Meta Invests Hundreds of Billions in AI Restructuring Over Past Year, Including $14 Billion Investment in Scale AI and Appointment of Wang as Chief AI Officer
On October 12, Meta released two models, Muse Wonder and Muse Wonder Full, claiming them to be the strongest models to date, achieving cutting-edge levels for personal and professional agent tasks. The company stated that the models are comparable to or surpass Claude Fable 5.1, Opus 5.5, and GPT-6 Astra on certain benchmarks, excelling in areas such as memory and encoding.
Muse Wonder is offered through the Meta Model API and Muse Code, priced at $4 per million input tokens and $20 per million output tokens, with a context window of up to 1 million tokens. Muse Wonder Full is only available to Muse Code subscribers. The models will first be launched to paid subscribers of Muse personal agents, while the free tier will continue to use the previous Muse Spark. New features include Active Memory (cross-conversation memory of relevant details) and Active Focus (continuously advancing, correcting, and confirming important actions in difficult tasks).
The models were trained in Meta's gigawatt-level Prometheus data center, based on pre-training and reinforcement learning. The internal code name is Watermelon. Meta stated that this is the first cutting-edge model release since Alexandr Wang restructured the AI work last August. The Muse app had previously ranked high in downloads on the US App Store and Android store.
Mark Zuckerberg and Alexandr Wang discussed the model on the Cheeky Pint podcast, hosted by Stripe co-founder John Collison. The podcast lasts about 2 hours and 14 minutes, covering Muse breakthrough use cases, model-as-product, methods of building cutting-edge models, over $100 billion capital expenditure bets, agent commerce, OpenRouter, forms beyond smartphones, financing, why AI is not a winner-takes-all, Meta Superintelligence Labs (MSL), and working for Zuckerberg.
Meta has invested hundreds of billions in AI restructuring over the past year, including a $14 billion investment in Scale AI and the introduction of Wang as Chief AI Officer to lead Superintelligence Labs. The company plans to exceed 1 gigawatt in the Prometheus cluster in Ohio and has a 5-gigawatt plan for Hyperion in Louisiana. Zuckerberg stated that the annual cost of operating a 1-gigawatt cluster is about $15 billion, with construction costs estimated at $50 to $60 billion.
After the release, Meta's stock price rose about 4% to $4.7 in intraday trading. The flow of funds is shifting from the advertising business to AI subscriptions and agent services, benefiting Meta, which can acquire paying users through Muse and API, while competitors relying on open-source or lagging models are under pressure. The models are closed-source, with the best versions sold directly to enterprises.
Source: Public Information
ABAB AI Insight
Previously, Meta focused on the open-source Llama series but lagged in cutting-edge capabilities. In 2025, it invested $14 billion in Scale AI and brought in founder Alexandr Wang to establish Meta Superintelligence Labs, with Wang serving as Chief AI Officer. In April 2026, Muse Spark was released as the first proprietary model, positioned as an "appetizer," followed by the rapid launch of Muse personal agents in September, which quickly topped app store download charts. Muse Wonder is the first model explicitly labeled as cutting-edge following this restructuring.
On the capital front, Meta has committed hundreds of billions for data centers (Prometheus and Hyperion) and talent recruitment, having planned gigawatt-level infrastructure before the model release. The pricing and subscription model (Muse Code and the upcoming Muse paid tier) aim to convert heavy capital expenditures into revenue. Wang and Zuckerberg discussed the decision to bet over $100 billion in capital expenditures before the model's existence and the talent reboot at MSL on the podcast.
Similar to OpenAI and Anthropic, Meta is monetizing through closed-source cutting-edge models combined with APIs/subscriptions. Meta is currently in a phase of catching up from open-source lagging to closed-source cutting-edge and productizing, leveraging its billions of user base to embed models into Muse agents and existing platforms.
This essentially represents a technological substitution. The capabilities of cutting-edge models shift personal and professional agent tasks from experimentation to scalable products, with Meta attempting to differentiate reliability in agents through features like Active Memory and Active Focus. The mechanism involves converting large-scale infrastructure and proprietary training into subscription and API revenue while shifting from advertising reliance to diversified AI monetization.
ABAB News · Cognitive Laws
Infrastructure precedes models; bets determine the speed of catching up.
Agents are products; models are the foundation.
User scale amplifies monetization; closed-source locks in cutting-edge price differentials.