Nvidia May Acquire AI21 Labs, a Large Language Model Startup, for $3 Billion

According to recent reports, NVIDIA is in advanced discussions to acquire AI21 Labs Ltd., a startup specializing in large language models and tools for developing AI agents.

A deal could be valued between $2 billion and $3 billion, Calcalist reported today. At the higher end, this valuation would more than double AI21’s last publicly disclosed funding round in 2023, which included investments from NVIDIA, Google LLC, Samsung Electronics, and other prominent backers.

AI21 quietly raised $300 million from investors earlier this year, with the post-money valuation reportedly close to its 2023 level, according to Calcalist. The report also noted that AI21 generates approximately $50 million in annual revenue.

Based in Tel Aviv, AI21 offers a family of open-source large language models called Jamba. The company claims these models can process long prompts up to 2.5 times faster than competing algorithms. This performance advantage stems from an architecture combining elements of transformer models with Mamba, a neural network design based on state space modeling.

Mamba replaces the attention mechanism used in traditional transformers—responsible for filtering data by importance—with state space models, a mathematical framework originally developed for modeling physical systems. This allows Mamba-based models to handle longer sequences using less memory, significantly boosting inference speed.

AI21 monetizes its technology through Maestro, a subscription-based software platform equipped with tools that help developers build and manage AI agents.

Prior to feeding datasets into AI agents, Maestro reorganizes them into formats better suited for analysis. The platform streamlines preprocessing by harmonizing documents across various file types, removing redundant or confusing information, and applying additional optimizations.

After AI agents generate real-time responses, Maestro validates the accuracy of the outputs. Developers can configure the platform to present results in specific formats, such as structured reports, and customize backend infrastructure usage for response generation.

NVIDIA may plan to integrate Maestro into its Nvidia AI Enterprise software suite, which comes bundled with its GPUs. This toolkit includes pre-packaged large language models (LLMs), development frameworks, and resources enabling researchers to train custom models.

The potential acquisition comes less than a week after NVIDIA signed a $20 billion agreement to license technology from Groq Inc., a venture-backed chipmaker. Groq produces processors optimized for running AI inference workloads. Under the terms, Groq's founding CEO and several key team members are expected to join NVIDIA.