Report: Nvidia could acquire LLM startup AI21 Labs for $3B – News

Nvidia Corp. is reportedly in advanced talks to acquire AI21 Labs Ltd., a startup that develops large language models and agent development tools.

Calcalist reported today that a deal could be worth between $2 billion and $3 billion. At the top end of the range, that’s more than double the valuation AI21 received after its last publicly disclosed funding round in 2023. The investment included the participation of Nvidia, Google LLC, Samsung Electronics Co. and other prominent backers.

According to Calcalist, AI21 quietly raised another $300 million from investors earlier this year. It’s believed the round closed at a similar valuation as the 2023 raise. The report also revealed that AI21 is generating about $50 million in annualized revenue.

Tel Aviv-based AI21 offers a series of open-source LLMs called Jamba. According to the company, its models can process long prompts up to 2.5 times faster than competing algorithms. The Jamba series owes its speed to an architecture that combines elements of a transformer model with a neural network design called Mamba.

Mamba replaces transformers’ attention mechanism, which they use to filter the data they process based on importance, with a so-called state space model. That’s a mathematical structure originally invented to study physical objects. It enables a Mamba model to process long prompts using significantly less memory than a transformer, which boosts inference speeds.

AI21 generates revenue with a paid software platform called Maestro. It includes tools that help developers build and manage artificial intelligence agents.

Before companies make a dataset available to an AI agent for processing, they organize it in a form that lends itself better to analysis. Maestro promises to ease the process. It can harmonize documents stored in different file formats, remove unnecessary information that might confuse the AI agent and apply other optimizations.

After an AI agent generates a prompt response, Maestro validates the accuracy of the output. Developers can have the platform display the processing results in a specific format such as a report. It’s also possible to customize other details such as the amount of infrastructure the platform uses to generate a prompt response.

Nvidia might be planning to integrate Maestro into Nvidia AI Enterprise, the AI software suite it ships with its graphics cards. The toolkit includes pre-packaged LLMs, development tools that enable researchers to build custom models and other resources.

The acquisition comes less than a week after Nvidia inked a $20 billion deal to license the technology of Groq Inc., a venture-backed chip developer. The startup sells processors optimized to run inference workloads. The deal will see Groq’s founding Chief Executive Officer and other key employees join Nvidia.

Photo: Nvidia