Nvidia Corp. is in advanced talks to acquire cloud startup Lepton AI Inc., The Information reported on Wednesday.

Sources familiar with the matter told the publication that the deal could be worth several hundred million dollars. That would represent a major exit for Lepton AI’s investors, CRV and Fusion Fund. The two venture capital firms led a $11 million seed round for the company about two years ago.

Lepton AI was founded in 2023 by Yangqing Jia and Junjie Bai, former researchers at Meta Platforms Inc.’s artificial intelligence lab. The duo helped develop several industry-standard AI tools, including PyTorch, the Facebook parent’s popular model development framework. Lepton AI reportedly has about 20 employees and counts multiple venture-backed startups among its customers.

The company’s namesake cloud platform is optimized for AI workloads. Customers can use it both to train AI models and perform inference, the task of running a neural network in production once it’s trained.

Lepton AI provides a visual interface for setting up training clusters in its cloud. The company offers a choice of several Nvidia graphics cards. For teams that train multiple models side-by-side, the platform provides a tool that can be used to manage centrally the amount of hardware assigned to each project.

During training, AI models sometimes encounter errors that can decrease their output quality. Lepton AI’s platform provides features for detecting such errors. It also spots more subtle technical issues, such as cases when a neural network might be using more memory than it should during training.

After development of a model is complete, Laptop AI customers can deploy it on inference-optimized instances in the company’s cloud. The platform promises to provide processing speeds of more than 600 tokens per second while keeping latency under 10 milliseconds. An auto-scaling feature automatically adds more hardware resources to customers’ inference environments when demand increases.

Lepton AI speeds up inference using open-source tools such as vLLM. The technology consolidates similar LLM prompts into a single request, which reduces the amount of processing a model must perform. Developers also use vLLM to reduce their models’ memory footprint.

It’s unclear whether Nvidia is seeking to acquire Lepton AI for its machine learning talent or plans to continue offering the startup’s cloud platform. The platform competes with the offerings of several major Nvidia customers. One of those customers, AI cloud operator CoreWeave Inc., plans to hold its initial public offering later today.

The report that Nvidia could buy Lepton AI comes days after word emerged about the company’s acquisition of Gretel Labs Inc., another venture-backed AI startup. That deal was reportedly valued above $320 million. Gretel provides tools for creating synthetic data, automatically generated data that can be used to train AI models.

Image: Nvidia