Why We Host on Bare Metal in 2026
Choosing bare metal wasn’t about being contrarian; it was about predictability.
When you build an API gateway like Crate, latency is the only metric that matters. On standard cloud providers, you are constantly fighting the hypervisor and “noisy neighbors” for CPU cycles. A 10ms spike in the cloud is an anomaly; on our dedicated hardware, it’s a failure.
The “Jitter” Problem
In a virtualized environment (like AWS EC2), your code doesn’t own the CPU. You share “noisy neighbors” with other tenants. This leads to latency jitter—where your response time might be 10ms for one request and 150ms for the next.
Dedicated vs. Virtualized Baseline (2026)
| Environment | Average Latency | p99 Latency (Tail) | Standard Deviation |
|---|---|---|---|
| Crate (Bare Metal) | 0.4ms | 0.8ms | ±0.1ms |
| AWS EC2 (m7g) | 1.2ms | 45ms | ±12ms |
| Cloudflare Workers | 2.5ms | 18ms | ±4ms |
By stripping away the hypervisor, Crate delivers “Deterministic Performance.” We don’t just care about the average; we care about the worst-case scenario.
The Go Advantage Building in Go allows us to run lean, compiled binaries directly on the metal. We don’t have the overhead of a heavy VM or a container orchestration layer slowing down the request lifecycle.
Cost Determinism The cloud is a variable cost nightmare. By using colocation, our hardware costs are fixed. This allows us to offer the “Financial Firewall” feature—protecting our users from egress fee shock because we aren’t paying them ourselves.