Nvidia in Talks to Acquire AI21 Labs for Up to $3 Billion

Nvidia Corp. is in advanced negotiations to purchase AI21 Labs Ltd., an Israeli developer of large language models and tools for building AI agents, according to reporting from Calcalist. The transaction could command a price tag between $2 billion and $3 billion.

At the upper end of that range, the deal would value AI21 at more than twice what it was worth following its most recent public funding announcement in 2023, when Nvidia, Google LLC, Samsung Electronics Co. and other major investors participated. Calcalist also disclosed that AI21 secured an additional $300 million from backers earlier this year at a valuation believed to match the 2023 round. The company is currently producing approximately $50 million in yearly revenue.

Jamba and the Mamba Architecture

AI21's product lineup centers on Jamba, a collection of open-source large language models. The company claims its models can handle extended prompts at speeds exceeding competing systems by 2.5 times. This performance advantage stems from Jamba's hybrid design, which merges transformer model components with Mamba, a different neural network architecture.

Mamba operates by substituting the attention mechanism found in transformers—which filters incoming data by relevance—with a state space model, a mathematical framework originally developed for analyzing physical systems. This substitution allows Mamba-based models to work through lengthy prompts while consuming substantially less memory than transformers, resulting in faster inference.

Maestro Platform

AI21 generates income through Maestro, a commercial software platform designed to help developers construct and oversee AI agents. The platform streamlines the preparation of datasets before AI agents process them by normalizing documents across multiple file formats, stripping out extraneous data that could mislead the agent, and executing other refinements.

Once an AI agent produces a response, Maestro can verify its correctness. The system also permits developers to shape how results are displayed—such as in report form—and to fine-tune other parameters, including the computational resources allocated to generate a response.

Industry observers suggest Nvidia could fold Maestro into Nvidia AI Enterprise, its software package bundled with graphics processors. This suite encompasses pre-built LLMs, tools for researchers to develop proprietary models, and supplementary resources.

Part of a Broader Strategy

The potential acquisition follows Nvidia's announcement just days earlier of a $20 billion licensing agreement with Groq Inc., a venture-funded chip manufacturer specializing in processors built for inference operations. Under that deal, Groq's founding chief executive and other senior personnel will transition to Nvidia.