$7 million monthly AWS bill. $11 million annual revenue.
That was Stability AI’s reality in 2024. The AI startup behind Stable Diffusion couldn’t pay their AWS bills despite massive funding rounds. They’re not alone – Figma spends $300,000 daily on AWS (that’s $100 million annually), and even successful companies struggle with cloud economics.
Here’s what’s really happening with your cloud costs. And what you can do about it.
Why Cloud Costs Explode
“Pay only for what you use” sounds great.
Until you realize what you’re actually using.
A simple web application with a database? Let’s count the real AWS costs. RDS db.t3.medium PostgreSQL in us-east-1: $52/month. Add Multi-AZ for production: $104/month. But wait. You need backups. Automated backups storage. Read replicas maybe. Now we’re at $200+/month just for the database.
Your EC2 instances? t3.medium at $30/month. But add the Application Load Balancer: $18/month plus $0.008 per LCU-hour. NAT Gateway for your private subnets: $33/month plus $0.045 per GB processed. EBS volumes: $0.08 per GB-month for gp3.
Suddenly your “simple” app costs $500/month before you even have any traffic.
The hidden multipliers are everywhere. Data transfer between availability zones: $0.01 per GB. CloudWatch Logs ingestion: $0.50 per GB. Route 53 health checks: $0.50 per health check per month. S3 API requests: $0.0004 per 1,000 requests.
These seem like small numbers until they suddenly aren’t.
Common Waste Patterns in the Wild
Development environments as production clones.
AWS actually recommends separate accounts per developer for isolation, but many startups implement this poorly. Instead of lightweight dev environments, they clone entire production stacks – RDS clusters, ElasticSearch, Redis. AWS’s own best practices suggest shared resources or scheduled environments that shut down after hours, saving up to 65% on costs. Yet teams keep running full clones 24/7.
Database instances sized for Black Friday that never comes.
RDS db.r6g.2xlarge with 8 vCPUs and 64 GB RAM: $570/month. Current connection count: 7. CPU usage: 3%. Memory usage: 2GB.
But “we might need it for scaling.”
Zombie resources accumulating for years.
Elastic IPs not attached to anything: $3.60/month each. Old AMIs with their snapshot storage. Forgotten EBS volumes from terminated instances. CloudWatch Logs that never expire. Load balancers from failed experiments.
Fathom Analytics discovered they were wasting $7,000/year on CloudWatch logs they never used, plus another $4,800/year on unnecessary CloudFront costs in EU regions. These zombie resources accumulate silently.
Multi-region setup for single-country startups.
Full disaster recovery in another region. Complete infrastructure duplication. For a B2B SaaS with 30 customers in California.
Cross-region data transfer alone: $0.02 per GB. Database replication, backup copies, everything doubles.
Petabytes of logs nobody reads.
CloudWatch Logs ingestion at $0.50 per GB. Application logs, access logs, error logs, debug logs. Everything at maximum verbosity.
A Stack Overflow user reported paying over $1,000/month for 2TB of CloudWatch logs required for PCI compliance – logs they had to retain but rarely accessed. The compliance checkbox cost them $12,000 annually.
Practical Optimization Strategies
Start simple.
Turn things off. Development environments should sleep at night. Lambda function with EventBridge to stop instances at 7 PM, start at 9 AM. Saves 58% on development costs immediately.
Right-size without the paralysis. Don’t need a month-long analysis. Look at CloudWatch metrics for the last 2 weeks. CPU under 20%? Drop an instance size. Memory usage under 50%? Drop another size.
t3.xlarge ($120/month) to t3.large ($60/month) to t3.medium ($30/month). Same performance for most applications.
Reserved Instances vs Savings Plans vs Spot.
One-year no-upfront Reserved Instances: 31% discount. Three-year no-upfront: 53% discount. But only for instances you’ll definitely run continuously.
Savings Plans more flexible but less discount. EC2 Instance Savings Plans: up to 72% off. Compute Savings Plans: up to 66% off.
Spot instances for batch processing: up to 90% off. But can be terminated anytime.
Rule of thumb: Reserved for databases, on-demand for application servers, spot for workers.
When to abandon ship.
Hetzner AX41: €39/month for 64GB RAM, 1TB NVMe storage. DigitalOcean droplet: $48/month for 8GB RAM, 160GB SSD. Vultr: $20/month for 4GB RAM, 80GB SSD.
Compare to AWS: t3.xlarge with 16GB RAM: $121/month. Plus EBS, plus bandwidth, plus everything else.
Sometimes the cloud isn’t the answer.
Architecture is Your Real Problem
Microservices sound modern. But every service needs its own infrastructure.
10 microservices. Each needs:
- ALB target group
- ECS tasks or K8s pods
- CloudWatch Logs stream
- X-Ray tracing
- Secrets Manager entries
- Parameter Store configs
Multiply everything by 10. Your $500/month monolith becomes $5,000/month distributed system.
Vercel discovered this the hard way. They slashed AWS Lambda costs by 95% just by reusing idle instances instead of spinning up new ones for each request. At billions of invocations, the architectural waste was costing them millions.
The NAT Gateway robbery.
NAT Gateway: $33/month plus $0.045 per GB processed. Every private subnet needs one for high availability. Three availability zones? $99/month just for NAT Gateways. Reddit users report actual bills from $2,745 to $9,600 monthly for NAT Gateways alone.
Alternative: NAT instances on t3.nano: $3.80/month each. Same functionality for small-scale applications.
Managed services vs self-hosted reality.
RDS PostgreSQL db.t3.medium: $104/month with Multi-AZ.
Self-managed PostgreSQL on t3.medium: $30/month.
Yes, you lose automated backups, failover, patching. But for early-stage startups? Often worth the trade-off.
ElasticSearch Service: starts at $80/month for t3.small.elasticsearch.
Self-hosted OpenSearch on t3.medium: $30/month.
Amazon MQ (managed RabbitMQ) m5.large: $230/month.
RabbitMQ on t3.small: $15/month.
The “managed” tax is real.
Why boring architecture saves money.
Single region. Single availability zone for development. Multi-AZ only for production RDS.
Monolithic application on a few larger instances instead of many small ones. PostgreSQL for everything until it truly can’t handle the load. Redis only when you have real caching needs.
No Kubernetes until you have a dedicated DevOps person. No service mesh ever (unless you’re Netflix). No multi-region until you have customers there.
Use S3 for files, CloudFront only if you need it, and SQS for queues - that’s all you really need.
The Reality Check
Consider these real-world examples:
37signals spent $3.2 million annually on cloud services. After moving to on-premise servers, they’re saving $2 million per year and expect to save over $10 million over five years. Their entire hardware investment of $700,000 was recouped in the first year.
Canva manages 230 petabytes on AWS but optimized aggressively. They saved $3.6 million annually just on S3 storage by moving infrequently accessed data to Glacier. They reduced compute costs by 46% through smart purchasing strategies.
Pinterest got hit with a $20 million overage during one holiday season when traffic spiked beyond projections. Adobe’s AWS bill jumped 64% year-over-year. These aren’t failing companies – they’re successful businesses struggling with cloud economics.
Cloud costs reflect architectural decisions.
Complex architecture means complex costs. Distributed systems come with distributed bills. Each managed service adds a monthly tax. Availability zones multiply everything. And regions? They double your costs.
Most startups die from running out of money, not from downtime.
Your massive AWS bill isn’t a success metric - it’s a warning sign. According to Flexera’s 2024 State of Cloud Report, 30% of cloud spending is wasted on average.
The solution isn’t better cost monitoring tools, FinOps processes, or reserved instance planning.
It’s simpler architecture.
Boring technology choices. Monolithic applications. Single-region deployments. Vanilla PostgreSQL. Basic compute instances.
You’re not Amazon. You don’t need Amazon-scale infrastructure.
What you need is to survive until product-market fit. And excessive cloud bills don’t help with that.
Keep it simple, boring, and cheap.
Your CFO will thank you. Your investors will thank you.
And you might actually build a business instead of a cloud bill.
Sources and Further Reading
- Stability AI’s AWS Crisis - The Register, April 2024
- Figma’s $300k Daily AWS Bill - Duckbill Group, 2024
- 37signals Cloud Exit - The Register
- How Canva Saves Millions - AWS Blog
- Fathom’s $100k Cost Reduction - Fathom Analytics
- Vercel’s Lambda Cost Optimization - The Register, July 2024
- AWS EC2 Pricing - Official AWS Documentation
- Flexera State of Cloud Report 2024 - Flexera
Cloud costs eating your runway? I help startups optimize their AWS spend without compromising reliability.