Meta is restructuring its generative artificial intelligence strategy following the failure of the Llama 4 Behemoth model and the loss of leadership in the open-weight models segment, which has shifted to China's DeepSeek. The reasons for the failure included erroneous choices in block attention, issues with expert selection routing, data quality, and scaling experiments. In a complete return to founder management, Mark Zuckerberg personally leads the restructuring, identifying two key issues: talent and computational resources.
The company acquired 49 percent of Scale AI at a valuation of about $30 billion – a move that directly addresses data issues. Zuckerberg is building the Superintelligence team from scratch and personally poaching leading researchers. The average offer ranges from $200 to $300 million over four years, which is tens of times higher than competitors' terms. Key hires include Nat Friedman (former CEO of GitHub), Alex Wang (former CEO of Scale AI), and Daniel Gross (co-founder of SSI).
The infrastructure is also being rethought. Instead of a traditional H-shaped project, Meta is building data centers in tent structures – prioritizing speed of launch over redundancy. In Ohio, a 1-gigawatt Prometheus cluster is being constructed, and in Louisiana, a 2-gigawatt Hyperion cluster, which will become the largest in the world by the end of 2027: 1.5 gigawatts of computing power in the first phase. Due to a lack of local energy, Meta is building two 200-megawatt gas power plants on-site.
Meta and Google have followed a gradual AI scaling strategy for years – integrating it into existing products to enhance advertising. This yielded financial results but did not help catch up with OpenAI and Anthropic in foundational models. Now, Meta is shifting from a lack of computational power to an abundance for each researcher, intending to compete with OpenAI in training scale.
Related articles
