Enterprise AI training and inference infrastructure
Applied Compute
Applied Compute helps enterprises train, deploy and improve AI models around their proprietary data and workflows. Its AC2, Applied Compute Agent Cloud, available in private beta, combines open-model training, dedicated inference and learning from production traces. AI teams can bring their own training harness, inspect model rollouts and use feedback from deployed agents in subsequent training. Its research agent Ari analyzes experiments and identifies failure modes. Customers include Cognition, DoorDash and Harvey.
Founded in 2025, the company is led by cofounders Yash Patil, CEO; Rhythm Garg, CTO; and Linden Li, Chief Architect. Its research includes Relevance-Masked Self-Distillation, a method that uses teacher–student probability differences and an AI judge to select which tokens receive training updates. This concentrates learning on desired behaviors rather than unrelated wording changes. Experiments on a synthetic task explored how models could learn unfamiliar behaviors while retaining existing capabilities.
By August 2026, Applied Compute had a reported $50 million annualized revenue run rate. Its business combines consulting services with charges for compute. In April 2026, it announced $80 million in financing led by Kleiner Perkins at a $1.3 billion post-money valuation, bringing total funding to $160 million.
3 talks
Newest first
4 speakers at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.
Start here
- Efficient Reinforcement Learning
Start here to understand how sampling variability, illustrated with a Qwen-30B experiment, complicates predictable training times.
Rhythm Garg · Linden LiAI Engineer Code 2025
- Learning on the job: the future of post-training
Learn how post-training changes when agents move from single-turn questions and synthetic environments to long-horizon enterprise tasks.
Raymond FengAI Engineer World's Fair 2026
- Bringing Continual Learning into Enterprises
Use Denton’s four-quadrant framework to compare learning from offline hints, production traces, and on-policy data.
Samuel DentonAI Engineer World's Fair 2026
Messages from the stage
Training efficiency depends on resource allocation
Rhythm Garg and Linden Li describe how stragglers leave GPUs underused during synchronous training. Their systems modeling connects inference and training resource allocation to throughput and latency.
Enterprise environments must support replay
Raymond Feng discusses orchestrator-and-sandbox environments for post-training and the difficulty of reproducing production conditions. The talk connects comparisons across rollouts with reinforcement learning in existing agent harnesses.
Shorter reasoning needs careful learning signals
Samuel Denton describes shortening agent reasoning while preserving SWE-bench test performance. Reward-shaping regressions, KL learning signals, and relevance-masked self-distillation expose technical choices behind that goal.
Affiliations reflect each recorded session, not necessarily current employment.
Company sources · checked 2026-08-27
- Applied Compute
- Blog — News, research and customer stories
- Applied Compute — Comcast NBCUniversal LIFT Labs
- The Advantage You Own
- Applied Compute Agent Cloud
- Applied Compute in talks to double valuation to $3B on open-source demand
- Bringing Capabilities in Distribution via Relevance-Masked Self-Distillation