Applied Compute

Applied Compute is an artificial intelligence company founded in 2025 by three former OpenAI technical staff — Yash Patil (CEO), Rhythm Garg (CTO) and Linden Li (Chief Architect) — all Stanford alumni. The company builds what it calls the "Agent Cloud" or "AI Supercloud": a platform for training, serving and continuously improving open-weight models that are specialized to an individual customer's data, workflows and evaluation criteria. Its stated thesis, which it labels "Specific Intelligence," is that general frontier models lack a given company's context, standards and judgment, and that enterprises should own models trained on their own operational reality.

The platform spans post-training of tool-using, long-horizon agents (including reinforcement learning inside customer-supplied harnesses and graders), production inference with dedicated single-tenant deployments, and a continuous-improvement loop that mines production traces for training data and reward signals, supported by online RL and self-distillation. Supporting components include a developer SDK covering training, inference, sandboxes, evals, memory and agent serving; rollout-level observability with automated detection of reward hacking, regressions and edge cases; elastic scheduling and fault tolerance; and support for open-weight models from 1B to over 1T parameters. Deployments can run serverless on Applied Compute's cloud or entirely inside a customer's VPC from a single control plane.

Alongside the software, Applied Compute positions an embedded research team of frontier-lab veterans that works with customer engineers on eval design, dataset/environment/harness construction and deployment, or in a managed model where the customer runs experiments and Applied Compute operates the training and serving infrastructure. According to a job posting cited by SiliconANGLE, the company trains models on a cluster of several thousand GPUs.