AWS to bare metal two years later: Answering your questions about leaving AWS

Original Article ↗ Hacker News Discussion ↗

When Bare Metal Makes Sense vs. Cloud

  • Many commenters agree the article’s key condition is decisive: a 24/7, steady baseload with high reservation coverage favors bare metal; bursty or unpredictable workloads still favor cloud elasticity.
  • Several note that most real-world systems are less “spiky” than people assume, so they unintentionally pay cloud premiums for workloads that would run fine on a few well-sized servers.
  • Hybrid patterns get praise: keep database / steady compute on colo or rented metal, use cloud for bursty components, CDNs, or hard-to-replicate services (e.g. CloudFront-like, SES, managed email).

Costs: Compute, Bandwidth, and Managed Services

  • Multiple concrete comparisons: Hetzner / OVH bare metal is often ~5–10× cheaper than equivalent AWS compute, with free or much cheaper egress.
  • Bandwidth is repeatedly called the “real killer” on AWS; NAT gateways and cross‑AZ traffic are singled out as nasty surprises.
  • Managed DBs, Kafka, and serverless offerings are described as excellent but “extremely expensive” at scale; some teams migrate off them to self-managed equivalents for cost reasons.
  • Others counter that S3 and some core AWS services can be cost‑competitive or cheaper than home‑grown equivalents at large scale, especially when you truly need their durability and geo‑replication.

Operational Complexity, Skills, and Org Dynamics

  • Strong disagreement over whether cloud reduces ops burden: many report bigger AWS ops teams and more DevOps toil (Terraform, IAM, CI/CD, FinOps) than when running on-prem.
  • Others argue bare metal reliably becomes a time sink: endless “little tasks” around hardware, backups, security, and upgrades that sap startup velocity, especially without strong infra talent.
  • Several say modern tooling (Kubernetes, Talos, Proxmox, Ansible, Kamal, etc.) plus LLMs has lowered the barrier to running your own infra; critics respond that k8s itself is fragile and overkill for many.

Reliability, Hardware, and Risk

  • Debates around ECC RAM, dual PSUs, and cheap hosts: some insist non‑ECC / single‑PSU is “a disaster waiting to happen,” others report decades on Hetzner/OVH with only a couple failures.
  • Consensus that hardware failure risk matters more at large fleet scale; for small setups, simple redundancy (two DCs, backups, occasional failover tests) is usually acceptable.
  • Some worry OneUptime’s earlier single‑rack phase was lucky; others note they maintained an AWS fallback and a second colo site, so total risk may be lower than typical cloud‑only shops during regional outages.

Lock‑In, Culture, and Cloud Economics

  • Recurrent theme: AWS’s real moat is organizational, not technical—certification culture, “nobody gets fired for buying AWS,” resume‑driven architectures, and fear of owning hardware.
  • Analysts’ “bear case” mentioned: value and margin may drift to higher‑level SaaS while hyperscalers become low‑margin server lessors.
  • Several predict rising demand (and consulting work) for de‑clouding and modern bare‑metal/hybrid setups as bills grow and the hype cycle cools.