Applied Compute finished a $350M round less than 18 months after starting. The valuation now sits at $3.25B, and the AI infrastructure provider says annualized recurring revenue has crossed $100M, with Microsoft and DoorDash among its customers.
Elad Gil led the raise, with Nvidia, Shopify Ventures, HubSpot Ventures and over a dozen other investors joining. The company was not hunting for capital out of desperation. The revenue milestone came first; the check followed.
Its platform serves teams training large language models and running inference. The technical difference sits in how training is orchestrated. Conventional stacks run data generation and weight adjustment in sequence, with each stage waiting on the other. Applied Compute uses asynchronous reinforcement learning so both proceed simultaneously, plus a self-distillation method that drops the need for separate teacher models.
An assistant inside the product automates research busywork, such as summarising which skills a model picked up during a training run, a task that otherwise demands reading long reasoning logs. For inference, an autoscaling engine sizes hardware to live traffic and an online reinforcement step keeps weights tuned to what users actually ask.
A router called Shift-1 sends each coding agent request to the cheapest model able to handle it, and can split complex jobs across several models.
The cash funds infrastructure purchases and market expansion. Applied Compute frames training, inference and routing as components of a larger stack, one intended to let companies own their own models and the systems that run and secure them.
