Experience Link:
play.mukunjin.com
Project Repo:
Game Overview
Objectives
Experience the entire AI large model development pipeline, from data collection to model training, and ultimately aim to train an AI model with the highest benchmark score.
Time System
- 1 real second = 1 game day (1x speed)
- Supports 1x / 2x speed, pause available (5x removed)
- Game runs continuously, not turn-based
Initial Resources
- Cash: $150,000,000 (150 million USD)
- Power Capacity: 1 MW
- Cooling Capacity: 1.5 MW
- Data Center: 200 rack slots
Core Gameplay
Complete Training Pipeline
- Collect Data → 2. Purchase GPUs → 3. Research Technologies → 4. Train Model → 5. Benchmark Test → 6. Deploy Model for Revenue
1. Data Collection (New Feature)
Data is the cornerstone of large models. You must collect training data from various sources before training:
| Data Source | Quality | Price | Category | Description |
|---|---|---|---|---|
| Web Crawler | 55% | $5M/10B | General | Large-scale internet text, uneven quality |
| Book Corpus | 85% | $15M/10B | Knowledge | High-quality published books, improves text understanding |
| Code Repository | 80% | $20M/10B | Programming | Open-source code from GitHub, improves coding ability |
| Academic Papers | 90% | $25M/10B | Reasoning | arXiv papers, enhances reasoning ability |
| Synthetic Data | 75% | $30M/10B | General | Model-generated high-quality training data |
| Multilingual Corpus | 65% | $10M/10B | Multilingual | Texts in Chinese, English, Japanese, Korean, etc. |
Data quality directly affects the model’s final score. It’s recommended to collect 50B+ tokens and mix multiple data sources for best results.
2. Purchasing GPUs
Open the purchase panel at the bottom via “GPU Management → Purchase GPU”.
- 1 rack = 8 GPUs
- 9 GPU models, unlocked progressively by company valuation:
| GPU Model | Unlocked at Valuation | Compute (TFLOPS) | VRAM | Power | Unit Price |
|---|---|---|---|---|---|
| A100 80GB | $0 | 312 | 80GB HBM2e | 400W | $15K |
| H100 | $200M | 990 | 80GB HBM3 | 700W | $30K |
| MI300X | $300M | 1307 | 192GB HBM3 | 750W | $20K |
| H200 | $500M | 2000 | 141GB HBM3e | 700W | $45K |
| MI325X | $800M | 1630 | 256GB HBM3e | 750W | $28K |
| B200 | $1B | 2250 | 192GB HBM3e | 1000W | $55K |
| B300 | $2B | 2500 | 288GB HBM3e | 1400W | $70K |
| GB300 NVL72 | $5B | 3750 | 288GB HBM3e | 1800W | $90K |
| Rubin | $10B | 6250 | 288GB HBM4 | 1800W | $100K |
- Max 2000 GPUs per model
- Minimum GPU calculation by model: Training and deployment require different GPU counts based on compute (TFLOPS) and VRAM (using H100 as baseline). Lower compute/smaller memory models require more cards. The UI shows “min N” for each model.
3. Removing GPUs
Click “Remove GPU” at the bottom, select model and quantity. Removal refunds 50% of the buy price.
- GPUs in use for training or model inference cannot be removed—stop training/undeploy model first
- The removal panel shows “removable quantity = total stock - used in training - used for inference” per model
4. Managing Power and Cooling
Power: GPUs require power to operate. Power usage is calculated based on actual load:
-
Training GPU: 95% of rated power
-
Inference GPU: 60% of rated power
-
Idle GPU: 15% of rated power
-
Total consumption = Actual GPU power x (1 + cooling factor 0.30)
-
Overload (actual total > capacity) causes a 3-day blackout; training pauses during blackout
-
Overload check runs when purchasing GPUs, starting training, or deploying models (all increase consumption)
-
Power expansion: $50,000,000 / MW
-
Blackout lasts 3 days, training pauses
Cooling: Cooling capacity must be >= 30% of actual GPU power usage.
- Cooling expansion: $20,000,000 / MW
- If cooling is insufficient (checked on training start), training efficiency drops 30%
5. Expanding the Data Center
- Starts with 200 rack slots (10 rows x 20 columns)
- Each expansion adds 2 rows x 4 columns
- Cost increases exponentially: $500M x 1.8^N
- No practical expansion limit, but cost becomes huge (5th expansion > $9B)
- Confirmation dialog shows current/after expansion size and cost breakdown
6. Technology Research
Click “Research Technology” to select research items. Technologies are organized in a 3-tier dependency tree across the pipeline (24 items):
Tier 1 (Basic, no prerequisites):
- Flash Attention / Mixed Precision / RoPE / Data Deduplication & Cleaning / Curriculum Learning / Sequence Packing / SwiGLU / RMSNorm
Tier 2 (Advanced):
- GQA / ZeRO-3 / Ring Attention / Sparse Attention / LoRA / Knowledge Distillation / RLAIF / Gradient Checkpoint / KV Cache Optimization
Tier 3 (Expert):
- MoE / MTP / 3D Parallelism / GRPO / Constitutional AI / QAT / Speculative Decoding
Research takes time and money; more researchers = faster progress. Max parallel research: (2 + researchers/3).
7. Model Training
Click “New Training” to configure training jobs:
Model Size: Free slider from 1B to 2T parameters, or choose presets
Data Quality: Derived automatically from collected sources; better sources yield better scores
Hyperparameters: Learning rate, batch size, seq length, warmup steps (all adjustable live)
GPU Assignment: Assign training GPUs by model; the minimum needed per model adjusts automatically based on TFLOPS
Tech Selection: Only already-researched/unlocked techs
Training Stages (and sub-stages):
- Pre-training (72%): Data prep → small-scale validation → full training → convergence check
- SFT (20%): Instruction data selection → multi-round training
- Alignment (8%): Preference alignment → safety eval
Training Events: Loss spike, GPU outage (hardware failure), data bottleneck (IO slows), gradient vanishing
Checkpoints: Auto-saves every 10% progress; can roll back after failed runs.
8. Hiring Researchers
Researchers boost training efficiency and research speed. 30-day cooldown after each hire. Higher levels unlock at higher valuations.
| Level | Monthly Salary | Training Bonus | Research Speed | Unlock Val | Max |
|---|---|---|---|---|---|
| Junior | $3,000,000 | +2% | +2% | $0 | 5 |
| Senior | $8,000,000 | +4% | +4% | $500M | 5 |
| Principal | $15,000,000 | +6% | +6% | $2B | 5 |
9. Benchmark Tests
After training, 6 major benchmark tests run automatically. The scoring is challenging; high scores require extensive resources:
| Category | Weight | Description |
|---|---|---|
| Reasoning | 25% | Math, logical reasoning |
| Programming | 20% | Code generation, debugging |
| Text Understanding | 20% | Reading comprehension, summarization |
| Multilingual | 15% | Chinese-English translation |
| Safety | 10% | Red teaming, refusal rate |
| Long Context | 10% | Needle-in-a-haystack |
Score formula: Base (model size) × Quality Bonus (data+tech+alignment, up to 1.3x) × Data Category Bonus (±10%) × Tech Bonus × Random fluctuation × Training interruption penalty
- Small models can hardly break 30 points even with everything maxed
- Midsize models can score 40-50 with good management
- Frontier models can reach 55-60 fully maxed out
- Different techs boost different categories (e.g., GRPO +5% reasoning, RoPE +4% long context)
- Data collection mix affects related category scores (up to ±10%)
10. Deploy Model (Inference)
After training, you can deploy models to provide inference services for revenue:
- Allocate inference GPUs by model (minimum assumes H100 baseline)
- Deployed inference GPUs are locked—not available for training/removal
- Deployed models earn daily API revenue (affected by DAU, token volume, open/closed source, tech/revenue boosts, and sufficient deployment GPU)
- Models can be “taken down” anytime to free GPUs
- Overloading power after deployment causes a 3-day blackout too
11. Funding
- Raise funds every 180 days
- Funding = $1B × (highest model score / 100) × 8
12. Revenue Sources
- API Revenue (daily): Closed models earn per DAU × tokens × price
- Enterprise Licensing (monthly): Based on model score
13. Bankruptcy
If company cash drops below -$50M, bankruptcy is triggered. The game pops up a notice showing your run’s days and final value, saves are deleted, and game resets to start screen.
3D Scene Controls
- Rotate camera: mouse drag (mobile: single finger)
- Zoom: mouse wheel (mobile: pinch)
- Click GPU rack: detail view
- Click power/cooling room: see capacity and current load
- Orange pipes: power lines
- Green pipes: cooling lines
- Semi-transparent building: data center hall
- Ground area scales with expansion, fog limits distant view
Random Events
Every 30–90 days you get a random event: hardware failure, grid instability, data breach, poaching, tech breakthrough, favorable policy, chip embargo, industry summit, algorithm breakthrough, power failure.
UI Overview
The right-side panel uses tab pages:
- Overview: Revenue/expense/profit/cash flow, infra status (shows actual/rated power)
- Products: List of deployed models, including score, size, open/closed source, daily revenue, 6 benchmark scores, used techs, inference GPU usage (H100 equivalent), can “take down” to free GPUs
- Training: Current training progress, phase/sub-phase, loss, GPU utilization, stability, checkpoints
- Research: Research progress, unlocked tech list
- Inventory: GPU inventory details
- Logs: Event log
Top bar: Company name | Cash | Valuation | Days | Speed control
- Pause / 1x / 2x: Game speed control
- Auto-save: Saves every 100 real seconds (also autosaves before refresh/close)
Bottom action bar: GPU Management | Infrastructure | Team | Data Collection | New Training | Funding | Delete Save
- GPU Management: Buy GPU / Remove GPU
- Infrastructure: Expand Power / Cooling / Data Center
- Team: Hire Researchers / Research Tech
- Group buttons expand drop-down, click item to open page
- Delete Save: Clears all progress, confirmation dialog before action
Stack
- HTML5 + CSS3 + JavaScript (ES6+)
- Three.js 0.148 (3D rendering)
- Tailwind CSS 3.x (UI styling)
- Frontend-only, no backend
- Mobile supported: 3D touch controls (rotate/zoom), bottom bar scrolls, dropdowns center, modals adapt
File Structure
model-rush/
├── index.html # Main page
├── README.md # This file
├── DESIGN_PLAN.md # Detailed design doc
├── css/
│ └── style.css # Custom styles
└── js/
├── config.js # Game config/constants
├── game.js # Game state & update loop
├── economy.js # Economy system
├── training.js # Training system
├── benchmark.js # Benchmark scoring
├── research.js # Technology system
├── data.js # Data pipeline
├── events.js # Random events
├── save.js # Save system (LocalStorage)
├── scene.js # Three.js scene
├── datacenter.js # 3D datacenter
├── ui.js # UI updates/interactions
└── main.js # Entry point
It’s still in testing, so lots of issues are expected. Please try it out and suggest improvements—I’ll definitely refine things. If you’re skilled, feel free to open issues or PRs. And you’re welcome to star the repo, thanks!
