To The Moon

Not only the big model itself, Meta also wants to become a major AI infrastructure company.

On June 10th local time, it was reported by the media that $Meta Platforms, Inc.(META)$ will acquire a 49% stake in Scale AI for $14.9 billion (equivalent to approximately 106.6 billion yuan). The co-founder of Scale AI, Alexandr Wang, will become the head of Meta's newly established "Super Intelligence Group."

Based on the equity ratio, Wang and his team could potentially receive $7.4 billion from this deal, making it the most expensive "talent poaching" in Silicon Valley. For comparison, Google's acquisition of the DeepMind team in 2014 cost only $600 million.

In an internal letter, Zuckerberg wrote, "We will build the future of AI together."

Amidst the setbacks of the Llama 4 model and the continuous loss of AI team members, what is Meta's motive behind this major bet on Scale AI?

With Scale AI and Alexandr Wang on board, can Meta regain its footing in the upcoming AI battle?

01 The Most Expensive "Free Agent"

As the fastest-rising company in Silicon Valley in the AI era, Scale AI's valuation has been soaring like a rocket, expanding to $13.8 billion in just five years. However, Meta's acquisition of a 49% stake in the company comes at a cost of $14.9 billion.

The 49% stake is clearly considered to avoid antitrust review, but what Meta and Zuckerberg truly want is Alexandr Wang, one of the co-founders. This 19-year-old entrepreneurial genius will become the head of Meta's newly established Super Intelligence Lab, leading Meta AI into a new era.

Interestingly, it is not entirely accurate to say that Meta has completely bought out Wang, as he will continue to serve as the CEO of Scale AI. This means that Wang and Scale AI will remain "independent." This could be the most expensive "straddling two boats" in history. If Scale AI continues its growth momentum, Wang may become the fastest-growing entrepreneur in Silicon Valley in terms of net worth, bar none.

Zuckerberg's eagerness to bet on Scale AI and Wang with such a rare amount of money reflects his anxiety over Meta's gradual falling behind in the AI race.

Despite launching the Llama 4 Behemoth with a parameter scale of 1.8 trillion in 2024, Meta still lags behind GPT-4.5 by about 12% in key indicators such as multimodal understanding and long-text reasoning. What's more embarrassing is that the quality issues of Llama's training data have been exposed: it is estimated that about 30% of the data comes from low-quality social media content, causing the model to frequently output incorrect information.

The Scale AI team, just two years after its establishment. Wang himself is on the far left. | Image source: Scale AI

"We are not short of computing power; what we lack is clean data and top-notch engineering talent," an anonymous Meta AI researcher complained. This explains why Zuckerberg is willing to spend a large sum of money to bring in Wang—a "construction king" known for his data annotation technology.

As the highest-valued data annotation company, Scale AI's rise to fame is not without reason. According to reports, Scale AI's competitive edge lies in its ability to transform raw data into fuel for AI:

Military-grade annotation accuracy: By combining human annotators with AI quality checks as a "double insurance," the company achieves a data error rate of only 0.3%, compared to the industry average of 5% (as claimed by the company).

Multimodal data monopoly: It possesses the world's largest video action annotation library (with 120 million human action data points) and a cross-lingual text dataset (covering 217 languages).

In fact, by spending $14.9 billion to acquire "half" of Scale AI and Wang himself, Meta's ambition goes beyond just the AI models themselves.

02 Transitioning to AI Infrastructure to Address B-Side Weaknesses

Data, computing power, and models are the three essential elements in the field of large models. As a social media giant, Meta has natural advantages in data and computing power. However, the term "data" should be put in quotation marks because, although Meta has a large volume of data, if the quality is poor, it is of little use for AI model training.

"Every GPT response you see is backed by 500 data points annotated by us," Wang said. This statement explains Meta's anxiety. While OpenAI uses Scale AI's data to train smarter models, Meta is stuck in the isolated island of its own social data. Acquiring Scale AI is equivalent to directly taking over the "ammunition depot" of its competitors.

Scale AI controls 35% of the global AI training data traffic and serves top clients ranging from the Pentagon to OpenAI. Engineers at the Meta Research Institute privately complained, "When we trained Llama 3, 30% of the computing power was wasted on cleaning junk data, while Scale AI's annotation accuracy can reach 99.7%."

With Scale AI's precise data cleaning and annotation, it is estimated that Meta will reduce its training data contamination rate from 15% to 2%. The training cycle for the next-generation Llama 5 is expected to be shortened by 40%. Insiders revealed that the "Llama 5 Behemoth" being tested has a parameter scale of 3 trillion, specifically aimed at tackling AGI.

Moreover, Scale AI's annotation system has been deeply integrated with Meta's custom AI chip architecture, forming a closed loop of "data annotation - model training - hardware optimization," which could potentially reduce the inference cost of the Llama model to one-third of that of GPT-4o.

It can be said that with the introduction of Scale AI, Meta's Llama model will see significant improvements in training quality, efficiency, and cost.

In fact, the integration of Scale AI could even reshape Meta's entire strategy in the AI competition. Compared to Google and Microsoft, Meta, which lacks a cloud computing platform, has always been limited to the consumer side. With Scale AI's capabilities, Meta plans to provide Scale AI data services through cloud platforms like AWS and Azure, creating an ecosystem similar to Microsoft's "Copilot + OpenAI" and turning competitors into customers.

If data is the oil of the new era, then by acquiring Scale AI, the largest "data refinery," Meta has already taken control of most of the AI infrastructure system.

Of course, whether competitors like OpenAI and Anthropic will buy into this is still unknown. Although Meta has only acquired half of Scale AI (and half of Wang), it is clearly enough to make OpenAI wary of Scale AI's neutral position. Therefore, OpenAI is also accelerating its cooperation with Handshake, a competitor of Scale AI.

However, given Scale AI's overwhelming advantage in data annotation, it is not realistic for companies like OpenAI to immediately cut ties with Scale AI. At least in the short term, AI giants will still need Scale AI's services.

Even if Scale AI's previous clients gradually reduce their orders, Meta and Scale AI have already planned new sources of revenue—government and defense clients. According to reports, Scale AI has already secured government contracts worth over $200 million from the US military. Meanwhile, Scale AI itself is expanding into vertical fields such as defense customization in the AI application layer, and Meta's enterprise sales capabilities and endorsement will undoubtedly provide sufficient momentum for Scale AI's future development.

Industry insiders have rumored that the huge deal between Meta and Scale AI also includes a hidden bet: if Scale AI's revenue growth rate falls below 80% in the next three years, Meta has the right to acquire the remaining shares at a discounted price. This means that Wang not only has to "make Meta AI great again" but also ensure that his own Scale AI continues to grow rapidly in revenue. The B-side business will clearly become a new source of growth for both parties.

For the Meta team, even if Wang joins as the head of the "straddling two boats" Super Intelligence Lab, he can still create a strong "catfish effect." In the Silicon Valley AI community, Meta has always been known for its strong academic atmosphere. The open-source and inclusive nature of Llama is a result of its academic thinking. However, Wang's strong advocacy of "data thinking" will undoubtedly have an impact and bring about changes to Meta's existing AI team.

According to media reports, as soon as Wang joined Meta, he cut three academic projects and pushed the team to transform in a more "realistic" direction.

If there is no interference from antitrust, this huge bet by Meta on Scale AI and Wang himself may reshape Meta's role and development direction in the fierce AI competition. Not only can it quickly close the gap with competitors in the model field, but it can also help this social media giant complete the transition from an application-focused role to an AI infrastructure role.

The essence of this gamble is Meta's attempt to rewrite the rules of AI competition with capital power. As Silicon Valley analyst Sarah Guo said, "When everyone is building cars, Meta has bought the entire highway—no matter who is in the car, they have to pay the toll."