Boundless set out to build a zero-knowledge proving network. We ended up building something bigger: a global network that coordinates GPU capacity at scale.
Today, we’re expanding into AI.
@sshankar on what’s next:
Most public inference benchmarks are wrong for production. I analyzed more than 55 million public request records across chat, code, RAG, reasoning, and agents, plus modeled batch-job examples.
All skewed heavily toward prefill before cache reuse: long inputs followed by short
Boundless CEO @sshankar spent his early career at Lyft and Grab, where the problem was idle cars and drivers with no way to convert time into income.
GPUs have the same problem: the workloads come in spikes, and even well-run infrastructure teams top out at 70–80% utilization.