A Series A fintech startup came to us with a problem that was eating their runway: an AWS bill that had grown from 2 lakhs to 8 lakhs per month over six months, with no corresponding increase in traffic or revenue. Their CTO knew the bill was bloated but did not have the bandwidth to investigate while also shipping features. We conducted a two-week cloud cost audit and implemented changes that brought the bill down to 3.2 lakhs — a 60% reduction — while actually improving application performance.
The biggest offender was over-provisioned compute. They were running six m5.2xlarge EC2 instances (8 vCPUs, 32GB RAM each) in an auto-scaling group that never scaled below four. CPU utilisation averaged 12% across the fleet. We right-sized to four m5.large instances (2 vCPUs, 8GB RAM) with auto-scaling rules that actually trigger on load, and purchased reserved instances for the baseline capacity. This single change saved 3.5 lakhs per month. The lesson is universal: most startups provision for the traffic they hope to get, not the traffic they actually have.
The second largest cost was an oversized RDS instance running PostgreSQL with Multi-AZ replication, provisioned IOPS, and 1TB of allocated storage — of which only 40GB was in use. We moved the database to Supabase's Pro plan, which gave them a managed PostgreSQL instance with connection pooling, automatic backups, and a generous free tier of API calls, for a fraction of the RDS cost. The migration took one weekend with zero downtime using pg_dump and logical replication.
Architectural changes delivered the remaining savings. We moved static assets to Cloudflare R2 (zero egress fees) instead of S3 with CloudFront, replaced a Redis ElastiCache cluster with Upstash's serverless Redis (pay-per-request), and converted three always-on background worker services into scheduled Lambda functions that run only when needed. The combined effect was not just lower costs but better architecture — the system is now simpler, easier to operate, and more resilient to traffic spikes because the serverless components scale automatically.