Applied Compute Inc., a fast-growing provider of artificial intelligence infrastructure, has raised $350 million in funding at a $3.25 billion valuation.

The startup announced the deal on Monday. Prominent angel investor Elad Gil led the round with participation from Nvidia Corp., Shopify Ventures, HubSpot Ventures and more than a dozen others.

Applied Compute operates a cloud platform that developers use to train large language models and perform inference, the process of running the models to provide responses to queries. Additionally, it has a team of machine learning researchers that can lend customers a hand with their projects. The group takes on tasks such as creating quality evaluations for newly developed models.

The software stack that AI researchers use to train LLMs usually comprises two main components. One creates multiple copies of the model, assigns a different task to each one and turns the results into training data. The other component uses that training data to refine the LLM’s weights. The two components usually run synchronously, which means that each one has to wait for the other to complete its work before launching.

Applied Compute’s platform uses a more efficient training approach called asynchronous reinforcement learning. It enables a training stack’s data generation and weight adjustment components to run at the same time, which speeds up LLM development. Additionally, the platform supports a second optimization method called self-distillation. It removes the need for teacher models, specialized neural networks that researchers use to supervise LLM training runs.

A built-in AI assistant automates repetitive aspects of customers’ work. It can, for example, summarize what new skills an LLM learned during a training run. Uncovering that information usually requires researchers to review lengthy reasoning logs.

Applied Computing’s platform also lends itself to running inference workloads. An autoscaling engine automatically adjusts the amount of hardware allocated to a model based on user traffic. A so-called online RL feature uses the prompts that the LLM receives to fine-tune its weights periodically, which keeps it aligned with customer requirements.

The company offers its training and inference features alongside a model router called Shift-1. It enables coding agents to send each user request to the LLM that can fulfill it in the most cost-efficient manner. Furthermore, it can split complex tasks among multiple LLMs.

Applied Computing disclosed in its funding announcement that its annualized recurring revenue has topped $100 million. The company, which launched less than 18 months ago, counts Microsoft Corp., DoorDash Inc. and other major tech firms as customers. It will use the newly raised capital to grow its market presence and buy more infrastructure.

“Training, inference, and routing are the first pieces of a much larger stack,” Applied Compute founders Rhythm Garg, Yash Patil and Linden Li wrote in a blog post. “We’re building the AI infrastructure every company will need to fully own their intelligence, from the models themselves to the systems that run, monitor and secure them.”

Photo: Applied Compute