Every lab wants an H100 and there aren't enough to go around. Meanwhile, a few hundred thousand consumer cards sit in drawers and basements, doing nothing. That gap is the whole idea.

A single H100 beats a 4090 on raw throughput — but twenty 4090s beat one H100 for the same money, and they're already plugged in and paid for. The mesh just needs to split the work so a rack of consumer cards can hold its own against a datacenter.

the trick is sharding

Training a frontier model still wants one big, fast, tightly-coupled cluster. But fine-tuning, embedding, rendering, and inference — most of what people actually run — shard beautifully. That's the wedge.

We're not claiming a 4090 is an H100. We're claiming the world already owns enough 4090s, and they're idle.

— Soren VossFounder, Mesh