Compute is the product

The datacenter game has grown fast. On 11 September alone it got a lot of new parts. There are storage buildings, AI training, deals with customers, market crashes, parts that wear out, shortages, heat and an ending.

The basic idea is the same as last time. Power, cooling, internet and chips go in, and computing power comes out. Buyers pay for it, and the price falls the more you sell. Your servers wear out while they run, so you have to replace parts. Servers packed close together need more cooling, so where you put things matters. Now and then power or chips run short for a minute. And sometimes the bubble wobbles, and every buyer pays half for ninety seconds.

You can also train your own AI models. To train one, you need data from the internet. But the more AI answers you sell, the more of that data is your own answers, and the worse your next model gets. It’s the one thing in the game that punishes you for winning. The ending is one giant final model. Keep it running for ten minutes and you win.

It feels like a real game now. You can pick from three places to build, and each one has different prices for power and chips. You can speed the game up two or four times. You can copy a group of buildings and place it again somewhere else. The game has sound now too.

Next is the demo. A new game walks you through the first ten minutes with a checklist. There are three save slots, a settings page and autosave.

The part I’m least sure about is the core idea. Every server you add lowers the price of what servers make. That’s the AI bubble in one rule. But it’s a strange rule for a building game, where the fun is building more. Playtests will tell me if players understand why their price fell.