h100
(5)
TensorPool
TensorPool (YC W25) is a San Francisco startup that turns cloud GPUs into a git-style command line. Instead of wrestling with SSH sessions, provider dashboards, and idle-instance billing, machine learning engineers describe a training job in a config file and run 'tp job push' - TensorPool spins up a multi-node cluster across partner clouds, streams logs, saves outputs, and shuts everything down, billing only for runtime. It advertises GPU access at roughly half the cost of major cloud providers, with H100s starting around $1.99/hr. Founded by three Stanford dormmates, the company was accepted into Y Combinator's Winter 2025 batch.
Modal
Modal is a serverless cloud platform that lets developers run compute-intensive AI and data workloads - inference, training, fine-tuning, batch jobs, and sandboxed code - without managing infrastructure. Built from scratch in Rust with a custom container runtime, file system, and GPU memory snapshotting, it delivers sub-second cold starts and elastic autoscaling from zero to thousands of GPUs, billed by the second. Founded in 2021 by Erik Bernhardsson and Akshat Bubna, the New York-based company reached roughly $300M in annualized revenue and raised a $355M Series C in 2026 at a $4.65B valuation.
The San Francisco Compute Company
The San Francisco Compute Company (SF Compute) runs a real-time marketplace for AI compute. It buys and operates large-scale, vetted GPU clusters - primarily Nvidia H100s wired with 3.2Tb/s InfiniBand - and sells them on flexible contracts, from a single hour to multiple years, that buyers can also resell. By turning long-term GPU capacity into a liquid spot market with transparent pricing, SF Compute lets startups, researchers, and enterprises buy exactly the compute they need without the multi-year commitments that dominate the industry, while pointing toward a future of cash-settled GPU futures.
Hyperbolic
Hyperbolic is an open-access AI cloud built for developers and researchers. It aggregates underutilized GPUs from data centers, mining farms and idle machines into an on-demand marketplace, and pairs that with serverless inference for state-of-the-art open-source models. The pitch is simple: fast, affordable access to compute and inference without sales calls or feature gates, at prices it says can run up to 75% below incumbents.
GMI Cloud
GMI Cloud is an AI-native GPU cloud built around NVIDIA's H100 and H200 accelerators. Founded in 2021 and headquartered in Mountain View, it sells on-demand and reserved GPU compute, an orchestration layer (Cluster Engine), and a low-latency Inference Engine to AI labs and enterprises building generative models. In 2025 NVIDIA named it one of seven Reference Platform Cloud Partners worldwide.