Applied Compute raises $350 million at a $3.25 billion valuation after passing $100 million in run-rate revenue

Applied Compute, an AI infrastructure startup founded by Rhythm Garg, Yash Patil and Linden Li, announced on November 2 that it has raised $350 million at a valuation of $3.25 billion. The round was led by investor Elad Gil. According to the company's announcement, Kleiner Perkins, Lux Capital, Nvidia, Blackstone, Shopify Ventures and HubSpot Ventures took part, along with more than a dozen other firms.

Revenue and customers

The company says its annualized revenue run rate has passed $100 million, a level it reached less than 18 months after its founding. It lists Microsoft, Shopify, Mercor, Harvey, DoorDash, Cognition, Clay and Decagon among its customers and says the number of enterprise clients has more than doubled since its previous raise. The team has tripled in size since January, with researchers and engineers making up 85% of staff.

Applied Compute also cites two customer results. It says a custom model it built for Harvey's Review Table product improved answer and citation quality at less than half the inference cost of a frontier model, and that a custom router cut Clay's production inference costs by up to 47%. These figures come from the company and have not been independently verified.

What the platform does

The startup's pitch is that businesses should own the models behind their products rather than rent general-purpose ones. It trains, deploys, evaluates and continuously improves custom models built on open-weight foundations. Its AC2 platform, introduced about three months ago, combines training, inference and ongoing improvement; the company says customers have run more than 1,700 training runs on it since July. Its inference service handles more than one trillion tokens a day, and about 40% of customers use it for both training and serving.

SiliconANGLE notes that the platform relies on techniques such as asynchronous and online reinforcement learning, as well as self-distillation, which avoids the need for a separate teacher model. A router called Shift-1 sends each request to the most cost-efficient model and can split complex tasks across several models.

Context

The company's growth rests on the bet that enterprises will prefer tailored open-weight models to off-the-shelf frontier models for at least part of their workloads, mainly on cost and control grounds. Applied Compute says the money will go toward more compute capacity, hiring and faster enterprise deployments. The founders describe training, inference and routing as only the first layers of a broader stack, without detailing what comes next.