deploybase

GPU Cloud Egress Fees: The Hidden Cost Nobody Talks About

Deploybase · May 8, 2025 · GPU Cloud

Contents

GPU Cloud Egress Fees

GPU cloud egress fees: The hidden killer.

Data leaving the data center costs $0.02-0.30/GB. Download a 50GB model and pay $4.50-$15.

Intra-region free. Cross-region and internet: expensive.

A $2/hour H100 can waste hundreds on egress.

  • Downloading trained models or inference results
  • Uploading datasets to external APIs
  • Syncing checkpoints to S3 outside the region
  • Transferring outputs to client systems
  • Backing up results to separate cloud providers

Egress structure varies dramatically by provider and region. Intra-region transfers often run free or cost 90% less than inter-region moves.

Provider-Specific Charges

AWS egress rates depend heavily on destination and region. Standard pricing:

  • Intra-region transfer: Free (EC2 to S3 in same region)
  • Cross-region transfer: $0.02/GB
  • Internet egress: $0.09/GB (US regions)
  • Internet egress: $0.12-0.15/GB (Europe, Asia)

A single 50GB model download to the internet costs $4.50 from US regions. The same transfer costs $6-7.50 from European data centers.

Google Cloud egress mirrors AWS pricing:

  • Intra-region: Free
  • Cross-region: $0.01/GB
  • Internet egress: $0.12/GB (most regions)

Azure rates differ significantly:

  • Intra-region: Free
  • Cross-region outbound: $0.02/GB
  • Internet egress: $0.087/GB (varies by region)
  • Special offers for Microsoft services reduce costs

RunPod and specialized GPU platforms often include generous egress allowances or offer free inter-provider transfers within their ecosystem.

CoreWeave pricing:

  • Outbound: $0.12/GB (standard)
  • Inbound: Free
  • API transfers: Reduced rates

Vast.AI operates on market-based pricing with user-specified egress rates. Competition keeps costs low, typically $0.03-0.08/GB.

Hidden Costs Beyond Egress

Egress fees represent one layer of hidden charges. Other costs accumulate quickly:

Data transfer within cloud ecosystems: Moving data from GPU instances to storage (S3, Blob Storage, Cloud Storage) incurs charges even within the same region in some cases. AWS charges $0.01/GB for EC2 to CloudFront transfers.

Storage persistence: Keeping training data in cloud storage adds $0.02-0.05 per GB monthly. A 1TB dataset costs $20-50/month just for storage.

Bandwidth for multi-GPU clusters: Intra-cluster communication on some platforms (CoreWeave, Lambda) charges per GB. A synchronization pass across 8 GPUs can cost $2-5.

API request pricing: Cloud providers charge per API call in some cases. Direct Connect and specialized network services cost thousands monthly if used.

Snapshots and backups: EBS snapshots cost $0.05 per GB per month. A 500GB instance snapshot costs $25 monthly to maintain.

Reserved capacity charges: Some providers charge upfront fees ($500-2000) to guarantee GPU availability. Reserved instances commit to multi-year terms.

Real-World Cost Scenarios

Scenario 1: Training and deployment

A team trains an LLM on H100s for one week:

  • Compute: 168 hours x $2/hour = $336
  • Storage (1TB training data): 7 days x $0.04/day = $0.28
  • Checkpoint egress: 50GB x $0.09 = $4.50
  • Final model download: 100GB x $0.09 = $9.00
  • Total: $349.78

Egress accounts for 3.8% of costs but compounds over multiple training runs.

Scenario 2: Inference serving

A startup runs inference at scale for 30 days:

  • Compute: 720 hours x $1.50/hour = $1,080
  • API request handling: 10M requests x $0.0001 = $1,000
  • Data egress to clients: 5TB x $0.12 = $600
  • Storage: 100GB x $0.04/month = $4
  • Total: $2,684

Egress jumps to 22% of monthly spend. Optimization here saves significant money.

Scenario 3: Multi-region failover

A company maintains H100 instances in three regions:

  • Compute (3x): 720 hours x 3 x $2 = $4,320
  • Data replication between regions: 500GB x $0.02 x 3 = $30
  • Client requests egress: 10TB x $0.12 = $1,200
  • Total: $5,550

Egress reaches 22% of total spend in multi-region deployments.

Optimization Strategies

Teams reduce egress costs through architectural choices:

Keep data inside the cloud: Instead of downloading models, serve them directly from cloud storage. AWS Lambda@Edge, Google Cloud CDN, and Azure CDN distribute content without egress charges to edge locations.

Use spot instances strategically: Spot GPU instances cost 50-70% less but add egress risk. Batch results and transfer them together, amortizing per-GB costs.

Compress outputs: Model inference outputs can compress 10:1 with appropriate encoding. Transferring 10GB instead of 100GB saves $9 per operation.

Request batching: Single large transfers cost the same per-GB as many small ones. Batch inference results and model downloads.

Regional clustering: Keeping GPU instances, storage, and users in the same region eliminates inter-region transfer costs.

Dedicated connections: Direct Connect (AWS) or Dedicated Interconnect (Google Cloud) offer fixed monthly costs for high-volume transfers. Break-even occurs around 10-20TB monthly.

Provider-specific features: AWS offers free transfer to CloudFront. Google Cloud provides free transfer to Firebase. Azure reduces charges for Microsoft services.

FAQ

How can I estimate egress costs for my workload? Calculate daily data movement (models downloaded, results uploaded, backups). Multiply by regional egress rates ($0.02-0.12/GB). Factor in compression and batching optimization.

Do intra-region transfers ever cost money? Generally no, but some services charge. AWS EC2 to EFS costs $0.01/GB if not in the same availability zone.

Can I negotiate egress rates with cloud providers? Large teams (10TB+ monthly) can negotiate AWS and Google Cloud rates. Standard discounts reach 20-40%.

What is the cheapest way to move data between cloud providers? Dedicated fiber connections cost $1000-5000/month for sustained transfers. For episodic transfers, compress and batch to minimize per-GB charges.

Do free tiers include egress allowances? AWS free tier includes 100GB outbound per month. Most providers count only a portion toward free tiers.

Sources

  • AWS EC2 Pricing Documentation
  • Google Cloud Compute Pricing (official)
  • Microsoft Azure Virtual Machines Pricing
  • CoreWeave Pricing & Terms
  • Vast.AI Market-Based GPU Pricing Model