One year after Meta rebuilt its AI organization: The super-intelligence ambition underpinned by triple advantages in data, talent, and computing.

行业背景：Meta的AI重建元年

2025年中，Llama 4的发布成为Meta AI史上一次重大挫折。模型表现不及预期，导致CEO马克·扎克伯格决定彻底重组AI组织，成立Meta超智能实验室（Meta Superintelligence Lab，MSL）。此后一年间，MSL经历了从组建到初步成果的转变，其最新模型Muse Spark于2026年4月发布，尽管在主流基准测试上落后于同期开源模型DeepSeek v4 Pro和Kimi K2.6，但评估员普遍认为，当前表现反映的是重建初期的“债务偿还”，而非长期能力上限。

市场影响：从“开源冠军”到“闭门追赶”

Meta过去以Llama系列开源模型闻名，Llama 3 70B和3.1 405B均为发布时的开源SOTA。但MSL的Muse Spark选择了闭源路线，且性能未达顶尖。这一转变向市场传递了明确信号：Meta暂时放弃了开源领先地位，转而集中资源追求前沿能力。投资者和客户需关注的是斜率而非截距——即MSL未来6个月的能力提升速度，而非当前基准表现。

竞争格局：OpenAI vs Anthropic双雄，Meta成潜在搅局者

当前前沿AI竞赛已明显集中于OpenAI与Anthropic之间。Google的Gemini 3 Pro和Nano Banana短暂亮眼后迅速失去势头，其Windsurf收购未能转化为有竞争力的Agent编码产品；微软在GitHub Copilot上的先发优势亦未保持。据SemiAnalysis分析，MSL拥有其他实验室难以复制的三重优势：

1. 数据：Meta通过追踪员工屏幕、键盘和鼠标活动，获取了全球最宝贵的RL环境数据之一。更关键的是，2026年5月Meta重组中，约3000名工程师（包括70%应届毕业生和大量资深人员）被划入新设立的“应用AI工程组织”，全职制作RL任务与环境。这一规模远超任何外部数据公司——例如，头部RL环境公司Mercor在2026年第二季度仅记录了4800全职等效小时的专家工时。

2. 人才：MSL从Scale AI“投资”14.3亿美元实质上挖走了其CEO Alexandr Wang及其安全、评估与对齐实验室（SEAL）核心团队，并提供数千万至数亿美元级别的薪酬包吸引顶级AI研究员。3. Computing Power: Meta has adopted an innovative "tent" data center design, achieving aggressive computing power expansion. According to SemiAnalysis's new Tokenomics model, Meta's AI computing investment growth rate is the fastest among major vendors.

Business Insights: RL Environment Data Becomes the New Scarce Resource

For corporate decision-makers, Meta's strategy reveals the key to the next phase of AI capability competition: high-quality RL environments and screen recording data. Traditional SFT and benchmark data are rapidly depreciating, while screen recording data that reflects real workflows (especially in specialized fields like finance, law, and advertising) becomes a strategic asset for training agent systems. Companies should assess the potential value of their own workflow data and consider establishing data partnerships with AI labs.

Furthermore, Meta's practice of forming a 3,000-person full-time RL environment team shows that even with abundant general computing resources, building customized RL environments still requires significant human investment. This reminds companies that the bottleneck for AI deployment has shifted from computing power to data engineering and task design capabilities.

Future Outlook: Can MSL Surpass Google in 12-24 Months?

SemiAnalysis believes that MSL has a strong possibility of surpassing Google's AI capabilities within 6 months. The core variables are: the maturity of the data feedback loop (the feedback cycle of screen recording data entering model training), the task output efficiency of the 3,000-person team, and the execution of computing power expansion as planned. If all goes well, MSL will enter the top three (OpenAI, Anthropic, Meta) by 2027, while Google may fall to the second tier.

In the long term (3 years), if MSL continues to maintain world-class levels in data, talent, and computing, it has the opportunity to challenge the leading positions of OpenAI and Anthropic. But the key prerequisites are: whether the internal culture can tolerate years of high investment and low visibility, and whether data privacy controversies will affect employee engagement.