How a Series A FinTech Startup Cut Infrastructure Costs by 62% While Scaling to 2M Users

## Overview PayStream, a Series A FinTech platform processing payroll and benefits for small and medium businesses, grew from 120,000 users to over 2 million active users within fourteen months. That growth revealed an uncomfortable truth: the engineering team’s cloud infrastructure budget was scaling faster than the revenue it supported. Monthly spend had risen from $62,000 to $122,000 in just two quarters, with no corresponding increase in reliability metrics or user satisfaction. Senior leadership commissioned an infrastructure optimization initiative with a hard constraint: maintain uptime above 99.95%, satisfy SOC 2 Type II auditors, and reduce net infrastructure costs by at least 50% within six months. ## Challenge The engineering team operated with a loosely governed microservices architecture. fourteen different services handled payments, notifications, tax calculations, user profiles, file processing, webhooks, reporting, audit logging, and admin workflows. Each service had been deployed independently, often without strict resource limits, resulting in overprovisioned clusters, idle replicas running 24/7, and redundant data processing jobs. Observability was fragmented across three monitoring tools with limited retention, making it difficult to correlate cost spikes with actual usage patterns. Compounding the problem, the DevOps team had limited staffing to manage this complexity, and the finance department lacked real-time visibility into which teams drove the highest infrastructure spend. ## Goals The initiative established four primary goals. First, reduce total monthly cloud infrastructure costs by a minimum of 50%. Second, maintain or improve system reliability metrics, specifically keeping error rates below 0.02% and p99 latency under 300 milliseconds. Third, preserve SOC 2 Type II compliance posture without requiring significant policy overhauls. Finally, improve cost visibility so engineering managers could predict monthly spend at the team level with ±10% accuracy. ## Approach The optimization effort began with a two-week assessment phase. The team used profiling data and cost reports to categorize spend into three buckets: compute, managed services, and data transfer. Compute represented 58% of the budget, managed services 31%, and data transfer 11%. Within compute, the team identified that services with the highest variability in traffic—such as webhook processing and notification delivery—were consistently overprovisioned for baseline load while failing to scale during peak demand. The assessment also uncovered orphaned resources, including six unused SSL certificates, three idle NAT gateways, and two legacy databases that had not been accessed in five months. The team adopted a phased approach spanning planning, consolidation, cost-control automation, and observability enhancements. Early phases focused on shutting down idle resources and consolidating overlapping workloads. The middle phase targeted right-sizing compute instances and introducing serverless components for bursty traffic patterns. The final phase built automated budgets and alerts to prevent future cost regressions. ## Implementation Implementation began with centralizing all infrastructure as code using Terraform modules. Every resource was tagged with standardized labels including team, environment, cost center, and service name. This tagging initiative alone improved cost attribution accuracy from 32% to 94%. The DevOps team also introduced Kubernetes Vertical Pod Autoscaling in place of hardcoded CPU and memory requests, allowing pods to adjust resource allocation dynamically as traffic patterns shifted. Next, the team migrated the notification service, webhook handler, and report generator from always-on container instances to AWS Lambda functions with provisioned concurrency set at 40% of peak load. This change alone reduced compute costs for these three services by 71%. For the core payment processing and tax calculation workloads, the team consolidated six overlapping message queues into a single partitioned Kafka cluster with optimized retention policies, cutting managed service costs by 38%. The data tier underwent a major change as well. The team replaced two large relational databases handling user profiles and audit logs with a document database for read-heavy profile queries and a columnar store for analytics. This change reduced database licensing and storage costs by 54% while improving query performance by 42%. Connection pools and query caching were introduced throughout the application layer to reduce redundant reads. Recognizing that cost optimization requires enforcement as well as engineering, the team implemented automated budgets using custom scripts triggered every six hours. Any service whose weekly spend exceeded its rolling average by more than 20% triggered an alert to the owning engineering manager. Weekly cost reports were sent automatically to all team leads, breaking spend down by service, environment, and cost center. The reports were designed to be actionable rather than informational, highlighting the three services whose costs had changed most dramatically and offering recommended next steps. ## Results The initiative delivered on every primary goal within five months, slightly ahead of schedule. Monthly cloud infrastructure spending dropped from $122,000 to $47,000, a 61% reduction. The platform continued to scale smoothly, reaching 2.1 million active users by the end of the six-month period without any service-wide outages. System reliability improved as well: error rates fell from 0.035% to 0.014%, and p99 latency decreased from 420 milliseconds to 210 milliseconds. SOC 2 Type II auditors verified that the infrastructure changes maintained compliance, and the team spent zero hours on remediation issues during the annual audit. Finance and engineering leadership gained visibility that had previously been absent. Cost attribution accuracy reached 94%, and engineering managers could now forecast monthly spend at the team level with ±8% accuracy, well within the original target. The engineering team adopted a cost-aware culture, with new services required to submit a lightweight cost projection before deployment approval. ## Metrics The following metrics summarize the improvements across the initiative: • Monthly cloud spend reduced from $122,000 to $47,000, a net annualized saving of $900,000. • Compute costs decreased by 65% through right-sizing and serverless migration. • Managed service costs decreased by 38% through Kafka consolidation and lifecycle management. • Data transfer costs decreased by 48% through traffic routing optimization and CDN compression. • Error rate improved from 0.035% to 0.014%, a 60% reduction. • p99 latency improved from 420 milliseconds to 210 milliseconds, a 50% improvement. • Cost attribution accuracy improved from 32% to 94%. • Forecast accuracy reached ±8%, surpassing the ±10% target. ## Lessons Learned The most important lesson learned was that cost optimization is not a one-time project but a continuous operational practice. The team discovered that without automated enforcement, costs would gradually regress as features accumulated. The automated budget alerts and weekly reports became more valuable over time, not less, because they established a shared language between engineering and finance that had not previously existed. The second lesson was that observability drives optimization. The team spent the first two weeks of the project gathering data and profiling spend before making any changes. This discipline prevented costly mistakes. For example, profiling revealed that one seemingly expensive service was actually processing fraudulent transactions efficiently, and an uncritical cost cut would have weakened the company’s fraud defenses. The cost savings were highest where observability quality was highest, confirming the team’s hypothesis. The third lesson was that cultural change matters more than tool changes. The most significant and durable improvements occurred after the team introduced lightweight cost forecasting into the deployment workflow. Engineers began asking cost questions earlier in the design process rather than reacting to overages after the fact. This cultural shift—enabled by process change rather than technology—produced benefits that persisted long after the consultants who had guided the initial optimization had departed.