Half-Price Cards Filled the Building
One hundred and twelve hours, and the fifth generation still had not reached the point where the game lets you start over stronger.
The game is an idle game about running an AI lab: you buy chips, build power, train models, and eventually prestige into a new generation with a perk or two carried over. To check the balance, I have bots that play it. They are simple planners, and when the lab has spare power, they buy whichever chip gives the most peak compute per dollar. On seed 21, with every perk bought in the order the bots buy them, generations one through four ended in hours. Generation five ran 112 and was still going.
The perk responsible is one I was rather pleased with. Secondhand chips: once a newer chip is on sale, the older one costs half price. Halving the price of a real 2017 gaming card, the GTX 1080 Ti, is enough for it to beat an H100 on compute per dollar. So from then on, every time the lab finished a new megawatt of power, the bots filled it with 1080 Tis.
They were not wrong about the dollars. They were wrong about the building. A 1080 Ti delivers about a thirtieth of an H100’s compute per watt. Each new megawatt arrived, was packed with the cheapest compute available, and was immediately full of chips doing a thirtieth of the work the power could have bought. The only thing left to do was build the next megawatt, and the next expansion took a real day to land. The bots kept making the locally excellent purchase all the way down.
The score was missing the scarce thing. Dollars were plentiful by generation five; power was not. A chip’s cost now includes the power it occupies, priced at what the next megawatt of the home site costs, server overhead and cooling included. When power itself is the binding limit, the bots compare chips per watt outright.
With that one change, the same five generations end after 41, 25, 16, 15 and 15 hours. Generation one, where power never binds, moved by three seconds.
The new balance test fails on the old scoring, with the bots buying 1080 Tis, and passes on the new. It also checks its own premise before anything else: that on sticker price alone, the secondhand card still wins. It does. That was never the problem.