Enabling "Maximum Enterprise Utilization" with AI

If you spend enough time reading AI papers and articles, you might have run into the acronym "MFU": Maximum FLOPS Utilization. Each GPU has a theoretical FLOPs capacity, but we often only get 20-40% MFU due to issues like sub-optimal scheduling, memory bottlenecks, etc. (See our podcast episode with Quentin Anthony of Eleuther AI (opens in a new tab) for a deeper technical breakdown)

You can calculate it with:

MFU=Actual FLOPsTheoretical Peak FLOPs MFU=\frac{\text{Actual FLOPs}}{\text{Theoretical Peak FLOPs}}


Theoretical Peak FLOPs=(Number of cores)×(Clock speed in Hz)×(FLOPs per cycle) \text{Theoretical Peak FLOPs} = (\text{Number of cores}) \times (\text{Clock speed in Hz}) \times (\text{FLOPs per cycle})

With the rise of AI copilots and agents in the enterprise, more and more tech leaders should be using a similar framework: what is the "Maximum Enterprise Utilization" (MEU) of my company, and how can I adopt AI to improve it?

MEU=Total ProductivityAll Available Work MEU = \frac{\text{Total Productivity}}{\text{All Available Work}}

The "All Available Work" for an enterprise is pretty self explanatory, it's all work tasks that need to be done in your company.

"Total Productivity" is the sum of all work throughput from all "Workers", which includes employees, contractors, and now AI agents:

Total Productivity=iWorkers(work hoursi×productivity ratei) \text{Total Productivity} = \sum_{i \in \text{Workers}} (\text{work hours}_i \times \text{productivity rate}_i)

AI is the first technology in decades that can increase Total Productivity without requiring more work hours by humans. I'm guessing the average MEU for a software company isn't much higher than 20-40%; if you lead an engineering team, a rough way to calculate it is your quarterly tickets throughput as a percentage of your whole backlog.

It's been really hard to improve on the MEU number historically:

60-80% of enterprise productivity in the future will likely be delivered by AI:

If you're a founder looking to build a company in the AI space selling to enterprises, you should ask yourself some of these questions:

Looking forward to reviewing this post at the end of 2024, which many are calling the year of "AI in production".

© Alessio Fanelli.RSS