, memory, or network could limit the performance of the

whole system. This aligns with how we view the importance of networking within the broader AI infrastructure stack

– if network performance is bottlenecking GPU performance, Cloud Providers, Model Builders, and Neoclouds need to

pay more attention towards optimizing it.

Who Fixes the MFU Problem? In our view, networking is a significant factor in improving MFUs. Increased network

efficiency allows more of the time to be spent on GPU-hours – rather than in the network, as described above. In this

context, these xAI Colossus clusters are built using NVIDIA's SpectrumX Ethernet switching fabric. It's an important

detail for the competitive narrative at Arista – we've long considered them to be the "best-of-breed" switching provider.

By no means is the 11% MFU directly contributable to a deficiency in SpectrumX's switching infrastructure at xAI –

as we noted above, there are challenges with scaling GPU clusters at the scale that xAI has done to this point (550k

GPUs). However, if Arista is able to demonstrate meaningfully higher utilization benefits, it's possible it could open

the door for them at Neocloud accounts that have previously been "all-in" on NVIDIA's full stack. Elsewhere, vendors