Mark Zuckerberg is going big on artificial intelligence. Reportedly, Meta Platforms has agreed to take a 49% stake in artificial intelligence startup Scale AI for $14.8 billion, The Information reported on Tuesday, citing two people familiar with the matter.
Scale AI is a San Francisco-based artificial intelligence infrastructure company founded in 2016 by Alexandr Wang and Lucy Guo. It specializes in providing high-quality labeled data and model evaluation tools that are essential for training AI systems. Scale’s services support a variety of sectors, including autonomous vehicles, national defense, finance, and large language models, helping organizations ensure their AI models are accurate, efficient, and safe.
Reportedly, the deal, which has not been finalized yet, appears to be beneficial for Scale AI’s investors including Accel, Index Ventures, Founders Fund and Greenoaks, as well as its current and former employees, the report said.
READ: Amid surging energy demands from AI and data centers, Big Tech turns to nuclear power (November 29, 2024)
The company gained prominence by partnering with major tech firms like OpenAI, Meta, and Microsoft, as well as U.S. government agencies, including the Department of Defense. It operates through platforms like Remotasks and Outlier to deliver human-in-the-loop data services. As of 2024, Scale AI raised $1 billion in a Series F funding round, reaching a valuation of nearly $14 billion.
This acquisition may be Meta’s way of playing catchup as both its recent AI offerings suffered setbacks as both the Llama 4 and the company’s Behemoth AI model supposedly did not meet expectations.
Meta delayed the release of its flagship “Behemoth” AI model due to concerns about its capabilities, the Wall Street Journal reported last month.
What is Llama 4:
Meta’s Llama 4, released in April 2025, marks a major leap in large language models. It uses a mixture-of-experts architecture that allows it to efficiently scale with fewer active parameters while maintaining powerful overall capacity. The model is multimodal, capable of understanding both text and images, and supports 12 languages, making it versatile across different applications. Llama 4 comes in several variants, such as Scout and Maverick, with parameters ranging from billions to hundreds of billions, enabling use cases from chatbots to complex reasoning tasks. Meta designed Llama 4 to compete with industry leaders like GPT-4o by excelling on benchmarks like LMArena. However, Meta faced some criticism for releasing a specially optimized version for testing, raising questions about benchmarking transparency. Overall, Llama 4 underscores Meta’s commitment to open AI development, combining scalability, efficiency, and multimodal capabilities to push forward the state of AI.
What is Behemoth AI:
Meta’s Behemoth AI is the company’s upcoming flagship large language model, poised to be one of the most powerful ever created. With an estimated 2 trillion parameters and 288 billion active parameters, Behemoth aims to handle highly complex tasks requiring advanced reasoning, understanding, and knowledge synthesis. Though still in training with no release date yet, Behemoth is expected to serve as the foundation for Meta’s other models, enhancing their capabilities in language and multimodal tasks. Its massive scale reflects Meta’s ambition to compete at the highest level in AI innovation. By investing in Behemoth, Meta hopes to advance applications in natural language processing, computer vision, and beyond, setting new standards for model performance and flexibility. This development signals Meta’s strategic push to lead the AI space through cutting-edge research and deployment of large, sophisticated AI systems.