benchmarks.
Independent weekly runs, published openly.
Every week we rent the same workloads on the mesh and on a hyperscaler, then publish the numbers. Nothing is cherry-picked — the raw logs are in the repo.
training throughput92 / 100
inference latency71 / 100
render · 4k83 / 100
cost per job66 / 100
read the fine print
Scores are relative to a hyperscaler baseline of 100. Cost is inverted — lower is better, so a 66 means the mesh does the same job for roughly two-thirds the price.