What is the most expensive thing in the 21st century? Talent!
Many years ago, Ge You's lines in "A World Without Thieves" had a high value, and it is still increasing.
On June 10, local time, media reports revealed that Meta will acquire a 49% stake in Scale AI for $14.9 billion (approximately 106.6 billion yuan). Alexandr Wang, the co-founder of the latter, will become the head of Meta's newly established "Super Intelligence Group."
Based on the equity ratio, Wang and the team could potentially receive $7.4 billion from this transaction, making it the most expensive "poaching" in Silicon Valley history—it's worth noting that Google acquired the DeepMind team for only $600 million in 2014.
Mark Zuckerberg wrote in an internal letter: "We will build the future of AI together." Given the reality that the Llama 4 model has failed and the AI team's personnel are continuously leaving, what is Meta betting on by investing heavily in Scale AI? With Scale AI and Alexandr Wang, can Meta find its place again in the upcoming AI war?
01 The Most Expensive "Wobbler"
As the fastest rising company in Silicon Valley during the AI era, Scale AI's valuation has skyrocketed at a rocket pace, swelling to 13.8 billion dollars in just 5 years. However, the cost for Meta to acquire 49% of the former's shares will be 14.9 billion dollars.
49% is clearly for antitrust review considerations, but what Meta and Zuckerberg want is the person of one of the co-founders, Alexandr Wang—this 19-year-old entrepreneurial genius will become the head of Meta's newly established superintelligence lab, leading Meta AI into a new era.
Interestingly, it's not entirely accurate to say that Meta has completely acquired Wang, because Wang will continue to serve as the CEO of Scale AI, meaning that Wang and Scale AI will still maintain their "independence." This could also be the most expensive "dual allegiance" in history, and if Scale AI continues to grow, Wang may become the fastest-growing entrepreneur in Silicon Valley, bar none.
Zuckerberg's urgent and impatient investment in Scale AI and Wang, with a rare amount from Meta, reflects his anxiety about Meta gradually falling behind in the AI race.
Although Meta launched Llama 4 Behemoth in 2024 with a parameter size of 1.8 trillion, it still lags behind GPT-4.5 in key metrics such as multimodal understanding and long-text inference. To add insult to injury, the quality of Llama's training data was exposed: an industry estimate that about 30% of the corpus comes from low-quality social media content, resulting in frequent error messages from the model.
The Scale AI team, which was established just 2 years ago, with Wang himself on the far left | Image source: Scale AI
"What we lack is not computing power, but clean data and top engineering talent." An anonymous Meta AI researcher expressed. This explains why Zuckerberg is spending heavily to hire Wang—a "infrastructure fanatic" known for his data labeling technology.
As the highest-valued data labeling company, Scale AI's rise to prominence is not without reason. According to reports, Scale AI's moat lies in its ability to convert raw data into fuel usable by AI:
Military-grade annotation accuracy: With the "double insurance" of mixed human annotators + AI quality inspection, its data error rate is only 0.3%, while the industry average is 5% (company statement).
Multimodal Data Monopoly: Possessing the world's largest video action annotation database (including 120 million human action data) and a cross-lingual text dataset (covering 217 languages).
In fact, by spending 14.9 billion dollars to acquire "half" of Scale AI and Wang himself, Meta's ambition is not limited to the AI large model itself.
02 Transforming AI Infrastructure to Address B-Side Shortcomings
Data, computing power, and models are the three essential elements in the field of large models. As a social media giant, Meta has a natural advantage in data and computing power, but the term "data" needs to be put in quotation marks because while Meta has a large amount of data, if the quality is poor, it won't be very useful for AI model training.
"Every GPT response you see is backed by 500 data points that we have labeled." Wang's statement explains Meta's anxiety. While OpenAI trains smarter models using Scale AI's data, Meta is trapped on its own social data island. Acquiring Scale AI is equivalent to directly taking over the competitor's "arsenal."
Scale AI holds 35% of the global AI training data flow, serving top clients from the Pentagon to OpenAI. An engineer from Meta's research institute privately complained: "When we trained with Llama 3, 30% of the computing power was wasted on cleaning up junk data, while Scale AI's labeling accuracy can reach 99.7%."
With Scale AI's precise data cleaning and labeling, it is estimated that Meta will reduce the training data contamination rate from 15% to 2%, shortening the training cycle of the next generation Llama 5 by 40%. Insiders revealed that the "Llama 5 Behemoth" currently being tested has a parameter scale of 30 trillion, specifically designed to tackle AGI.
At the same time, Scale AI's annotation system has been deeply adapted to the Meta 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 1/3 of GPT-4o.
It can be said that after the introduction of Scale AI, Meta's Llama model will achieve significant optimization in terms of training quality, efficiency, and cost.
In fact, Scale's access may even reshape Meta's entire strategy in the AI race. Compared with Google and Microsoft, Meta, which lacks a cloud computing platform, has been able to go wild on the C side. With Scale's capabilities, Meta plans to provide Scale AI data services through cloud platforms such as AWS/Azure to build an ecological closed loop similar to Microsoft's "Copilot + OpenAI" and convert competitors into customers.
If data is the oil of the new era, then Meta has already mastered a large part of the AI infrastructure by purchasing Scale AI, the largest "data refinery".
Meta is gradually falling behind in the AI competition | Image source: Meta
Of course, it remains to be seen whether competitors like OpenAI and Anthropic will actually take the bait. Although Meta has only acquired half of Scale AI (and half of Wang), it is clearly enough to make the former wary of Scale AI's neutral position. Therefore, OpenAI is also intensifying its collaboration with Scale AI's competitor Handshake.
However, given Scale AI's overwhelming advantages in data labeling, it is unrealistic for companies like OpenAI to immediately sever ties with Scale AI. At least in the short term, the AI giants still need Scale AI's services.
Even as Scale AI's previous customers are tapering down on orders, Meta and Scale AI are already looking for new revenue streams — government and defense customers. According to reports, Scale AI has partnered and has received more than $200 million in government orders from the U.S. military. At the same time, Scale AI itself is also expanding to the AI application layer in vertical fields such as defense customization, and Meta's enterprise-level sales capabilities and endorsements will undoubtedly provide enough impetus for Scale AI's future development.
Insiders say that there is a hidden bet in the huge deal between Meta and Scale AI: if Scale AI's revenue growth in the next three years is less than 80%, 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 that his own Scale AI must continue to grow rapidly in terms of revenue. The B-end business will clearly become a new source of growth for both parties.
For the Meta team, Wang's joining as the head of the "dual-role" super-intelligent laboratory can create a strong "catalyst effect." In the AI scene of Silicon Valley, Meta has always been known for its strong academic atmosphere, and the open-source and inclusive nature of Llama is a result of this academic thinking. However, the "data-driven thinking" that Wang strongly advocates will undoubtedly impact and transform Meta's existing AI team.
According to media reports, Wang has just joined Meta and immediately cut three academic projects, pushing the team to transform towards a more "realistic" direction.
If we do not consider the antitrust obstacles, Meta's substantial bet on Scale AI and Wang himself may reshape Meta's role and development direction in the fierce AI competition, not only allowing Meta to quickly close the gap with competitors in the model field but also enabling this social giant to complete the transition from application to AI infrastructure role.
The essence of this gamble is that Meta is attempting to rewrite the rules of AI competition using capital power. As Silicon Valley analyst Sarah Guo stated: "When everyone is building cars, Meta bought the entire highway - no matter who is in the car, they all have to pay the toll."
The content is for reference only, not a solicitation or offer. No investment, tax, or legal advice provided. See Disclaimer for more risks disclosure.
100 billion sky-high price, Zuckerberg buys "half a genius" and the future of Meta AI
Author: Jingyu
What is the most expensive thing in the 21st century? Talent!
Many years ago, Ge You's lines in "A World Without Thieves" had a high value, and it is still increasing.
On June 10, local time, media reports revealed that Meta will acquire a 49% stake in Scale AI for $14.9 billion (approximately 106.6 billion yuan). Alexandr Wang, the co-founder of the latter, will become the head of Meta's newly established "Super Intelligence Group."
Based on the equity ratio, Wang and the team could potentially receive $7.4 billion from this transaction, making it the most expensive "poaching" in Silicon Valley history—it's worth noting that Google acquired the DeepMind team for only $600 million in 2014.
Mark Zuckerberg wrote in an internal letter: "We will build the future of AI together." Given the reality that the Llama 4 model has failed and the AI team's personnel are continuously leaving, what is Meta betting on by investing heavily in Scale AI? With Scale AI and Alexandr Wang, can Meta find its place again in the upcoming AI war?
01 The Most Expensive "Wobbler"
As the fastest rising company in Silicon Valley during the AI era, Scale AI's valuation has skyrocketed at a rocket pace, swelling to 13.8 billion dollars in just 5 years. However, the cost for Meta to acquire 49% of the former's shares will be 14.9 billion dollars.
49% is clearly for antitrust review considerations, but what Meta and Zuckerberg want is the person of one of the co-founders, Alexandr Wang—this 19-year-old entrepreneurial genius will become the head of Meta's newly established superintelligence lab, leading Meta AI into a new era.
Interestingly, it's not entirely accurate to say that Meta has completely acquired Wang, because Wang will continue to serve as the CEO of Scale AI, meaning that Wang and Scale AI will still maintain their "independence." This could also be the most expensive "dual allegiance" in history, and if Scale AI continues to grow, Wang may become the fastest-growing entrepreneur in Silicon Valley, bar none.
Zuckerberg's urgent and impatient investment in Scale AI and Wang, with a rare amount from Meta, reflects his anxiety about Meta gradually falling behind in the AI race.
Although Meta launched Llama 4 Behemoth in 2024 with a parameter size of 1.8 trillion, it still lags behind GPT-4.5 in key metrics such as multimodal understanding and long-text inference. To add insult to injury, the quality of Llama's training data was exposed: an industry estimate that about 30% of the corpus comes from low-quality social media content, resulting in frequent error messages from the model.
The Scale AI team, which was established just 2 years ago, with Wang himself on the far left | Image source: Scale AI
"What we lack is not computing power, but clean data and top engineering talent." An anonymous Meta AI researcher expressed. This explains why Zuckerberg is spending heavily to hire Wang—a "infrastructure fanatic" known for his data labeling technology.
As the highest-valued data labeling company, Scale AI's rise to prominence is not without reason. According to reports, Scale AI's moat lies in its ability to convert raw data into fuel usable by AI:
Military-grade annotation accuracy: With the "double insurance" of mixed human annotators + AI quality inspection, its data error rate is only 0.3%, while the industry average is 5% (company statement).
Multimodal Data Monopoly: Possessing the world's largest video action annotation database (including 120 million human action data) and a cross-lingual text dataset (covering 217 languages).
In fact, by spending 14.9 billion dollars to acquire "half" of Scale AI and Wang himself, Meta's ambition is not limited to the AI large model itself.
02 Transforming AI Infrastructure to Address B-Side Shortcomings
Data, computing power, and models are the three essential elements in the field of large models. As a social media giant, Meta has a natural advantage in data and computing power, but the term "data" needs to be put in quotation marks because while Meta has a large amount of data, if the quality is poor, it won't be very useful for AI model training.
"Every GPT response you see is backed by 500 data points that we have labeled." Wang's statement explains Meta's anxiety. While OpenAI trains smarter models using Scale AI's data, Meta is trapped on its own social data island. Acquiring Scale AI is equivalent to directly taking over the competitor's "arsenal."
Scale AI holds 35% of the global AI training data flow, serving top clients from the Pentagon to OpenAI. An engineer from Meta's research institute privately complained: "When we trained with Llama 3, 30% of the computing power was wasted on cleaning up junk data, while Scale AI's labeling accuracy can reach 99.7%."
With Scale AI's precise data cleaning and labeling, it is estimated that Meta will reduce the training data contamination rate from 15% to 2%, shortening the training cycle of the next generation Llama 5 by 40%. Insiders revealed that the "Llama 5 Behemoth" currently being tested has a parameter scale of 30 trillion, specifically designed to tackle AGI.
At the same time, Scale AI's annotation system has been deeply adapted to the Meta 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 1/3 of GPT-4o.
It can be said that after the introduction of Scale AI, Meta's Llama model will achieve significant optimization in terms of training quality, efficiency, and cost.
In fact, Scale's access may even reshape Meta's entire strategy in the AI race. Compared with Google and Microsoft, Meta, which lacks a cloud computing platform, has been able to go wild on the C side. With Scale's capabilities, Meta plans to provide Scale AI data services through cloud platforms such as AWS/Azure to build an ecological closed loop similar to Microsoft's "Copilot + OpenAI" and convert competitors into customers.
If data is the oil of the new era, then Meta has already mastered a large part of the AI infrastructure by purchasing Scale AI, the largest "data refinery".
Meta is gradually falling behind in the AI competition | Image source: Meta
Of course, it remains to be seen whether competitors like OpenAI and Anthropic will actually take the bait. Although Meta has only acquired half of Scale AI (and half of Wang), it is clearly enough to make the former wary of Scale AI's neutral position. Therefore, OpenAI is also intensifying its collaboration with Scale AI's competitor Handshake.
However, given Scale AI's overwhelming advantages in data labeling, it is unrealistic for companies like OpenAI to immediately sever ties with Scale AI. At least in the short term, the AI giants still need Scale AI's services.
Even as Scale AI's previous customers are tapering down on orders, Meta and Scale AI are already looking for new revenue streams — government and defense customers. According to reports, Scale AI has partnered and has received more than $200 million in government orders from the U.S. military. At the same time, Scale AI itself is also expanding to the AI application layer in vertical fields such as defense customization, and Meta's enterprise-level sales capabilities and endorsements will undoubtedly provide enough impetus for Scale AI's future development.
Insiders say that there is a hidden bet in the huge deal between Meta and Scale AI: if Scale AI's revenue growth in the next three years is less than 80%, 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 that his own Scale AI must continue to grow rapidly in terms of revenue. The B-end business will clearly become a new source of growth for both parties.
For the Meta team, Wang's joining as the head of the "dual-role" super-intelligent laboratory can create a strong "catalyst effect." In the AI scene of Silicon Valley, Meta has always been known for its strong academic atmosphere, and the open-source and inclusive nature of Llama is a result of this academic thinking. However, the "data-driven thinking" that Wang strongly advocates will undoubtedly impact and transform Meta's existing AI team.
According to media reports, Wang has just joined Meta and immediately cut three academic projects, pushing the team to transform towards a more "realistic" direction.
If we do not consider the antitrust obstacles, Meta's substantial bet on Scale AI and Wang himself may reshape Meta's role and development direction in the fierce AI competition, not only allowing Meta to quickly close the gap with competitors in the model field but also enabling this social giant to complete the transition from application to AI infrastructure role.
The essence of this gamble is that Meta is attempting to rewrite the rules of AI competition using capital power. As Silicon Valley analyst Sarah Guo stated: "When everyone is building cars, Meta bought the entire highway - no matter who is in the car, they all have to pay the toll."