01
Build a model step by step
Pick the model size, the data and the compute, and the game shows straight away how good it will be, how long training takes and what it costs. Quality comes from a real scaling law from AI research.
An AI company tycoon / free alpha 0.5.0, 20 September 2026
Scaling Laws : a free AI company tycoon about timing hardware, not buying it
January 2022. Twelve million dollars. No product. Train models, fund research and pay for compute, while your rivals keep moving whether you are ready or not. The expensive decision is rarely just what to buy. It is when.
Current build
Unity / current development build
HCK Labs
What you do in the game
You run an AI lab. This is what it looks like.
Eight things you do from the first day. All of them are in the free 0.5.0 build.
02
Research takes time
Money speeds research up, but you cannot buy the calendar. New architectures, better data and even bigger offices unlock one after another.
03
Watch the temperature
Eight cards are not always better than seven. A fan instead of a card can give you more power. You buy the cards yourself and fit them into the cabinets in the basement.
04
Ship an update
A quarter of your users move to an update on day one. Careful, though: if it is worse than the last one, the rest never come.
05
Sell it to people
A good model is not enough. You set the price, decide how much to give away for free and pick the advertising: social media, press, creators or television.
06
The world happens
Twenty-three real news events reach you and your rivals on the days they actually happened. After a serious incident, the regulator gives you five days before deciding on a penalty.
07
Hire people
Emil walks you through your first hour and pays for your first research. He runs a small lab of his own, too. Then you hire a team, and every person has their own expectations and working hours.
08
Find the money
Investors, loans and grants. Each has a price: a share of your company, a monthly fee, or conditions you have to keep.
State programmes
When a government starts to trust you
A government can lend you up to 10 billion dollars, but you pay back almost three times as much, and while you owe it, nobody can buy your company. After five years without a serious incident, it can also hand you eight sectors of a country, from paperwork to defence. They pay more than the market while your model stays near the top, and a failure costs billions.
$10Bthe largest loan
5 yearsof clean history before a state trusts you
8sectors of a country to run
The competition
Play dirty. Just be careful, they may notice.
Fourteen labs grow alongside you. Each one remembers what you did to it, even when nobody can prove it was you.
4 levels of smearing
You pay for stories about a rival. The pricier the level, the harder it hits and the easier it is to get caught. If they trace it, you get a letter from a lawyer, a phone call, and maybe a lawsuit.
Lawsuits
You can sue a rival. The more you demand, the lower your chance of winning. You pay the costs up front and nobody gives them back.
Poaching people
Open a lab, see who works there, and make a better offer. Every attempt costs you some of the relationship.
Takeovers
Buy a rival's shares. Once you hold the majority, you can buy the whole company, newest model included.
Who is on the board
Parodies of real labs, and one invented lab from Radom. Thirteen start where the real ones stood in 2022, and then each goes its own way.
- OpenSI2015
- Antropic2021
- DeepThink2010
- Infinity2013
- Astral2023
- DeepSearch2023
- zAI2023
- Swen2023
- Grob2022
- StableAI2020
- IntroduceAI2022
- Algho Alpha2019
- Gohere2019
- E-Solutions2021
In motion
Recorded from the builds, not animated for the page.
Eleven clips, the five newest from 0.5.0. Every figure on screen came out of the simulation on the day it was recorded.
New in 0.5.0
Shipped today, not next time.
Rendered from the 0.5.0 build, released 20 September 2026. All four are in the download.
Humble beginnings
It starts in a house, not a glass tower.
The first company view is deliberately modest: a desk, a bed, a living room, a garage and a cheap car. Growth is something you see, not only a number: bigger offices, people at desks and your own infrastructure, and one day somebody offers to buy the whole company.
The founder walks around the rooms, and your staff sit at their desks during the hours you set. You research, rent and then buy new offices. The city of Bayview around those buildings is finished and arrives in the next build.
The loop
Research. Train. Release. Earn.
Every decision here costs money and time. A better model only helps if the company is still around on the day it is finished.
Unlock it
New architectures, data and model types unlock through research. Money helps, but some things you simply have to wait for.
Train a model
Pick the size, the data and the compute. The creator shows how good the model will be, how long training takes and what you will pay.
Pick the launch day
A finished model can wait. Sometimes it is worth waiting out a rival's launch, but the market keeps moving meanwhile.
Make it pay
Set the price, the free plan and the advertising. A company that ships one model and stops growing falls behind.
Under the interface
Real rules, made playable.
The game is built on published AI scaling research and real hardware data. Where it has to guess the future, it says so.
~20×
Tokens per parameter
A model trains best when it gets about twenty tokens of data for every parameter. Too little or too much data and the same budget buys a worse model.
10× → +10
The compute race
Ten times the compute buys about ten points of quality. That is why your rivals keep pulling away, even when your last model was expensive.
2×
Hardware loses value
Cards lose value with time, and again when a new generation arrives. Buy too early and your money is stuck in hardware the market has already passed.
Engineering notes and source are public in the Scaling Laws repository.
Current build
Playable from start to finish. The world around it is still growing.
Version 0.5.0, released 20 September 2026, plays end to end, from the first day of January 2022 through a campaign that lasts years. It is an early version and some screens are still plain. Everything below is in the build you can download today.
Model creation
A step-by-step creator that shows quality, training time and cost as you go.
Research
A technology tree and your own architectures. Every unlock costs money and time.
Company setup
Founder skills and traits, your company name and country, and every country has its pros and cons.
Business layer
Prices, customers, market share, advertising, investors, loans, the ranking and paid intelligence.
Physical site
The starter home, offices you rent and then buy, and a server room you fill cabinet by cabinet.
Safety and regulators
Safety on every model, and a regulator that gives you five days before deciding.
Rivals and relations
Fourteen labs: shares, takeovers, poaching, smearing and lawsuits, in both directions.
Testers
MiNatix played Scaling Laws for hours, went through this site and PC Workman, and wrote down every place that was unclear or broke. Half the fixes on this page are hers. She is in the game's credits.
Questions people ask
Is Scaling Laws free?
Yes. Every build so far is free for Windows on itch.io. A Steam release is planned for October to November 2026.
What is the current version?
Version 0.5.0, released 20 September 2026, an alpha. The economy was rebuilt so that serving people pays for itself, the parts warehouse has an order window with express delivery, and the creator says how many people your own cluster can keep served.
Which AI labs are in the game?
Fourteen parody labs: OpenSI, Antropic, DeepThink, Infinity, Astral, DeepSearch, zAI, Swen, Grob, StableAI, IntroduceAI, Algho Alpha, Gohere, and E-Solutions, the invented lab of the player's cousin in Radom. Thirteen start where their real counterparts stood in 2022, and then each goes its own way.
Can I buy or attack rival labs?
Yes. You can buy shares and take over a lab with a majority, hire its staff, pay for smear stories at four levels, and sue it. Rivals can sue back. Every action changes the relationship.
Are alliances or mergers in the game?
Not yet. Alliances between labs and mergers are planned and not in any build.
Is the source code open source?
It is source available, not open source. The code can be read on GitHub, but from 19 September 2026 it may not be redistributed. Earlier versions stay under PolyForm Noncommercial 1.0.0.
Who makes Scaling Laws?
Marcin "HCK" Firmuga, a solo developer from Radom, Poland, publishing as HCK Labs. He also made PC Workman, a Windows monitor on the Microsoft Store.
Built in public
A game built in the open, one week at a time.
Scaling Laws is being built openly by Marcin "HCK" Firmuga under HCK Labs. The source, engineering notes and development history stay visible while the game changes.