Meta Pursues Dual AI Strategy
Meta is advancing a dual strategy in artificial intelligence, engaging in acquisition talks with Perplexity AI before securing a multibillion-dollar deal with Scale AI for data labeling. The company is also continuing its open-source push, announcing new models like Llama 4 and "Scout". This approach blends proprietary scale with a commitment to the open-source ecosystem.
- The $14.3 billion deal with Scale AI secured Meta a 49% stake in the data-labeling startup and brought its 28-year-old CEO, Alexandr Wang, to Meta to lead a new "Superintelligence" lab. This move addresses a critical bottleneck in AI development: access to massive, high-quality datasets for training foundation models. - Meta's Llama 4 Scout model utilizes a Mixture-of-Experts (MoE) architecture, with 17 billion active parameters out of 109 billion total. It features a 10-million-token context window and is natively multimodal, processing both text and images, yet is efficient enough to run on a single NVIDIA H100 GPU with quantization. - The pursuit of Perplexity AI, which ultimately did not result in an acquisition, was part of an aggressive talent and technology push by CEO Mark Zuckerberg to catch up with rivals like OpenAI and Google. This talent war has seen Meta offer signing bonuses as high as $100 million to recruit top AI researchers from competitors. - Meta's open-source strategy is championed by Chief AI Scientist Yann LeCun, who argues that an open approach accelerates innovation and safety through broader community involvement, contrasting with the more "walled garden" strategies of competitors. However, LeCun has also publicly stated he believes Large Language Models are a "dead end" for reaching human-level AI, indicating a divergence in technical roadmaps even within Meta. - By open-sourcing its Llama models, Meta effectively shifts the significant computational and infrastructure costs of deployment to the broader developer community. This strategy aims to establish Llama as a foundational industry standard, fostering a wide ecosystem and preventing a single competitor from achieving a proprietary monopoly. - In August 2025, Meta restructured its AI division under the new Meta Superintelligence Labs (MSL), placing Alexandr Wang in direct control of all senior AI leaders and organizing efforts into four teams: research, training, products, and infrastructure. - To support its AI ambitions, Meta is planning a massive infrastructure expansion, with capital expenditures projected to reach up to $65 billion in 2025. A new "Meta Compute" initiative, co-led by Head of Infrastructure Santosh Janardhan, will centralize the build-out of data centers capable of supporting these intensive AI workloads. - The Scale AI deal has disrupted the AI data labeling market, causing competitors like OpenAI and Google to reportedly wind down their work with the company to avoid sharing insights into their data priorities with a key rival.
