You are running a small AI datacenter.

Every time someone asks an AI a question — a coding question, a search, an image generation — a computer somewhere actually does the work. That computer is a GPU (graphics processing unit), and one full "ask" run through the model is called an inference.

Your job: own enough GPUs to handle the day's inference demand, price them right, and keep the power cheap without poisoning the town.

   ┌───────────────────┐
   │ ████ NVIDIA  A100 │   workhorse · $8k · 300 inf/day
   │ ████ ░░░░░░  ▓▓▓▓ │   the steady volume tier — indie devs,
   │ ████ HBM:80G ▓▓▓▓ │   cheap inference, RAG pipelines.
   └───────────────────┘

   ┌───────────────────┐
   │ ▓▓▓▓ NVIDIA  H100 │   premium · $25k · 900 inf/day
   │ ▓▓▓▓ ░░░░░░  ████ │   the standard for production LLM APIs.
   │ ▓▓▓▓ HBM:80G ████ │   pay-for-quality customers ride these.
   └───────────────────┘

   ┌─────────────────────┐
   │ ██▓██ NVIDIA  B200  │   frontier · $60k · 2,000 inf/day
   │ ██▓██ ░░░░░░  ▓███▓ │   the new generation. frontier labs and
   │ ██▓██ HBM:192G ████ │   long-context customers fight for these.
   └─────────────────────┘

How an inference happens:

1. A user types a prompt: "write me a python script"
2. The prompt is broken into tokens (~750 words per 1000 tokens).
3. The GPU loads the model (tens to hundreds of GB) into its HBM memory.
4. It runs a forward pass — billions of matrix multiplications.
5. It streams tokens back to the user, one at a time.
6. That single response = roughly 1 "inference" in this game.

Why the tiers matter:

• A100 — older but cheap and plentiful. Can't fit the biggest models. Good for fine-tuned 7B-class workloads.
• H100 — current production standard. Most paying customers will only sign for H100-or-better.
• B200 — newest, fastest. Frontier-research workloads need these or they go elsewhere.

The catch: GPUs burn a lot of electricity. A rack of B200s pulls more power than a small neighborhood. Where you get that power is the real game — and the regulators are watching.

Hit BACK and pick a difficulty. Easy = short campaign with more starting cash; Hard = 60 days, you're under-capitalized.

You run an AI compute datacenter for 30 days. Make money. Don't go bankrupt. Don't poison the town.

Each day:

1. Read the news ticker. Model releases drive demand spikes.
2. Buy GPUs (click +A100 / +H100 / +B200). Each one fills a slot on the floor.
3. Pick power: diesel (cheap, fines), grid (medium, blackouts), PPA (clean, costly).
4. Set your $/inference. Charge too much, customers leave. Too little, you bleed.
5. Hit RUN DAY. Watch your floor run. Adjust tomorrow.

Score = ending cash + (reputation × $1k) − fines paid. Top of the leaderboard wins bragging rights.

Tip: premium customers (H100 / B200 buyers) won't sign if reputation drops below 30. Stay clean enough for them to take your calls.