How Much Should a SaaS Spend on Cloud Infrastructure?

The question is usually asked as a ratio: what percentage of revenue should go to cloud? It is a reasonable starting point and a poor stopping point, because the range across healthy SaaS companies is genuinely wide.

Why the benchmarks vary so much

• Workload shape: a data- or inference-heavy product has structurally higher infrastructure cost than a CRUD application, at any level of efficiency.
• Stage: pre-product-market-fit companies over-provision deliberately, because engineering time is scarcer than money.
• Pricing model: seat-based pricing decouples revenue from usage; usage-based pricing couples them.
• Tenancy: per-tenant isolated infrastructure costs far more than a well-designed shared tier — sometimes for good contractual reasons.

Better questions than the ratio

Ask instead: what does it cost to serve one customer, and is that number going down as you grow? Healthy infrastructure economics show declining cost per unit of usage over time, because fixed platform costs amortise and efficiency work compounds. A flat or rising cost per customer means you are scaling linearly with infrastructure — which caps your gross margin.

The three numbers worth tracking

• Cost to serve per active customer or per unit of usage, trended monthly.
• Non-production spend as a share of total — above roughly a quarter usually indicates environments nobody turns off.
• Spend growth rate versus usage growth rate; the gap between them is your efficiency trend.

When spend is too low

Underspending is a real failure mode. Single-AZ databases, no staging environment, no observability tooling and undersized instances all save money until the first serious incident. Reliability has a price, and paying it deliberately is cheaper than paying it during an outage.

What to do with the answer

If your spend is uncomfortable, the useful next step is attribution, not a target. Once you can see spend by service, environment and team, the conversation stops being about a percentage and becomes a set of specific engineering decisions with numbers attached.