
Run an AI inference business for one year
The Inference Machine is a browser-based simulation game from MakerMaker.ai in which you run an AI inference business for one in-game year. Starting with $10,000 and no experience, you rent GPUs from two providers (Nbibia and AMB), decide which open-weights models to host and when, and dedicate spare GPUs to Elise RSI, an in-game research agent that finds speedups for your inference engine. The goal is to maximize profit by year end, with a Top 10 leaderboard threshold of $66M.
Rent GPUs from two legally distinct providers, Nbibia and AMB, and manage a fleet of rented hardware with per-GPU daily costs.
Choose which open-weights models (such as GigaParrot 120B) to deploy and serve to customers, balancing price per token, demand, and GPU requirements.
Assign idle GPUs to Elise, an R&D agent that runs optimization work to improve your inference engine's speed and profitability.
A news feed delivers world events over the 52-week run, and you make strategic decisions in a changing market.
Track daily profit, total revenue and costs, GPU utilization, price per token, demand, and profit-per-GPU, with charts comparing performance with and without Elise.
Aim for a $66M profit milestone to reach the Top 10, with additional GPUs and upcoming models unlocking as your fleet grows (at 16, 64, and 256 GPUs).
You begin the one-year run with $10k capital, no deployed models, and no GPUs.
Deploy open-weights models from the Serving tab; without a deployed model, users cannot use your service.
Rent GPUs from Nbibia and AMB, splitting the fleet between serving traffic and idle capacity.
Give spare GPUs to Elise RSI to discover speedups for the inference engine instead of letting them sit idle.
Over 52 in-game weeks, adjust pricing, hardware, and R&D to maximize profit, aiming for the $66M Top 10 threshold.
It is a browser game by MakerMaker.ai where you run an AI inference business for one in-game year: renting GPUs, choosing which open-weights models to serve, and using spare GPUs for inference-engine optimization.
You begin with $10,000 and no GPUs. You rent GPUs from Nbibia and AMB, deploy a model under the Serving tab so users can use your service, and optionally assign idle GPUs to Elise R&D under the R&D tab.
Maximize your profit by the end of the 52-week year. Reaching $66M in profit advances you to the Top 10 on the leaderboard.
Elise RSI is the in-game R&D agent. When you assign spare GPUs to her, she works on finding speedups for your inference engine, which improves serving efficiency. The monitoring charts show your profit with and without Elise's optimizations.
The game features open-weights models such as GigaParrot 120B, plus several upcoming models that unlock as your GPU fleet grows (at 16, 64, and 256 GPUs).
The stats view shows revenue minus GPU rent per day and per GPU, assuming 100% utilization. The game notes that in practice utilization can be lower when your market share is small, so real profits may differ from the stated figures.
No. It is a browser game that you play directly on the MakerMaker.ai site.
The site links to a MakerMaker.ai GitHub page, but the page does not state whether the game itself is open source.