Applied Compute has raised $350 million in a funding round that values the artificial intelligence infrastructure startup at $3.25 billion. The company announced the financing on Monday, saying its annualized recurring revenue has surpassed $100 million.
Angel investor Elad Gil led the round. Nvidia, Shopify Ventures, HubSpot Ventures and more than a dozen other investors also participated. Applied Compute said it will use the new capital to expand its market presence and acquire additional infrastructure.
Launched less than 18 months ago, Applied Compute provides a cloud platform for training large language models (LLMs) and running inference, the process of using a trained model to generate responses. The company also has machine learning researchers who assist customers with projects, including developing quality evaluations for new models. Microsoft and DoorDash are among its customers.
A platform for model training and inference
The company’s training system is built around asynchronous reinforcement learning. In a typical LLM training stack, one component generates data by running copies of a model on different tasks, while another uses that data to adjust the model’s weights. When the components operate synchronously, each must wait for the other to finish before proceeding.
Applied Compute says its approach allows data generation and weight adjustment to run at the same time. The goal is to reduce waiting between those stages and accelerate model development. The platform also supports self-distillation, an optimization method that the company says can remove the need for teacher models—specialized neural networks used to supervise training.
A built-in AI assistant handles some repetitive research tasks. One example is summarizing the capabilities an LLM acquired during a training run. Researchers would otherwise need to examine lengthy reasoning logs to identify what the model learned.
Applied Compute also offers infrastructure for inference workloads. Its autoscaling engine adjusts the hardware assigned to a model in response to user traffic. For customers seeking to refine models after deployment, an online reinforcement learning feature periodically uses prompts received by an LLM to fine-tune its weights, with the stated aim of keeping the model aligned with customer requirements.
Routing requests across models
The training and inference services are offered alongside Shift-1, Applied Compute’s model router. The product can direct requests from coding agents to an LLM selected to handle each task cost-efficiently. It can also divide more complex work among multiple models, rather than sending the entire task to a single LLM.
The combination places training, deployed model workloads and request routing within the company’s product portfolio. Applied Compute’s founders have described these capabilities as early components of a broader AI infrastructure stack. The company has not detailed in the announcement what additional products that longer-term effort may include.
Passing the $100 million annualized recurring revenue mark gives the startup a reported measure of its current revenue run rate, while the financing provides funds for expansion and infrastructure purchases. The announcement did not disclose a breakdown of the round’s terms beyond its total size and valuation, or specify how the new infrastructure will be deployed.
For cloud providers and AI developers, the company’s product focus reflects the operational demands of building and serving LLMs: training systems must coordinate data generation and model updates, inference capacity must respond to changing demand, and requests may need to be routed among different models. Applied Compute is seeking to serve those needs through a combined platform, while using the new funding to grow its business and infrastructure base.
