Why we ditched AWS for local bare metal at Flowcast

Last September, my credit card got hit with a $4,200 AWS bill for Flowcast. We were mostly routing API calls and caching LLM responses, yet we were paying a premium for virtual machines that spent half their time waiting on network I/O. I looked at our Rupiah revenue, calculated the margin, and realized we were paying a lazy tax for convenience we did not actually need. ## The cloud is just someone else's overpriced computer Most devs think you need AWS or GCP to build something reliable. They spin up EC2 instances, RDS databases, and NAT gateways without calculating the egress fees. In Indonesia, those dollar-denominated bills eat your margin fast when your customers pay you in IDR. We bought two physical servers in a Jakarta data center for the cost of three months of our AWS bill. They have 128GB of RAM and dedicated NVMe drives, not the shared, throttled virtual disks Amazon sells you. ## Virtualization is a tax on performance People love talking about horizontal scaling, but they forget that hypervisors introduce latency. For an AI gateway like Flowcast, every millisecond matters when streaming tokens from OpenAI or local models. When we moved to bare metal, our internal API latency dropped by 40%. We no longer share CPU cycles with noisy neighbors on the same cloud hardware. Our database queries hit physical disks directly, making our cache hits almost instantaneous. ## Local bandwidth is practically free Cloud providers make money by letting you in for free and charging you to leave. Egress fees are the silent killer of tech startups in Southeast Asia. At Flowcast, we process millions of API requests daily. On AWS, we were billed for every gigabyte that left their Singapore region to reach users in Jakarta. By co-locating our own hardware in the Cyber 1 building, we get unmetered local bandwidth. We pay a flat monthly fee for the rack space and the port, regardless of how many terabytes we push. ## Managing physical hardware is easier than you think The biggest lie cloud sales reps tell you is that bare metal requires a massive operations team. They want you to believe you need a certified Kubernetes expert just to keep the lights on. For Flowcast, we use a simple setup: Debian, Docker, and basic systemd services to manage restarts. If a drive fails, the data center staff swaps it out under our support agreement. During my time at Deadsec monitoring CVEs, I learned that keeping your stack simple is the best security policy. Less proprietary cloud configuration means fewer vectors for misconfiguration. ## Calculate your egress tax today Open your last cloud bill and highlight the charge for Data Transfer Out and Provisioned IOPS. If that number is more than 30% of your total spend, search for colocation providers in Jakarta today and ask for a quote on a 1U server slot.