Former OpenAI researchers launch Applied Compute with $80M in funding – News

Three former OpenAI researchers have launched a new startup, Applied Compute Inc., that plans to develop custom artificial intelligence models for enterprises.

Applied Compute announced on Wednesday that it has raised $80 million in funding. The capital was provided by Benchmark, Sequoia, Lux and a group of angel investors.

The company didn’t disclose its valuation. Last month, The Information reported that it was in the process of raising capital at a $500 million valuation. In July, sources told Upstarts Media that the company had closed a $20 million round.

Applied Compute was launched a few months before its initial raise by Yash Patil, Rhythm Garg and Linden Li. Patil, the company’s chief executive, previously worked on OpenAI’s Codex programming assistant. Garg and Li helped build the ChatGPT’s o1 reasoning model and AI training infrastructure, respectively.

The startup will train custom AI models optimized for each customer’s use cases. According to the company, every custom model will be trained on the data of the organization that commissioned it. It says that such a training workflow can provide higher output quality than what users receive from general-purpose language models such as GPT-5.

A job posting for an infrastructure engineer indicates that Applied Compute plans to train AI models using reinforcement learning, or RL. That’s one of the most popular approaches to training reasoning models. In an RL project, researchers give a neural network a set of sample tasks and award it points when it completes a task correctly. That feedback enables the model to refine its responses.

Applied Compute already has multiple customers. It’s working with DoorDash Inc., programming assistant developer Cognition AI Inc. and AI training data provider Mercor Inc., which received a $10 billion valuation on Monday. Applied Compute staffers wrote in a blog post that the company is helping customers build custom AI models and agents in days instead of the months usually required for the task.

One of the most common ways developers speed up AI training is by applying a method known as LoRA. It involves extending a pretrained model with a small number of additional parameters. When it’s time to customize the model, developers only train those additional parameters instead of the entire model.

Applied Compute’s infrastructure engineer job posting revealed that it’s training AI models on a cluster of several thousand graphics processing units. Large-sale GPU clusters are expensive to operate even when they’re rented from a cloud provider rather than build from scratch. As a result, it’s possible the startup will raise additional funding in the near future to finance its development efforts. 

Image: Unsplash