Meta would plan to invest more than $ 10 billion in Scale AI, a startup specializing in data annotation for artificial intelligence, according to information reported by Bloomberg. If the operation is materialized, it would not only become one of the most important private funding ever made in the sector, and would also mark a strategic turning point for Meta, historically attached to an internalized R&D.
TL; DR – META Puts SCALE AI: A major strategic turn to IA infrastructure
👥 For whom is it important?
- Strategic AI and cloud infrastructure decision -makers
- Tech investors monitoring capital movements in AI
- Public and private actors involved in civil or military uses of AI
- Startups positioned on the upstream value chains of AI
💡 Why is it strategic?
- Meta breaks with its internalized R&D doctrine by considering investing> $ 10 billion in Scale Ai
- SCALE AI masters a critical layer: the data annotated on a large scale, pillar of generative AI
- The operation would strengthen Meta against Microsoft and Amazon, already integrated at all levels of the Stack IA
- It is part of an increasing convergence between civil and military (Defense Llama, DOD contract)
🔧 What it changes concretely
- Meta would consolidate her IA training chain without depending exclusively on her internal infrastructure
- SCALE AI would become a key supplier for large models and autonomous agents
- The annotation market is industrialized: current fragmentation, but rise in power of strategic platforms
- Companies will have to combine internal tools and specialized service providers to secure their IA pipeline
A break in Meta’s strategy in terms of AI
For several years, Meta has been distinguished from its competitors by its choice of an autonomous and open source approach in the development of its AI models. Unlike Microsoft, Amazon or Google, which have invested in Openai, Anthropic and other proprietary models respectively, Meta has bet on Llama, its internal infrastructure and an open ecosystem.
The investment envisaged in Scale AI seems to be derogating from this rule and the amount mentioned, greater than $ 10 billion, would place this commitment at the same level as that of Microsoft in Openai. It would be the biggest external operation ever carried out by Meta in AI whose exact format of the deal is not known at this stage.
SCALE AI, essential link in the training chain
Founded in 2016 by Alexandr Wang, SCALE AI provides structuring and annotation of data to cause automatic learning models. The company is collaborating with Microsoft, Openai, and more recently, with the American Department of Defense. The startup has been in hypercroissance since its creation and generates significant income, with $ 870 million in turnover in 2024, and a forecast at 2 billion for 2025, according to Bloomberg.
The startup was valued at $ 14 billion in its last round in 2024. A secondary operation, mentioned at the start of the year, suggested a valuation of $ 25 billion. At this point, none of the actors concerned wanted to comment on the negotiations.
The main asset of Scale AI is based on its ability to produce, structure and make massive volumes of data. These data are used to cause large -scale models, but also to calibrate AI agents intended for specific, civil as military uses. This mastery of labeled data is a critical resource of generative AI.
A convergence around defense uses
This rapprochement is also part of a context of increasing cooperation between technological companies and the defense sector. Meta recently announced a partnership with Andundil Industries for the development of caps increased for military use. It now explicitly authorizes the use of its LLAMA models by American government agencies.
At the same time, Scale has strengthened its presence in this area. The company has signed a contract with the Department of Defense to develop AI agents and works on a militarized version of Llama, called “Defense Llama”, in partnership with Meta. This shift to duals (civil and military) uses could explain, in part, Meta’s strategic interest in Scale.
Behind the models, the battle of infrastructure
The attention paid to AI models (GPT-4, Claude, Gemini) masks a more structural reality: The performance of an AI system depends largely on the infrastructure on which it is based. This infrastructure is not limited to fleas. It includes several critical layers, more and more interdependent:
- Data : Without annotated, specialized, representative datasets, models can neither learn nor generalize. This is the role of actors like Scale AI or Sama.
- The calculation : GPUs, in particular those of Nvidia (H100, A100), remain at the heart of training and inference. Tensions on supply, TSMC dependence, and the emergence of alternatives like the TPUs of Google or the Puces of Amazon (Trainium, Inferentia) reconfigure the balance of power.
- Software frameworks : tools like Pytorch, Tensorflow or Jax allow researchers and engineers to build and experiment quickly. Their control becomes strategic, as shown in Pytorch’s handling by Linux Foundation.
- Orchestration and deployment : Training a model is only the first step. Version management, supervision, scaling and integration into business flows are ensured by platforms such as Weights & Biasses, Hugging Face, Mosaicml or specialized cloud environments (Azure ML, Sagemaker, Vertex AI).
- Security and governance : The use of autonomous agents in sensitive contexts (health, defense, finance) requires new requirements in terms of traceability, robustness and human supervision.
In this landscape, Scale Ai embodies a actor of the “upstream” infrastructurethat which precedes training but conditions its validity. The emergence of these intermediate layers marks an increasing industrialization of AI, where differentiation no longer goes only by models, but by mastery of the entire algorithmic logistics chain.
The data annotation market, between fragmentation and specialization
The data annotation market remains fragmented, structured around several complementary approaches. SCULE AI occupies a central position on large -scale projects, with a high degree of automation, proprietary software infrastructure and orientation to strategic, public and private customers. Other actors bet on differentiated positions, Samafor example, outsourcing its operations in Africa and is positioned on an “ethical” approach to annotation, with an anchoring in social impact. On the other hand, Snorkel ai offers an annotation without human intervention, based on programmatic rules, in a logic “Data-Centric AI”.
Other solutions such as Labelbox Or Hive Ai Allow the internal teams of companies to keep hands on their data pipelines, via SaaS platforms or hybrid models. Finally, providers like Cloud offer outsourced, low -cost annotation services for high volume and standardization needs.
The current trend is at combination of several approaches, Conjugant internal tools for sensitive or owners, and specialized providers to speed up scaling. In this landscape, Scale AI is distinguished by its ability to treat massive corpus while adapting to complex use cases, especially in the military, industrial or regulatory fields.
By choosing to support Scale rather than a competing model editor, Meta bets on a critical infrastructure, difficult to reproduce on a large scale, and capable of serving its ambitions both commercial and strategic. The envisaged investment would rebalance its positions in a competition where Microsoft and Amazon have already taken a step ahead thanks to their double presence on the models and on the cloud.
