Overview

You're planning for next year's infrastructure needs. How many database servers do you need? How much network bandwidth? How much CPU? You make a guess. Sometimes you're right. Usually you're over-provisioning for some things and under-provisioning for others.

Money gets wasted on infrastructure you don't use. Or performance suffers because you didn't allocate enough. Neither is ideal.

AI can help you model this more accurately. Not by predicting the future (no one can), but by exploring scenarios and helping you think through bottlenecks before they happen.

Why Capacity Planning Is Hard

You have too many variables. Your system has hundreds of components. Each component has capacity constraints. They interact in complex ways. Scale one thing up, and suddenly something else becomes the bottleneck.

Example: You scale your API servers 10x to handle more traffic. But now the database is overwhelmed. You scale the database. But now the cache hit rate drops. You add more cache. But now memory usage spikes on your caching layer. You scale that. But now network bandwidth becomes the bottleneck.

Add business uncertainty on top: How many new features will you ship? Which ones will be popular? How will customer usage patterns change? Will enterprise customers with high data needs join? No one knows.

The result: capacity planning is mostly guesswork with a spreadsheet.

The Cost of Being Wrong

Under-provision: Systems get slow, customers get upset, you lose revenue. Over-provision: You pay for infrastructure you don't use, your margins drop, investors get upset.

Most organizations err on the side of over-provisioning (safer politically) and waste 20-40% of infrastructure cost.

How AI Changes Capacity Planning

Scenario Modeling

Instead of one forecast, build multiple scenarios: optimistic (10x usage growth), realistic (2x growth), conservative (1.2x growth). For each scenario, model: Where are the bottlenecks? What breaks first? When does it break?

AI can help you build these models. You give it your current system topology, usage patterns, and growth assumptions. It simulates: "If you scale to 10x users with current architecture, here's what happens: API server CPU hits 85% utilization in month 8. Database query latency hits 1 second in month 10. Cache miss rate increases from 5% to 25%. At month 12, you need to upgrade the database."

Example Scenario Analysis

Conservative scenario (1.2x growth): All components stay in healthy range for 18 months. You can coast. Invest elsewhere.

Realistic scenario (2x growth): Database becomes bottleneck in 12 months. Database CPU hits 80%. Query latency increases. You should plan to upgrade in month 9-10, giving time for procurement and testing.

Optimistic scenario (10x growth): Everything becomes bottleneck in 4-6 months. API servers, database, cache, network all need upgrades. You need a serious capacity plan.

Bottleneck Analysis

Your system has a critical path. Some components matter more than others. Your bottleneck is what prevents you from scaling. Spend money on the bottlenecks, not on everything equally.

AI can analyze your architecture and identify the critical path. "You're over-provisioning web servers (you have 3x headroom, they're never above 30% utilization) but under-provisioning the database (you're at 70% capacity, growing 15% per month). The database is your bottleneck. That's where to invest." A Series B fintech company did this: they were spending $80k/month on infrastructure. AI analysis showed: API servers were 20% utilized, database was 75% utilized. They were spending $20k/month on unnecessarily large API servers. They right-sized to $12k/month and invested the $8k/month savings in database optimization. Same performance, $96k/year in savings.

Bottleneck Identification Framework

For each component: measure utilization under realistic load. Components at > 70% utilization are bottlenecks. Components at
The Key Insight: You can't predict the future. But you can model the present and extrapolate based on different growth scenarios. AI helps you do this systematically, testing what-if scenarios, rather than using spreadsheets and gut feel. The difference is having a plan vs. reacting in a panic when things break.

Implementing AI-Driven Capacity Planning

Step 1: Model Your Current System

Document your architecture: What components do you have? What are their capacities? How are they related?

This sounds like a lot of work, but you probably already have this in architecture diagrams. Get it into a format AI can analyze.

Step 2: Define Your Constraints**

• Budget constraints: How much can you spend per year?
  - Performance constraints: What latency/throughput do you need?
  - Availability constraints: What uptime SLA do you need?
  - Operational constraints: Can you handle complex infrastructure?

Step 3: Build Scenarios**

Create 3-5 scenarios: conservative growth, realistic growth, aggressive growth. For each, estimate: users, traffic, data volume.

Step 4: Run Analysis**

Feed your model and scenarios to AI. Ask: "For each scenario, where are the bottlenecks? What breaks first? What do we need to invest in?"

Step 5: Plan Investments**

Based on the analysis, plan your capacity investments. "In scenario 1, we need to upgrade the database in 12 months. In scenario 2, we need to upgrade in 8 months. In scenario 3, we need it in 4 months. Let's plan conservatively and assume scenario 2."

Step 6: Continuous Monitoring**

As the year progresses, compare actual growth to your scenarios. Are you tracking conservatively, realistically, or aggressively? Adjust your plans accordingly.

What Good Capacity Planning Looks Like

You Know Your Bottlenecks

You can answer: "What's going to break first if we 5x our users?" Database? API servers? Storage? You know. And you've planned for it. You don't wait for Black Friday or product launch to discover your bottleneck.

You're Not Over-Provisioning

You're not spending money on infrastructure you don't use. You have enough headroom for growth, but you're not paying for 3 years of unused capacity. Most companies over-provision by 20-40%; good capacity planning cuts that to 10-15%.

You Can Answer "What If" Questions

Product asks: "What if we add this new feature that will increase traffic by 30%?" You can model it immediately. "We're fine for 6 months, then we need to upgrade the cache layer. Budget it for Q2." No surprises.

You Have a Growth Plan

You're not upgrading reactively when things break. You're upgrading proactively, on a schedule, with time to test and validate. Reactive upgrades cause mistakes and outages. Proactive upgrades are smooth.

Failure Mode to Avoid: A team did capacity planning once, got comfortable with their plan, then didn't revisit for a year. Real growth exceeded their aggressive scenario. Their database hit 95% capacity with no warning. They had a week to panic-upgrade. Lesson: quarterly reviews are mandatory. Growth doesn't follow forecasts. Monitor constantly.

Case Study: Fintech Company Prevents Infrastructure Crisis

A fintech company using AI capacity planning caught a critical bottleneck 6 months early. Their transaction database was on track to hit capacity in Q3. By catching it in Q4 of prior year, they: (1) Negotiated hardware procurement (4 week lead time), (2) Tested the upgrade in staging environment, (3) Planned a smooth migration with zero downtime. If they'd missed this, hitting capacity unexpectedly in Q3 would have caused an emergency, outage costs, customer churn, regulatory penalties worth $2M+. The capacity planning exercise took 2 weeks, cost $15k. Prevented value: $2M+.

What to Do Monday Morning

• Document your current infrastructure and their current utilization (measure from your monitoring tools)
  - Estimate your growth over the next 18 months (conservative 1.2x, realistic 2-2.5x, aggressive 5-10x)
  - List the top 5 components by importance and sensitivity
  - Use Claude to model: For each scenario, what hits capacity first? When? What happens?
  - For your top 3 bottlenecks, research upgrade options, costs, and lead times (procurement, testing, deployment)
  - Create a capacity plan: "Q2 we'll upgrade X (4 week lead time), Q3 we'll upgrade Y (2 week lead time), Q4 we'll upgrade Z"
  - Set up quarterly reviews: compare actual growth to your scenarios. Adjust plan if needed.

FAQ

Q: What if our growth is unpredictable?

A: That's fine. Model multiple scenarios anyway. The point isn't accuracy; it's knowing where your constraints are so you can react faster when things change. Unpredictable growth is why you have conservative, realistic, and aggressive scenarios.

Q: How often should we re-plan?

A: Quarterly. Check if you're tracking the scenario you expected. If you're exceeding it, adjust your plans and accelerate upgrades. If you're below it, maybe you can delay upgrades and save money.

Q: What if we use cloud infrastructure that auto-scales?

A: Capacity planning is still valuable. You want to know: how much will auto-scaling cost? What's the maximum monthly bill if you do 10x? What will your bill be under different scenarios? Cloud auto-scaling is great for coping with peaks, but you still need to understand your cost trajectory.

Q: Can we use historical data to improve forecasts?

A: Yes. If you've been tracking metrics for a year, you have growth rate trends. Use that as the basis for your realistic scenario. But: past growth rate doesn't guarantee future growth rate. New features, sales campaigns, or market changes can change everything. Use historical data to inform, not determine, your scenarios.

Q: What if the AI's forecast is wrong?

A: It probably will be, at least partially. The goal isn't to predict the future accurately; it's to explore scenarios and identify where your constraints are. Even an imperfect model is better than a spreadsheet and gut feel. Refine the model quarterly as you get more data. The best approach: use conservative scenario for actual planning, realistic for monitoring, aggressive for knowing when to panic.

Real Examples of Capacity Planning Impact

Example 1: Database Bottleneck (Caught Early) A fintech company used AI capacity modeling and discovered their PostgreSQL database would hit capacity limits in 14 months under realistic growth. They planned a 12-month migration to a horizontally-scalable database, completing before the crisis hit. Without planning, they would have discovered this in a production outage, costing 8 hours of downtime and $500k in lost revenue. The 2 months of migration planning work saved them from that disaster.

Example 2: Cache Over-Provisioning (Cost Savings) An e-commerce company discovered through capacity analysis they were over-provisioning caching (Redis) by 3x. They were paying $60k/month for cache utilization at only 25%. They right-sized cache to $20k/month, then invested the $40k/month savings in database optimization, reducing query load by 40%. Same performance, $480k/year in cost savings, through systematic capacity analysis instead of "use big numbers for safety."

Example 3: Multi-Component Optimization A SaaS company modeled growth scenarios and discovered: (1) API servers hit capacity at 6x users, (2) database at 8x users, (3) cache at 4x users. Cache was the tightest constraint. Instead of scaling everything equally (wasting money on oversizing API servers and database), they invested in cache optimization. This bought them 2x the growth runway in the bottleneck component, enabling them to hit 8x user growth before infrastructure investment was needed.

Key Insight

Good capacity planning means knowing your bottlenecks before they become problems, investing in the right things, and growing without over-provisioning. AI helps you model scenarios and identify bottlenecks faster than spreadsheets ever could. The teams that do this well grow faster (better performance), cheaper (no over-provisioning), and smoother (no production crises). It's not glamorous work, but it's foundational infrastructure thinking.

On This Page

Watch the Lecture
Why Capacity Planning Is Hard
How AI Changes It
Implementing
What Good Looks Like
Monday Morning Action
FAQ
Real Examples
Key Insight

Chapter Details

Part ofChapter 3