How much will your cloud bill be each month?
Estimate monthly hosting cost for a web app or SaaS product from your users, traffic, data and architecture. It includes the lines startups most often forget: staging environments, observability, backups, and the model usage and GPU time that AI features now add to cloud bills.
Size your cloud bill
Start from a preset, then describe your traffic, architecture and AI features. Key unit rates are editable at the end.
- AI models and GPUs$1,600
- Database$230
- Bandwidth and CDN$210
- Staging and preview$110
- Observability$71
- Compute$60
- Networking and gateways$55
- Storage and backups$14
- Single server$2,300
- Managed containers$2,400
- Serverless$2,300
- Kubernetes$2,600
- At 10 times the users (200,000), this setup would cost about $19,000 a month, or $0.09 per user.
- This configuration is already lean. Set budget alerts and review the bill monthly so growth does not surprise you.
- AI features account for 68% of the bill ($1,600 a month), about $0.08 per active user.
- Routing all model calls to a small fast model would bring model spend to about $240 under these assumptions. In practice, route routine calls to a small model and escalate the rest.
- Cache repeated prompts and responses, and cap usage per user, so a few heavy users cannot drive the bill.
Assumes peak traffic at 5 times the average, 25% extra compute for background jobs, and generic unit prices. Excludes support plans, taxes, third-party SaaS and engineering time.
Indicative only. We will send the inputs above with your message so a senior engineer can sanity-check them.
What really determines SaaS hosting cost
Compute is rarely the only surprise. These are the categories the estimator adds up, and where they tend to catch teams out.
Compute and architecture
Always-on servers are sized for peak traffic; serverless charges per request. Kubernetes adds cluster overhead and ops time before it saves anything.
Database and high availability
A standby replica roughly doubles database cost. It is still worth it once customers depend on you, but it is a deliberate choice.
Bandwidth and regions
Media-heavy products pay mostly for data leaving the cloud. Every extra region repeats compute and adds database replicas.
Environments nobody turns off
Staging and per-pull-request preview environments are valuable, and often run all night and all weekend with nobody using them.
Observability and backups
Logs, traces and long retention grow with traffic. Point-in-time backups and off-site copies grow with data. Both are easy to underestimate.
AI models and GPUs
LLM calls scale with how much each user engages with AI features, not just user count. A single always-on GPU can cost more than the rest of an early-stage stack.
How the cloud cost estimate is built
The estimator turns monthly active users and requests per user into peak requests per second, assuming the busiest hour runs at five times the average. For server-based architectures, that peak sets how many vCPUs you need per region, plus headroom for background jobs and failover. Serverless is priced per million requests instead. Database size follows both traffic and data volume.
Every unit price is a generic, editable default rather than any provider’s price list, because real rates vary by provider, region, instance family and negotiated discounts, and they change often. Replace the defaults with your own rates and the whole estimate updates. Commitment discounts apply only to steady compute, database compute and GPUs, as they do in practice.
AI costs are modelled from the number of language model calls per user, a typical token count per call and a blended price per million tokens for the model tier you choose. Embeddings add a vector index that grows with your data, and dedicated GPUs are priced as always on, which is how most teams first run self-hosted models.
The range runs from 15% below to 30% above the mid estimate, since real bills more often surprise on the high side. It excludes support plans, taxes, third-party SaaS and, importantly, the engineering time needed to run everything.
Single server, containers, serverless or Kubernetes?
The cheapest architecture on paper is not always the cheapest to run. Ops effort and reliability belong in the comparison.
| Cost profile | Ops effort | Good fit | |
|---|---|---|---|
| Single server | Lowest fixed cost, sized for peak | Low, but outages on every restart | MVPs and internal tools |
| Managed containers | Moderate, scales with load | Low to moderate | Most SaaS products from launch to scale |
| Serverless | Near zero when idle, rises per request | Lowest day to day | Spiky or low, uneven traffic |
| Kubernetes | Cluster overhead before any traffic | High, needs platform skills | Large teams with many services |
General characteristics, not a ranking. A good architecture review looks at your traffic shape and team skills as well as the bill.
Cloud cost optimisation habits worth building early
Tag resources and review the bill monthly
Costs you cannot attribute to a product, team or environment are costs nobody will reduce.
Schedule non-production environments
Shutting staging and previews down outside working hours is one of the simplest savings available.
Commit only to your stable baseline
Reserved capacity and savings plans pay off on predictable load. Leave growth and experiments on demand.
Route AI calls by difficulty
Send routine prompts to small fast models, cache repeated answers and escalate only hard requests to frontier models.
Set per-user limits on AI features
A handful of heavy users can drive most of your model spend. Usage caps and fair-use limits keep unit economics intact.
Question every always-on GPU
Unless a GPU is busy most of the day, API-based inference or scale-to-zero GPU hosting is often cheaper.
Questions we often hear
How much does cloud hosting cost for a startup?
An early MVP with a few thousand users commonly runs from under $100 to a few hundred dollars a month. A growing B2B SaaS product with high availability, staging and proper monitoring often lands in the low thousands. AI features, media-heavy traffic and multiple regions can multiply those figures quickly.
What is a reasonable hosting cost per user for SaaS?
For a typical B2B SaaS product without heavy AI or media use, hosting often works out to a few cents to a few tens of cents per monthly active user. What matters more is that the cost per user falls, or at least holds steady, as you grow. If it rises, your architecture or AI usage needs attention.
How do AI features change cloud costs?
They add a cost that scales with engagement rather than infrastructure. Every language model call has a token price, embeddings need vector storage, and self-hosted models need GPUs that are expensive to leave running. It is now common for AI to be the largest line on a product’s bill, so model choice, caching and per-user limits matter.
Is serverless cheaper than containers?
At low or spiky traffic, usually yes, because you pay almost nothing when idle. At sustained high traffic, per-request pricing can overtake always-on containers. Compare both at your expected volume and at ten times that volume, which is what the estimator’s architecture bars show.
Are these AWS, Azure or Google Cloud prices?
No. The estimator uses generic component prices so it stays useful across providers and does not go out of date when price lists change. Replace the unit rates with your provider’s current prices for a closer figure, and use the provider’s own calculator before committing to a budget.
Related reading and tools
Make sure the bill scales with you.
Send us your estimate or your current bill. We will review the architecture, point out where spend is likely to grow faster than revenue, and suggest practical changes, including how to keep AI and GPU costs under control.
Working with companies globally · Response within 24 hours