Skip to main content

Cloud vs. On-Premise Infrastructure

The Infrastructure Decision​

Building Physical AI systems requires significant computational resources. You have two main options:

  1. On-Premise (High CapEx): Buy hardware upfront
  2. Cloud (High OpEx): Rent compute by the hour

This chapter helps you choose based on budget, usage patterns, and requirements.

Cost Comparison​

Scenario: 13-Week Course (1 Student)​

On-Premise Option​

Initial Investment:

  • Workstation (RTX 4080, 64GB RAM): $3,500
  • Jetson Orin Nano kit: $700
  • Total: $4,200 upfront

Quarterly Cost: $0 (after initial purchase) Electricity: ~$20-30/quarter (negligible)

Total Course Cost: $4,200

Amortized over 3 years (6 quarters):

  • $4,200 / 6 = $700/quarter
  • Plus hardware retains resale value (~$2,000 after 3 years)

Cloud Option​

Simulation (AWS g5.2xlarge):

  • Instance: $1.50/hour
  • Usage: 10 hours/week × 13 weeks = 130 hours
  • Cost: 130 × $1.50 = $195

Storage (EBS):

  • 500GB for datasets/models: $50

Edge Deployment (Still need Jetson):

  • Jetson Orin Nano kit: $700

Total Course Cost: $945 ($195 + $50 + $700)

Breakeven Analysis​

On-premise breakeven: 4-5 quarters of usage

Quarter 1: Cloud cheaper ($945 vs. $4,200)
Quarter 2: Cloud cheaper ($1,890 vs. $4,200)
Quarter 3: Cloud cheaper ($2,835 vs. $4,200)
Quarter 4: Tie ($3,780 vs. $4,200)
Quarter 5+: On-premise cheaper

Conclusion:

  • Short-term (1-2 quarters): Cloud is cheaper
  • Long-term (3+ quarters): On-premise is cheaper

Detailed Comparison​

On-Premise Workstation​

Advantages​

  1. No ongoing costs: Pay once, use forever
  2. Full control: No usage limits, no throttling
  3. Low latency: Local access, no network delay
  4. Privacy: Data stays on your machine
  5. Offline capable: No internet dependency
  6. Resale value: Hardware retains 30-50% value after 3 years

Disadvantages​

  1. High upfront cost: $3,500-4,000
  2. Maintenance: You handle hardware failures
  3. Limited scalability: Can't easily add more GPUs
  4. Depreciation: Hardware loses value over time
  5. Space/power: Requires desk space, reliable power

Best For​

  • Students planning 2+ years of robotics work
  • Labs with multiple students (shared resource)
  • Developers working on long-term projects
  • Those with reliable power and space

Cloud Computing​

Advantages​

  1. Low barrier to entry: No upfront cost
  2. Scalability: Spin up 10 GPUs for parallel training
  3. Flexibility: Pay only for what you use
  4. No maintenance: Provider handles hardware
  5. Latest hardware: Upgrade to new GPUs easily
  6. Global access: Work from anywhere

Disadvantages​

  1. Ongoing costs: $200-300/quarter adds up
  2. Usage limits: May need to manage quotas
  3. Latency: Network delay for data transfer
  4. Internet dependency: Unusable offline
  5. Data egress costs: Downloading large datasets expensive
  6. No physical robot control: Can't control real robot from cloud (latency)

Best For​

  • Short-term projects (1-2 quarters)
  • Students with budget constraints (can't afford $4k upfront)
  • Those needing massive parallelization (RL training)
  • Travelers without permanent workspace

Cloud Provider Comparison​

AWS (Amazon Web Services)​

GPU Instances:

  • g5.xlarge: 1x A10G (24GB VRAM), 4 vCPUs, 16GB RAM - $1.01/hour
  • g5.2xlarge: 1x A10G (24GB VRAM), 8 vCPUs, 32GB RAM - $1.21/hour
  • g5.12xlarge: 4x A10G (96GB VRAM), 48 vCPUs, 192GB RAM - $5.67/hour

Storage:

  • EBS (gp3): $0.08/GB/month
  • S3: $0.023/GB/month (cheaper for datasets)

Pros:

  • Largest selection of instance types
  • Spot instances (up to 90% discount)
  • Mature ecosystem (AMIs for Isaac Sim, ROS 2)

Cons:

  • Complex pricing (data transfer, storage)
  • Steeper learning curve

Google Cloud Platform (GCP)​

GPU Instances:

  • n1-standard-4 + T4: 1x T4 (16GB VRAM), 4 vCPUs, 15GB RAM - $0.63/hour
  • n1-standard-8 + A100: 1x A100 (40GB VRAM), 8 vCPUs, 30GB RAM - $3.67/hour

Pros:

  • Simpler pricing
  • Good for TensorFlow/JAX (Google tools)
  • Preemptible VMs (80% discount)

Cons:

  • Fewer robotics-specific resources
  • Less Isaac Sim support

Microsoft Azure​

GPU Instances:

  • NC6s_v3: 1x V100 (16GB VRAM), 6 vCPUs, 112GB RAM - $3.06/hour
  • NCads_A100_v4: 1x A100 (80GB VRAM), 24 vCPUs, 220GB RAM - $3.67/hour

Pros:

  • Good Windows support (if needed)
  • Azure Kinect integration
  • Student discounts (Azure for Students)

Cons:

  • More expensive than AWS/GCP for equivalent specs

Cost Optimization Tips​

Use Spot/Preemptible Instances​

Spot instances (AWS) / Preemptible VMs (GCP):

  • 60-90% cheaper than on-demand
  • Can be terminated anytime (when demand is high)
  • Good for training (can checkpoint and resume)
  • Not good for interactive development (annoying to be kicked off)

Example:

  • AWS g5.2xlarge on-demand: $1.21/hour
  • AWS g5.2xlarge spot: $0.36/hour (70% savings)

Stop Instances When Not Using​

Mistake: Leave instance running 24/7

  • Cost: $1.21 × 24 × 30 = $871/month

Correct: Stop instance when done

  • Usage: 10 hours/week = 40 hours/month
  • Cost: $1.21 × 40 = $48/month

Savings: $823/month (94%)

Use S3/Cloud Storage for Datasets​

Don't keep large datasets on expensive EBS:

  • EBS: $0.08/GB/month
  • S3: $0.023/GB/month

1TB dataset:

  • EBS: $80/month
  • S3: $23/month
  • Savings: $57/month

Right-Size Your Instance​

Don't use overkill GPU for simple tasks:

TaskInstance NeededCost
Debugging ROS 2 nodesCPU-only (t3.medium)$0.04/hour
Training simple CNNg5.xlarge (1x A10G)$1.01/hour
Training large VLA modelg5.12xlarge (4x A10G)$5.67/hour

Mistake: Using g5.12xlarge for all tasks → $5.67/hour Correct: Use appropriate instance per task → $0.50/hour average

Best of both worlds:

Phase 1: Development (Cloud)​

  • Use cloud for initial development and experimentation
  • Low commitment, flexibility to try different approaches
  • Cost: ~$50-100 for first month of exploration

Phase 2: Production (On-Premise)​

  • Buy workstation once you're committed to the project
  • Amortize cost over remaining course + future work
  • Use cloud for occasional heavy parallelization

Edge Deployment (Always Local)​

  • Jetson Orin Nano is a one-time $249 purchase
  • Cannot be replaced by cloud (need local robot control)

Decision Matrix​

Your SituationRecommendationCost
1 quarter course onlyCloud + Jetson$945
2-3 quarters of workCloud or on-premise (tie)$1,890-2,835 vs. $4,200
4+ quarters (degree program)On-premise + Jetson$4,200 (cheaper long-term)
Uncertain commitmentStart cloud, switch laterFlexible
Lab (5+ students)On-premise (shared)$4,200 / 5 = $840/student
Need massive parallelizationCloud (scale up/down)Variable
Traveling/no fixed workspaceCloud$200-300/quarter

Cloud Setup Guide (Quick Start)​

AWS Setup for Isaac Sim​

Step 1: Create AWS account

  • Go to aws.amazon.com
  • Sign up (requires credit card)

Step 2: Launch g5.2xlarge instance

# Use NVIDIA Deep Learning AMI (Ubuntu 22.04)
# Instance type: g5.2xlarge
# Storage: 100GB gp3 EBS
# Security group: Allow SSH (port 22)

Step 3: SSH and install Isaac Sim

ssh -i your-key.pem ubuntu@<instance-ip>

# Install Omniverse Launcher
wget https://install.launcher.omniverse.nvidia.com/installers/omniverse-launcher-linux.AppImage
chmod +x omniverse-launcher-linux.AppImage
./omniverse-launcher-linux.AppImage

# Install Isaac Sim via launcher
# (Follow GUI instructions)

Step 4: Install ROS 2

# Follow standard ROS 2 Humble installation
# (See Module 1 notes)

Step 5: Transfer code

# From local machine
scp -i your-key.pem -r my_robot_package ubuntu@<instance-ip>:~/

Step 6: Work

# SSH with X11 forwarding for GUI
ssh -X -i your-key.pem ubuntu@<instance-ip>

# Or use VNC for better performance
# Or run headless (no GUI)

Step 7: Stop instance when done

# In AWS Console: Stop instance (not terminate)
# Or via CLI:
aws ec2 stop-instances --instance-ids i-1234567890abcdef0

Monitoring Costs​

AWS Cost Explorer​

  • AWS Console → Cost Management → Cost Explorer
  • View daily/monthly spending
  • Set budgets and alerts

Example alert:

  • Alert if spending > $50/month
  • Receive email notification

Resource Tags​

Tag resources for cost tracking:

Project: robotics-course
Student: john-doe
Module: isaac-sim

Final Recommendation​

For This Course (13 weeks)​

Option A: Budget ($700)

  • Borrow/use lab workstation for simulation
  • Buy Jetson Orin Nano kit ($700)
  • Deploy to Jetson for edge AI

Option B: Cloud ($945)

  • AWS/GCP for simulation (130 hours × $1.50 = $195)
  • Jetson Orin Nano kit ($700)
  • Storage ($50)

Option C: On-Premise ($4,200)

  • Buy workstation ($3,500)
  • Jetson Orin Nano kit ($700)
  • Best long-term value if continuing robotics

Recommendation​

Start with cloud ($945):

  • Low risk, try before committing
  • After 1-2 months, decide if you want to invest in hardware
  • If you continue, buy workstation (breakeven by month 5)

Next Steps​

You've completed the Hardware & Infrastructure section! Now explore the course assessments.

👉 Next: Assessments →

Cost Saving Pro Tip

Student discounts:

  • AWS Educate: $100 free credits
  • Azure for Students: $100 free credits
  • GCP Education: $50-300 credits
  • NVIDIA Inception: Startups get cloud credits

Total free credits: $250-500 (covers 150-300 hours of GPU time)

Further Reading​

Last updated on