deploybase

Serverless vs Reserved GPU Instances: Cost Breakdown

Deploybase · June 26, 2025 · GPU Pricing

Contents

Serverless vs Reserved GPU Defined

Reserved: developers buy capacity by the hour. Pay whether developers use it or not. Serverless: developers rent execution time. Pay only when code runs.

Reserved = consistent workloads. Serverless = intermittent processing.

Cost Structure Analysis

Reserved Instance Costs

H100 PCIe (Lambda Labs): $2.86/hr = $2,087/mo = $25,000/yr. Eight H100s? $16,700/mo. 1-3 year commits get 20-40% discounts.

Serverless GPU Costs

Serverless GPU platforms (RunPod Serverless, Modal, Replicate): typically $0.01-0.025 per request for a 2-5 second inference job. 1,000 requests at $0.025 each = $25/mo. At low request volumes, this is dramatically cheaper than a reserved H100 ($2,087/mo). Note: AWS Lambda does not natively support GPU instances; use AWS SageMaker Serverless or container-based alternatives for GPU serverless workloads.

Break-Even Analysis

Reserved instances win above roughly 200 GPU-hours monthly at high utilization. Below that threshold, serverless pricing is cheaper because you only pay for actual execution time.

Performance Characteristics

Reserved: consistent latency, guaranteed performance, large models ok, batch jobs work great. Serverless: auto-scale, zero idle cost, global distribution, quick deploy. But: 30-60s cold start, model size limits, 15-60 min timeout, shared hardware jitter.

Workload Matching

Reserved Instances

Inference APIs (consistent traffic), training jobs (hours-long runs), chat/code generation (low latency needed), daily batch jobs (predictable schedule).

Serverless GPU

Event-triggered tasks (uploads, messages, webhooks), periodic processing (weekly reports, monthly analytics, quarterly retraining).

Viral content, flash sales, unpredictable traffic? Serverless scales. Prototyping? Serverless is cheap.

Migration Strategies

Hybrid: route baseline to reserved, spikes to serverless. Or: reserved during peak hours, serverless off-peak. Or: inference on reserved, background jobs on serverless. Check AWS pricing and SageMaker options.

FAQ

Can serverless functions use multiple GPUs? No. Reserved instances only.

What if serverless exceeds timeout? Terminates, charges incur. AWS Lambda: 15-min max. Longer jobs need reserved.

Do reserved instances need upfront payment? AWS and Lambda Labs: hourly billing, no upfront. Azure: commitment required.

How fast do serverless functions scale from zero? 30-120 seconds to 1000 concurrent. Reserved instances scale instantly.

Can reserved instances autoscale? Yes. AWS and Azure both support it.

Sources