Why share AI infrastructure

A graphics card is not expensive in itself; it is expensive when idle—that was the panel’s main economic takeaway. Small institutions risk building their own mini-centers where GPUs run at only 10–20% and sit idle the rest of the time. A shared, centrally managed model makes it possible to allocate capacity flexibly and increase its utilization across agencies. It also reduces procurement duplication and shortens the time before work on solutions can even begin.

A practical lesson came from a business example: the company initially assigned individuals their "own" cloud GPUs and quickly ran into unsustainable bills. The shift came when, instead of hardware, it started providing access to already-deployed models and only to the portion of capacity teams needed. Operating costs fell by tens of percent, while experimentation remained accessible. The same sharing logic suggests itself for the state as well.