AI Datacenter Tycoon · 2020 → 2028

GIGAWATT

Buy real GPUs. Design your own racks. Fight physics for every kilowatt. Wire a supercomputer and lease it to the labs racing to train the next frontier model.

Download free View on GitHub

Linux · Windows · macOS  ·  plays on Steam Deck with a controller

GIGAWATT top-down view of a datacenter hall: racks labelled by hardware and customer, with the Overview drawer open on the right 690+ real parts live thermal state train & serve
690+real parts from 100+ vendors
70+ready-to-order system templates
97months, Dec 2020 → Dec 2028
4cooling classes: air, rear-door, liquid, immersion
330automated tests keep the economy honest

THE LOOP Four moves, a thousand tradeoffs

1

Design

Pick a platform or build a tray and rack from individual CPUs, DIMMs, GPUs, NICs and PSUs. The validator tells you what breaks.

2

Order

Real lead times, allocation caps, vendor relationships and volume discounts. Buy DIY, an OEM rack, or a turnkey one.

3

Build

Power, cooling, floor loading, internet circuits. Air, rear-door, direct-liquid or immersion. Overcommit and racks trip.

4

Lease

Training customers rent a coherent cluster by the GPU-hour. Inference customers pay per million tokens. Miss the SLA and they leave.

FEATURES Everything is a real engineering constraint

Procure like an operator

A guided planner (platform → how many → who builds it) on top of a full parts bin. Specs, street prices that drift with the AI boom, lead times that stretch when demand spikes.

  • HGX A100 to Vera Rubin NVL144, AMD Instinct, Gaudi, Cerebras, Tenstorrent
  • Vendor frameworks and relationships unlock discounts and allocation
  • Open orders can be cancelled or trimmed, with a restocking fee
The Procure drawer showing platform cards and a build planner

Design at the motherboard level

A tray bench with sockets, DIMM banks, GPU baseboards, NICs, NVMe and PSUs, plus a 42U rack elevation. Live validation and a five-year cost advisor.

  • Power, weight and cooling interface are checked before you order
  • Save designs and reuse them across halls
The Design Lab tray bench

Wire a real supercomputer

Group racks into one machine, auto-design the fabric or hand-tune it, and see what it can train or serve: per-model tokens per second, MFU and time to first token.

  • Fat-tree bisection, scale-up domains, oversubscription
  • MoE and dense models behave differently, as they should
The Supercomputer builder showing a fabric and model fit table

Customers who actually want things

Offers show a live fit verdict and exactly what you still need to buy. Hit Easy Path and the game orders, installs, deploys and assigns for you as parts arrive. Good service earns renewals.

  • Training vs inference, SLAs, ramp windows, three rival operators
  • Sell idle capacity on the spot market, or sign power purchase agreements
The Customers drawer with market offers and a build suggestion

Run it, age it, sell it

Staffing scales with racks, liquid cooling and fabric size. Hardware loses value every year and every GPU generation. Know when to replace a rack and when to ride it out.

  • Burn-in and maintenance policy, spares, failures, repairs
  • Bankruptcy after three months of negative cash; a scored report at the end
The Facility drawer with power and cooling gauges and the machine ledger

HEAT LAB Can your hall cool that rack?

A back-of-the-envelope version of the question the game asks you every month. Rack powers are rounded vendor figures; the game computes them from the exact parts you choose.

0racks fit in the power budget
0accelerators

STEAM DECK Play on the couch

Full controller support: D-pad navigates every drawer, sticks fly the camera, and the UI scales up on small screens. A Steam release is planned.

  • A select
  • B back / close drawer
  • X inspect the rack under the reticle
  • Y next month
  • LBRB switch drawers
  • LTRT zoom
  • L-stick pan R-stick orbit
  • D-pad navigate
  • Start menu View game speed
A campus-scale hall in thermal view
Fly the camera across hundreds of racks with the sticks; the in-game Codex explains every part as you meet it

UNDER THE HOOD A deterministic engine you can test and train against

Godot 4Vulkan front-end, 3D hall, drawers, controller
⇄
Node sidecarJSON line protocol over a pipe
⇄
TypeScript engineevery rule: power, cooling, fabric, markets

Same seed, same game

The monthly tick is an ordered pipeline of systems. Golden trajectory tests pin full campaigns bit for bit.

Scripted players

Six personas (idle, random, operator, tycoon, cautious, spot-only) play full 97-month campaigns and check whole-state invariants every month.

An RL environment

The engine exposes a step/reset interface with a Python wrapper, so agents can learn to run a datacenter.

DOWNLOAD Free, no account

Loading the latest release…

Builds are unsigned, so your OS will warn you the first time. Each release lists SHA-256 checksums. Everything is built by GitHub Actions from the public source.