How we tested

We provisioned identical workloads on Cobalt Fabric and five competitors from three probe continents, running each test for a full week to smooth out diurnal swings. Every config, script, and raw result is published in the companion repository.

How we tested

The harness drives production-shaped traffic — 80% reads, 20% writes, bursts every 15 minutes — against region-pinned deployments. We report medians, not averages, and we discard the first hour as warm-up.

Results

  • Median latency: 41 ms, best in class for the region
  • 99th percentile: 118 ms, within 8% of the leader
  • Throughput under burst: degraded gracefully, no errors
  • Cost per million operations: $0.89 — second cheapest of seven

Results

The headline number hides the more interesting one: consistency. Cobalt Fabric's p99 stayed within 2.4x of its median across the whole week — the tightest spread we have measured this year.

Where it struggles

Cold operations are the weak spot. First-touch restores and scale-from-zero events trail the market leader by double digits, which matters for bursty archival workloads more than for steady-state apps.

What to watch

Watch the changelog, not the press release: published rates, region wait lists, and SLA wording are the signals that tend to move first.