You run the physical layer of the cloud. Racks earn money while they are powered and cool — everything else is the fight to keep those two things true.
Sister game of Server Survival: that one teaches the logical layer (services, routing, scaling); this one teaches the physical layer it all runs on — megawatts, heat, cooling, PUE.
Real capture, sped up with the game's own fast-forward: the chain goes live → the rack block cooks itself past 45 °C and throttles → CRACs answer it (and push PUE up) → the city grid dies, and only the UPS-backed half of the room stays lit until the standby generator picks up. Throughout, badges name the building responsible for every kilowatt you fail to serve — and the ledger totals it in dollars.
-
Power is wired. Grid Feed → Transformer → UPS → PDU → Rack. Every link has a kW capacity. Overload one and it clips its whole subtree proportionally — push it further and its breaker opens, because real gear does not dim forever. A UPS bridges a blip in seconds; a standby generator carries the rest, for exactly as long as there is fuel in it.
-
Heat spreads. Every kW a rack draws becomes heat in the cells around it. CRAC units cool a radius — and draw power themselves, idle draw included, so two half-loaded units cost more than one working one. Above a certain size a chiller making chilled water for many CRAH heads beats cooling everywhere at once; below it, the plant's pumps cost more than they save. That is why your PUE (total facility power ÷ IT power) is both your score and your power bill. Press T for the thermal overlay and watch hot aisles form from your own layout. Drag (right/middle mouse) or use WASD / the arrow keys to pan the camera around the floor.
-
The plant drinks. A chiller's cooling tower rejects heat by evaporating water — around 1.8 litres per kWh, which is what the industry calls WUE and what the HUD shows beside PUE. It is cheap enough that the loop still wins, right up until a drought prices water twelve times over: then the air-cooled CRAC, strictly worse on power at every size, is the cheaper room. PUE is the number everyone quotes; WUE is the one that gets a datacenter into the local newspaper.
-
A run can be handed to someone else. Free play takes a seed —
?seed=KYIV, or type one on the menu — and every crisis, every contract and every price window in that run is drawn from it. Two people playing?seed=KYIVget the same brownout at the same second, so "I held PUE 1.19 for four minutes" stops being a claim and starts being something you can check. The seed rides along in the address bar; the game-over screen hands you the link. Without a seed nothing changes — the game is as random as it always was.
Overheated racks throttle, throttled racks miss SLA, missed SLA drains
reputation and money. And every kilowatt you fail to serve is named:
badges float the cause over the building responsible, and the ledger totals
the run in dollars — Where the $65 went: AT CAPACITY 61%, TOO HOT 20%….
Thirteen levels in five chapters, each teaching one mechanic and each proven — by machine-played tests — to be winnable with that mechanic and losable without it.
- Chapter 1 · Power & Heat — the delivery chain, the hot aisle, PUE and placement, a grid sag that only headroom rides out, and a blackout that only a charged UPS bridges.
- Chapter 2 · Backup — a standby generator, its transfer switch, and the fuel gauge that is the real capacity of your backup.
- Chapter 3 · Diagnosis — levels that hand you a room that is already running and already wrong: four CRACs burying the PUE, four racks sharing one bus until its breaker opens, cooling installed in the wrong corner, and two "redundant" feeds that turn out to share a substation.
- Chapter 4 · Scale — one chiller plant feeding many cooling heads: cheaper than cooling everywhere at once, right up until the day the plant stops and every head on it stops together.
- Chapter 5 · Serviceable — scheduled work orders with a deadline: Tier III isn't "has a backup", it's concurrent maintainability — any element can be pulled for planned work while the load keeps running, which is a statement about spare capacity, not about spare parts.
Optional bonus objectives sit on a different axis than the level's own goal — serve through the sag, or bridge three blackouts and still finish with money in the bank.
The Lab sits below them and is open from the first minute, because a rehearsal room behind thirteen wins is a trophy. It is a working hall — one chain with a UPS, a standby generator already wired to the transformer, three racks and a CRAC — with live knobs for demand, ambient temperature and the tariff band, and buttons that fire a heatwave, a brownout, a grid outage, a peak-price window or a CRAC failure on demand. There are no objectives, no clock and nothing to lose. Every crisis in the rest of the game is on a schedule, so understanding the transfer switch costs you 220 seconds of waiting and gets you one look at it; here you can watch the same one until you can predict it.
- The datacenter power chain, and why a breaker opening is not the problem
- Thermal design: hot and cold aisles emerge from placement, not from rules
- PUE as a profit lever, and why over-cooling is as expensive as under-cooling
- Redundancy that is real (independent substations) versus redundancy that is decoration
- Batteries bridge seconds, generators carry hours, fuel is the actual limit
- Shared cooling is shared efficiency and shared blast radius — the same property, and scale decides which one you get
- WUE, and that efficiency has a second bill: the loop buys its power advantage with evaporated water, and a drought is what makes that matter
- Diagnosis: reading an attribution ledger instead of guessing
To play, click PLAY NOW. Nothing to install, nothing to build, no account. That is the whole setup.
No build step anywhere: the repo is served raw by GitHub Pages — native ES
modules, Three.js from CDN, not a line of it compiled. To run your own copy
(a fork, or a change you just made) you need a static server rather than the
file itself, because browsers fetch module scripts in CORS mode and a
file:// origin is opaque to them. Double-clicking index.html gets you a
HUD over an empty screen. One line fixes it, and it is already on your
machine:
python3 -m http.server 8000 # then open http://localhost:8000For contributors: npm i && npm run check runs ESLint and the Vitest suite.
The simulation is tested headless (power conservation, heat conservation,
loss attribution that must sum exactly, breaker timing, no-NaN invariants),
and every campaign level is machine-played in both directions. npm run demo:capture && npm run demo:gif re-records the README animation from a
real playthrough.
Thirteen campaign levels across five chapters plus The Lab, nine buildings, 592 tests. Simulation: wired power with inverse-time breakers, a diffusing heat field with part-load cooling, a shared chilled-water loop metered for water on a WUE, UPS buffers, standby generators with fuel, grid sags, per-substation outages, time-of-use and peak tariffs, droughts, peak shaving off the UPS battery, rolling contracts, seeded shareable runs, and per-cause loss attribution. English and Ukrainian.
Peak shaving is a toggle, not an upgrade. A charged UPS can serve its subtree from the battery instead of the meter, but the battery gives back less than it takes, so the round trip only pays if you spend it into a band that is dearer than the one you buy it back in. Left switched on at a flat price it loses money and leaves you with no ride-through — the FAQ's Shaving tab has the numbers, generated from the same config the simulation bills from.
The Lab was the last open item on the roadmap (#5); it is in. What comes next is whatever a level turns out not to be able to teach.
The bar is that a mechanic has to teach something true — a mechanic that plays well and models the physics dishonestly is a bug. Everything a contributor needs is in CONTRIBUTING.md, with the invariants behind it in docs/ARCHITECTURE.md.
Starting points: anything tagged good first issue is a real, verified gap with the fix already located. Questions and mechanic proposals go to Discussions.