Blacksmith enforces its org-only rule by leaving the job queued forever

Blacksmith's quickstart mentions "organization" 33 times and "personal repo" twice, and the second of those two mentions is where it says the product is unavailable for personal repositories ^[9]. So the rule is documented; the enforcement is a job the scheduler will never place, sitting in `queued`, which is also what waiting for capacity looks like.

The next failure mode was a cancellation. On a later run both jobs started, and the Blacksmith job was killed after about five minutes with no steps executed, which the author read as an account-verification hiccup rather than anything about the runner image ^[14]. "None of these were show-stoppers, but if you're evaluating Blacksmith on a deadline, budget for at least one queued-forever mystery and one silently-cancelled run before you get a clean baseline," the author wrote ^[15].

The timing is better built than most runner comparisons. Each trial was measured with `date +%s%N` around the `node` invocation instead of bash's `time` builtin or the wall-clock job duration GitHub reports, and one untimed warmup run was discarded before the loop started ^[16][2]. The workflow file and the raw run logs are public, at `.github/workflows/blacksmith-benchmark.yml` and `.github/scripts/run-trials.sh` in github.com/Shed-engineering/gzip-threshold-json-api ^[19].

The result that survives hardest is the separation. On client.js, the slowest of Blacksmith's ten trials came in at 166ms and the fastest of the github-hosted trials at 368ms ^[17]. Take the comparison most favourable to the shared runner and Blacksmith is still 2.2x ahead ^[1].

The narrower spread is a weaker finding than it looks. Blacksmith's ranges were 51ms and 14ms wide against github-hosted's 100ms and 51ms ^[18]. Absolute width falls with runtime, so a runner three times faster starts ahead on that measure even when its variability is identical. If the 51ms width is the client.js pair, it is 31 percent of that job's slowest trial, against 27 percent for the 100ms width on trials no faster than 368ms ^[2].

For 3.17x to transfer, the workload has to be what this one is: a small CPU-bound process invoked over and over. Blacksmith runs Firecracker microVMs on bare metal with gaming CPUs, and `runs-on: blacksmith-...` is the entire migration ^[5]. Dependency installs spend their time on network and disk, where a faster core moves the total less. The author says a workflow dominated by dependency installs, Docker layer builds or a slow test suite could land at a very different ratio, better or worse ^[20].

The company raised a $45M Series B led by Peak XV Partners at a $550M valuation in March 2026, on top of a $3.5M seed from GV and Y Combinator ^[6]. It claims more than 6,000 companies running CI on it, has one independent dev.to write-up about cutting CI costs, and otherwise appears in generic runner comparison listicles next to Depot and WarpBuild, with no Show HN or Launch HN ^[7]. The free tier is 3,000 minutes a month with no credit card required at signup ^[11].