Affordable and open intelligence
More intelligence for less energy, open to everyone.
The mission is to maximize useful intelligence per joule.
We measure progress in tokens per joule because energy is the physical input to machine intelligence and tokens are a specific metric we can track.
Cost is important, but it is shaped by electricity markets, hardware prices, cloud margins, subsidies, and geography. Energy efficiency is more fundamental. It tells us how much intelligence a system can produce from a finite physical resource, regardless of where it runs or who is paying for it.
It also gives every layer of our work a common direction. Advances in chips, systems, compilers, models, and inference software should all help produce more useful intelligence from each joule.
This matters because energy is a constraint on both scale and access. More efficient inference means greater capacity within a data centre, longer operation on a device, and useful intelligence in places where power and infrastructure are scarce. It opens applications that are currently too expensive, too power-hungry, or simply impossible.
Cost tells us whether intelligence is affordable today. Energy efficiency determines how abundant it can become.
Achieving our mission makes intelligence available to far more people, in far more places, for far more purposes. This will drive forwards mankind.