Cloud Cost Optimization: FinOps Basics, Right-Sizing

Cloud bills grow silently until they consume 15-25% of revenue and everyone is surprised.

Cloud Cost Optimization: Where the Money Actually Goes, and How to Cut It

Cloud cost optimization (FinOps) is the practice of continuously reducing infrastructure spend without sacrificing performance or reliability. It sits between engineering, finance, and product. At startup scale the goal is not sophisticated allocation models — it is finding the 5-10 line items that account for 80% of spend and attacking them directly. Every 1% of revenue saved here is 1% of gross margin, which is more valuable than the equivalent in top-line growth for a fundraise.

First: know where the money goes

Before optimizing, get to a monthly breakdown by service, environment, and (ideally) product feature. AWS Cost Explorer / Billing Console, GCP Billing, Azure Cost Management give you the raw data; tools like Vantage, CloudZero, and Infracost overlay allocation and forecasting. Tag every resource by environment (prod/staging/dev), service, and team from day one — untagged resources are almost always the source of mystery bills six months later. If tagging is broken, fixing that is the prerequisite to every other optimization.

The usual top-five line items

For most SaaS startups: (1) Compute (EC2/GCE/App Engine) — usually over-provisioned; right-sizing saves 20-40%. (2) Data egress — cross-AZ, cross-region, and to-internet transfer; can be 15-25% of AWS bills at scale. (3) NAT Gateway — a common surprise line item; consolidate or replace with VPC endpoints where possible. (4) Data warehouse (Snowflake/BigQuery/Redshift) — pay-per-query surprises from unrestricted analytics access; add per-role quotas. (5) Managed databases (RDS/Aurora) — overprovisioned storage and IOPS; enable autoscaling storage, review IOPS quarterly. Attack the top item first; do not spread effort thin across ten small ones.

Right-sizing without breaking things

The temptation is to shrink instance sizes aggressively and hope. The discipline: measure actual CPU, memory, and IOPS utilization over 14 days at p95, size for that plus 30% headroom, use autoscaling for burst. Tools: AWS Compute Optimizer, GCP Recommender, Datadog Cloud Cost. Move stateless workloads to spot/preemptible instances with graceful shutdown handling — 60-80% savings for compatible workloads. Move dev/staging to a scheduled auto-shutdown outside business hours — 60%+ savings on those environments. Never right-size production databases without a load test.

Commitment discounts and Savings Plans

Cloud providers offer 20-70% discounts for committing to spend levels: AWS Savings Plans and Reserved Instances, GCP Committed Use Discounts, Azure Reserved Instances. Rule of thumb: commit to your steady-state baseline (the workload level you are confident you will run for 1-3 years). On-demand covers the variable portion above baseline. Start with 1-year no-upfront commitments while you learn the pattern; move to 3-year and partial-upfront once confident. Do NOT over-commit — an unused commit is a worse outcome than paying on-demand.

Architecture is the biggest lever

Right-sizing and commits are tactical; architecture is strategic. Big multipliers: (1) Multi-tenant vs. per-tenant deployments — per-tenant on managed services is 3-10x more expensive at low utilization; consolidate where the customer contract allows. (2) Cache aggressively — every DB read served from Redis/memcache instead of Postgres is 100x cheaper. (3) Batch expensive operations (LLM calls, image processing) via queues to smooth cost. (4) Region choice — us-east-1/us-central1 are cheapest in most providers; other regions carry 10-30% premium. (5) Managed vs. self-hosted trade-off shifts with scale — cheap when small, expensive above a threshold; review annually.

Frequently asked questions

When do we hire a FinOps person?
When cloud spend exceeds ~$1M/year or 10% of revenue, whichever comes first. Before that, a single engineer with 4 hours a week and a shared dashboard is enough. Do not build a FinOps org before the bill justifies it.
How much does AI/LLM inference cost?
At scale it becomes a top-3 line item. Rough per-token pricing keeps dropping (Claude, GPT, open-source via Together/Fireworks) but usage grows faster. Cache identical prompts, use the smallest model that satisfies quality, batch requests where latency permits, and negotiate reserved throughput with providers when monthly spend exceeds ~$50K.
What about egress fees for AI providers?
If you self-host models on GPUs, cross-cloud egress can dominate cost. Co-locate model serving in the same region as the source data. Cloudflare R2 (zero-egress) and similar 'no egress' object stores exist specifically to attack this cost.

Related fundraising guides (40)

Investor directory · Fundraising library · Articles A–Z · Company funding database