Podcast Episode
Wang, who now serves as Meta's Chief AI Officer, has assembled a formidable team by recruiting researchers from OpenAI, Anthropic, and Google DeepMind. The lab represents Meta's most ambitious effort yet to develop frontier AI models that can compete with the industry's leading systems.
Beyond these flagship projects, Wang has indicated the lab is working to improve the coding capabilities of Meta's language models. The team is also conducting early research on world models, a type of AI system that can understand its surroundings through visual data and spatial relationships. This technology could have significant applications in robotics, augmented reality, and autonomous systems.
Internal reports suggest that the Avocado model is targeting a 60 percent score on the SWE bench Verified benchmark, a critical metric for measuring autonomous software engineering capabilities. This would represent a substantial improvement over Meta's current language models.
To put this in perspective, a gigawatt represents one billion watts of electrical power, roughly equivalent to the output of a large nuclear power plant. Meta's plans suggest an infrastructure buildout on a scale unprecedented in the technology industry.
The Meta Compute initiative will be led by Santosh Janardhan, Meta's head of global infrastructure and co-head of engineering, alongside Daniel Gross. Dina Powell McCormick, Meta's president and vice chairman, will coordinate government partnerships for data centre investments and financing.
Meta has indicated it may spend 60 to 70 billion dollars in capital expenditure on AI compute infrastructure during 2026 alone. Speaking at an event in Davos, Bosworth compared these massive AI investments to the 19th century railroad expansion, predicting that whilst some companies would fail, consumers and society at large will be the true beneficiaries.
This pivot marks a significant change in priorities for a company that had previously bet heavily on the metaverse as the future of social interaction. The reallocation of resources underscores the urgency Meta feels in catching up to AI competitors.
However, not everything has gone smoothly. Meta's most advanced Llama 4 model, codenamed Behemoth, has faced development delays and received a mixed reception compared to frontier models from competitors. The challenges have sparked internal debates within Meta Superintelligence Labs about whether to continue the company's traditional open source approach or shift towards proprietary models.
The appointment of Wang has also generated controversy. Yann LeCun, one of the godfathers of AI who previously worked at Meta, publicly called Wang young and inexperienced, though he acknowledged that Wang learns fast and knows what he doesn't know. The comments reflect broader tensions about Meta's AI strategy and leadership.
The pressure to deliver has intensified scrutiny of Meta's AI efforts. With the delivery of the first models from Meta Superintelligence Labs, the company has demonstrated tangible progress. However, the real test will come when Mango and Avocado are released publicly in the first half of 2026.
The success or failure of these models could have far reaching implications not just for Meta's AI ambitions, but for the broader competitive dynamics of the artificial intelligence industry. With billions of dollars at stake and the company's strategic direction on the line, Meta is making one of the biggest bets in its history on the promise of artificial intelligence.
Meta's AI Lab Delivers First Models as Company Bets Billions on Infrastructure
January 21, 2026
Audio archived. Episodes older than 60 days are removed to save server storage. Story details remain below.
Meta Platforms has achieved a significant milestone in its efforts to compete in the artificial intelligence race, with the company's newly formed Meta Superintelligence Labs delivering its first key AI models internally this month. Chief Technology Officer Andrew Bosworth described the models as very good during an announcement at the World Economic Forum in Davos, Switzerland.
The Birth of Meta Superintelligence Labs
Meta Superintelligence Labs emerged from a major restructuring of the company's AI operations in 2025. The reorganisation came amid growing concerns that Meta's AI capabilities were falling behind competitors like OpenAI and Google. To lead this initiative, CEO Mark Zuckerberg recruited Alexandr Wang, the 29 year old former CEO of Scale AI, after Meta acquired a 49 percent stake in his startup for 14.3 billion dollars.Wang, who now serves as Meta's Chief AI Officer, has assembled a formidable team by recruiting researchers from OpenAI, Anthropic, and Google DeepMind. The lab represents Meta's most ambitious effort yet to develop frontier AI models that can compete with the industry's leading systems.
New AI Models in Development
The lab has been developing two major AI models with fruit themed codenames. Mango is designed for image and video generation, whilst Avocado focuses on text processing and language understanding. Both models are planned for release in the first half of 2026.Beyond these flagship projects, Wang has indicated the lab is working to improve the coding capabilities of Meta's language models. The team is also conducting early research on world models, a type of AI system that can understand its surroundings through visual data and spatial relationships. This technology could have significant applications in robotics, augmented reality, and autonomous systems.
Internal reports suggest that the Avocado model is targeting a 60 percent score on the SWE bench Verified benchmark, a critical metric for measuring autonomous software engineering capabilities. This would represent a substantial improvement over Meta's current language models.
Massive Infrastructure Expansion
The AI model developments coincide with Meta's announcement of a sweeping infrastructure initiative. On 12 January 2026, Zuckerberg unveiled Meta Compute, a new top level organisation dedicated to expanding the company's AI infrastructure to tens of gigawatts this decade, with plans for hundreds of gigawatts or more over time.To put this in perspective, a gigawatt represents one billion watts of electrical power, roughly equivalent to the output of a large nuclear power plant. Meta's plans suggest an infrastructure buildout on a scale unprecedented in the technology industry.
The Meta Compute initiative will be led by Santosh Janardhan, Meta's head of global infrastructure and co-head of engineering, alongside Daniel Gross. Dina Powell McCormick, Meta's president and vice chairman, will coordinate government partnerships for data centre investments and financing.
Meta has indicated it may spend 60 to 70 billion dollars in capital expenditure on AI compute infrastructure during 2026 alone. Speaking at an event in Davos, Bosworth compared these massive AI investments to the 19th century railroad expansion, predicting that whilst some companies would fail, consumers and society at large will be the true beneficiaries.
Strategic Shifts and Internal Tensions
The AI push has come at a cost to other parts of Meta's business. Earlier in January, the company laid off more than 1,000 employees from its Reality Labs division, which focuses on virtual reality and metaverse products. The job cuts represent a strategic shift as Meta redirects resources from VR towards AI wearables and smartphone features.This pivot marks a significant change in priorities for a company that had previously bet heavily on the metaverse as the future of social interaction. The reallocation of resources underscores the urgency Meta feels in catching up to AI competitors.
However, not everything has gone smoothly. Meta's most advanced Llama 4 model, codenamed Behemoth, has faced development delays and received a mixed reception compared to frontier models from competitors. The challenges have sparked internal debates within Meta Superintelligence Labs about whether to continue the company's traditional open source approach or shift towards proprietary models.
The appointment of Wang has also generated controversy. Yann LeCun, one of the godfathers of AI who previously worked at Meta, publicly called Wang young and inexperienced, though he acknowledged that Wang learns fast and knows what he doesn't know. The comments reflect broader tensions about Meta's AI strategy and leadership.
The Competitive Landscape
Meta's aggressive moves come as the AI industry enters an increasingly competitive phase. OpenAI, Google, and Anthropic have all made significant advances in recent months, raising the stakes for companies trying to develop cutting edge AI systems.The pressure to deliver has intensified scrutiny of Meta's AI efforts. With the delivery of the first models from Meta Superintelligence Labs, the company has demonstrated tangible progress. However, the real test will come when Mango and Avocado are released publicly in the first half of 2026.
The success or failure of these models could have far reaching implications not just for Meta's AI ambitions, but for the broader competitive dynamics of the artificial intelligence industry. With billions of dollars at stake and the company's strategic direction on the line, Meta is making one of the biggest bets in its history on the promise of artificial intelligence.
Published January 21, 2026 at 1:11pm