There were quite a few of these around when I was learning to program: I and a couple of friends competed in one where you provided a robot that had to kill survive longer than the rest. Each action (movement, shooting at an enemy, using your shield, scanning to see anything) took energy and sources of that were unevenly distributed around the simple landscape (so a robot could camp on a large source for a while as an example tactic). My robots didn't tend to do very well unfortunately.

There must be modern equivalents around, though it sounds like you want to go the other way: providing the challenge to the AIs rather than an AI itself, so perhaps you could create your own.

I've considered it - it is one of the many projects on my "when I win the lottery or otherwise have a lot of spare time all of a sudden" project list. There are two main options: run the AIs on your platform in which case one of the constraints is processing resource (which you somehow need to hand out fairly) or just provide an API through which the AIs can submit moves (in which case you can't govern how much processing resource each AI can throw at the problem, but the game service would be very resource efficient to run. Of course with the latter option if there are limits to what the AI can see (fog of war or similar) then you have to play the whole game out before letting spectators see it otherwise an AI could abuse the spectator interface to gain advantage, if the AIs are hosted on your resources you can constrain such visibility much more readily.