Course contentsShow
← System Design
Lesson 45 of 1,9192. Scaling FoundationsFree lesson

Comparing Cost Structures

Vertical scaling has exponential cost curves; horizontal scaling offers more linear cost growth.

Comparing Cost Structures

What you'll learn: Why vertical scaling becomes increasingly expensive while horizontal scaling costs grow more predictably.

The Cost Curve Problem

When you choose between vertical scaling (upgrading a single machine) and horizontal scaling (adding more machines), you're also choosing between two very different cost models.

Vertical Scaling: The Exponential Price Tag

Vertical scaling follows an exponential cost curve. The first upgrades are relatively cheap—doubling RAM from 8GB to 16GB might cost $100. But as you push toward the limits of hardware, costs skyrocket. Going from 256GB to 512GB RAM could cost thousands. A top-tier server with maximum CPU, RAM, and storage can cost 10-50x more than a mid-tier machine, yet only provides 2-3x the performance.

Why? High-end hardware requires:

  • Specialized components (error-correcting memory, enterprise-grade processors)
  • Custom engineering and manufacturing
  • Diminishing returns as you approach physical limits

Horizontal Scaling: Linear Growth

Horizontal scaling offers linear cost growth. If one server costs 100/monthandhandles1,000QPS,then10serverscost100/month and handles 1,000 QPS, then 10 servers cost 1,000/month and handle 10,000 QPS (roughly). Each additional server provides predictable capacity at predictable cost.

The Practical Reality

Most systems use both approaches strategically: vertical scaling for simplicity early on, then horizontal scaling when growth demands it. The crossover point typically occurs when the cost of upgrading one machine exceeds the cost of multiple smaller machines plus the engineering effort to distribute load.

Key Takeaway: Vertical scaling has exponential cost curves that hit a ceiling quickly; horizontal scaling provides more linear, predictable cost growth that scales with your actual needs, making it the economical choice for high-scale systems.