Lab Infrastructure Options
Building a "Physical AI" lab is a significant investment. You must choose between building a physical On-Premise Lab (High CapEx) versus running a Cloud-Native Lab (High OpEx).
This guide helps you evaluate the trade-offs and choose the right infrastructure for your teaching environment.
Option 1: On-Premise Lab (High CapEx)
Best for: Institutions with upfront budget, long-term programs, hands-on hardware access.
Infrastructure Components
| Component | Specification | Quantity | Est. Cost per Unit | Notes |
|---|---|---|---|---|
| Workstation | RTX 4080 + i7 + 64GB RAM + Ubuntu 22.04 | 1 per student | $2,500 - $3,500 | Core simulation rig |
| Edge Kit | Jetson Orin Nano + RealSense D435i | 1 per student | $700 | Physical AI deployment |
| Robot (Shared) | Unitree Go2 or G1 | 1 per 4-6 students | $3,000 - $16,000 | Shared physical platform |
| Networking | Gigabit switches, Wi-Fi 6 access points | Lab-wide | $500 - $1,000 | Low-latency robot control |
| Storage | NAS for datasets and models | 1 per lab | $1,500 - $3,000 | Shared Isaac Sim assets |
Sample Lab Configuration (20 Students)
| Item | Quantity | Unit Cost | Total Cost |
|---|---|---|---|
| Workstations | 20 | $3,000 | $60,000 |
| Jetson Edge Kits | 20 | $700 | $14,000 |
| Unitree Go2 Robots | 4 | $3,000 | $12,000 |
| Networking Equipment | 1 set | $1,000 | $1,000 |
| NAS Storage (10TB) | 1 | $2,000 | $2,000 |
| Total CapEx | $89,000 |
Advantages
- Zero Latency: Direct hardware access for real-time control
- Predictable Costs: One-time investment, no recurring fees
- Hands-On Learning: Students physically interact with robots
- Data Privacy: All data stays on-premise
- Offline Capability: Works without internet dependency
Disadvantages
- High Initial Investment: $89,000+ upfront cost
- Maintenance Burden: IT support for hardware issues
- Space Requirements: Dedicated lab space needed
- Hardware Deprecation: Equipment becomes outdated
When to Choose On-Premise
- You have upfront budget but limited operational budget
- You plan to run this course for 3+ years
- You want students to deploy code to physical robots
- Your institution has IT support for lab maintenance
- You prioritize data privacy and offline capabilities
Option 2: Cloud-Native Lab (High OpEx)
Best for: Rapid deployment, remote learning, students with weak laptops.
Cloud Architecture
┌─────────────────────────────────────┐
│ Student Laptop (Any OS) │
│ - SSH/Web Browser │
│ - VS Code Remote │
└──────────────┬──────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ AWS/Azure Cloud Instance │
│ - g5.2xlarge (A10G GPU, 24GB) │
│ - Ubuntu 22.04 │
│ - Isaac Sim + ROS 2 + Gazebo │
└──────────────┬──────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ Local Edge Kit (Still Required) │
│ - Jetson Orin Nano │
│ - RealSense D435i │
└─────────────────────────────────────┘
Cloud Workstation Specifications
Instance Type: AWS g5.2xlarge or Azure NC6s_v3
- GPU: NVIDIA A10G (24GB VRAM)
- vCPUs: 8
- RAM: 32GB
- Storage: 200GB EBS volume
- Software: NVIDIA Isaac Sim on Omniverse Cloud (requires specific AMI)
Cost Calculation (Per Student)
| Item | Usage | Rate | Cost |
|---|---|---|---|
| Cloud Instance (g5.2xlarge) | 10 hours/week × 12 weeks | $1.50/hour | $180 |
| Storage (EBS 200GB) | 3 months | $20/month | $60 |
| Data Transfer | ~100GB outbound | $0.09/GB | $9 |
| Jetson Edge Kit (Still needed) | One-time | $700 | $700 |
| Total per student | $949 |
Sample Lab Configuration (20 Students)
| Item | Quantity | Unit Cost | Total Cost |
|---|---|---|---|
| Cloud Instances (12 weeks) | 20 | $249 | $4,980 |
| Jetson Edge Kits | 20 | $700 | $14,000 |
| Unitree Go2 (Shared) | 2 | $3,000 | $6,000 |
| Total for 12-week course | $24,980 |
Advantages
- Low Initial Investment: No hardware purchase needed
- Scalability: Easy to add/remove students
- Remote Learning: Students work from anywhere
- Latest Hardware: Always access to newest GPU instances
- No Maintenance: Cloud provider handles infrastructure
Disadvantages
- High Recurring Costs: $250/student/semester
- Latency Issues: Cannot control physical robots from cloud in real-time
- Internet Dependency: Requires stable, high-bandwidth connection
- Data Transfer Costs: Large simulation files expensive to download
- Limited Customization: Restricted to cloud provider's offerings
The Latency Trap (Hidden Cost)
Critical Issue
Simulating in the cloud works well, but controlling a real robot from a cloud instance is dangerous due to latency (50-200ms delays).
Solution: Students train in the cloud, download the model weights, and flash them to the local Jetson kit for physical deployment.
When to Choose Cloud-Native
- You need rapid deployment (start in days, not months)
- You have operational budget but limited capital budget
- Students are remote or have weak personal laptops
- Course is experimental or pilot phase
- You want to avoid hardware maintenance overhead
Option 3: Hybrid Approach (Recommended)
Best balance: Cloud for simulation, on-premise for physical deployment.
Architecture
┌─────────────────────────────────────┐
│ Cloud: Training & Simulation │
│ - AWS g5.2xlarge instances │
│ - Isaac Sim, Gazebo, Unity │
│ - Model training (RL, VLA) │
└──────────────┬──────────────────────┘
│ Download Models
▼
┌─────────────────────────────────────┐
│ On-Premise: Physical Deployment │
│ - Jetson Edge Kits (20 units) │
│ - RealSense Cameras │
│ - Shared Robots (4 Unitree Go2) │
└─────────────────────────────────────┘
Cost Breakdown (20 Students)
| Item | Cost | Notes |
|---|---|---|
| Cloud Instances (12 weeks) | $4,980 | Simulation and training |
| Jetson Edge Kits | $14,000 | Physical AI deployment |
| Shared Robots (4 units) | $12,000 | Hands-on robot interaction |
| Total Hybrid Cost | $30,980 | Best of both worlds |
Advantages
- Eliminates need for expensive workstations ($60k saved)
- Students train in cloud, deploy to physical edge devices
- Scalable simulation, hands-on robot experience
- Lower total cost than pure on-premise
Disadvantages
- Still requires physical lab space for robots
- Complexity of managing two environments
- Students need training on both cloud and edge workflows
Comparison Matrix
| Factor | On-Premise | Cloud-Native | Hybrid |
|---|---|---|---|
| Upfront Cost | $89,000 | $5,000 | $26,000 |
| Per-Semester Cost | $0 | $25,000 | $5,000 |
| Latency | Zero | High (50-200ms) | Low (edge local) |
| Scalability | Fixed (20 seats) | Unlimited | High |
| Maintenance | High | Zero | Low |
| Physical Robots | Yes | Limited | Yes |
| Remote Learning | No | Yes | Partial |
| 3-Year TCO | $89,000 | $150,000 | $56,000 |
Recommendation by Use Case
Choose On-Premise If:
- You have $90k+ upfront budget
- Course will run 3+ years
- Physical robot interaction is mandatory
- Data privacy is critical
Choose Cloud-Native If:
- You need rapid deployment (< 1 month)
- Students are remote/distributed
- You have operational budget ($25k/semester)
- Course is experimental/pilot phase
Choose Hybrid If:
- You want best of both worlds
- Budget is $30-60k total
- You want scalable simulation + physical deployment
- Long-term course with gradual hardware investment
Implementation Roadmap
Phase 1: Pilot (Semester 1)
- Use cloud instances for all students
- Purchase 4-6 Jetson Edge Kits for demos
- Share 1-2 robots among the class
Phase 2: Scale (Semester 2-3)
- Purchase Jetson kits for all students
- Add 2-4 more shared robots
- Continue using cloud for simulation
Phase 3: Maturity (Year 2+)
- Evaluate workstation purchase if cost-effective
- Transition high-usage students to on-premise
- Keep cloud for overflow and remote students
Next Steps
- Review Hardware Requirements for detailed specs
- Check Software Setup for installation guides
- See Student Kit Guide for Jetson configuration