Loophole Labs

Open-source AI compute orchestration platform unifying AI workloads across clouds, Kubernetes and GPU providers.

Thunder Compute is a systems research lab building GPU virtualization software, founded in 2022 and headquartered in San Francisco, and a member of Y Combinator's Summer 2024 batch. Its core technology, GPU-over-TCP, abstracts GPUs as network resources rather than bare-metal allocations, running invisibly beneath workloads to improve scheduling and utilization across data centers and partner GPU fleets. The company positions itself as VMware for GPUs and operates its own on-demand GPU cloud (per-minute billing, one-click VSCode connection) as a production testbed alongside enterprise software partnerships.

Prime Intellect, founded in 2023 and headquartered in San Francisco, builds decentralized AI training infrastructure. The company offers a full-stack toolkit covering compute, reinforcement learning, and evaluation, enabling enterprises to train their own agentic systems without relying on frontier AI labs. Customers include Ramp and Zapier, with an annualized revenue run rate of approximately $100 million.

Sail Research, founded in 2025 in San Francisco, USA, provides inference infrastructure and sandbox environments (Sailboxes) for long-horizon AI agents, rebuilding the inference stack from the chip level up and claiming up to 10x reduction in per-token cost.

Clastix, based in Naples, Italy, is an infrastructure software company behind the open-source Kubernetes multi-tenancy project Capsule and the hosted control plane engine Kamaji, combined into its commercial fleet-scale platform kMetal. It helps AI labs, telecoms, cloud providers and the public sector run Kubernetes at scale, with customers including Mistral AI, NVIDIA, Fastweb and OVHcloud.

Infron AI is a San Francisco AI inference infrastructure company offering one OpenAI-compatible API across 400+ models from 100+ providers, with intelligent routing, automatic failover, caching, unified billing and a 99.9%+ SLA for production AI and agent workloads.