Meta plans to extend its self-developed chips to AI model training
Meta's Chief Financial Officer Susan Li stated that the company remains ambitious in developing its own custom chips, including plans to develop processors that can train future artificial intelligence models. She mentioned that Meta's self-developed chips are currently primarily deployed in ranking and recommendation workloads, "but we anticipate and hope to expand this scope over time, eventually extending to the training of AI models."
Meta plans to develop custom chips to train its artificial intelligence models.
On Wednesday, March 4th, Meta's Chief Financial Officer Susan Li stated at a technology industry conference hosted by Morgan Stanley that the company remains ambitious in developing its own custom chips, including plans to develop processors that can train future AI models.
Susan Li mentioned that Meta's self-developed chips are currently mainly deployed in ranking and recommendation workloads. She said:
But we expect and hope to expand this scope over time, including eventually extending to the training of AI models.
Meta positions itself as one of the large-scale operators of data centers for training and running AI models, even though the company itself is not a cloud service provider.
In the weeks prior, Meta has successively reached large-scale procurement agreements with chip market leaders NVIDIA and its competitor AMD to support the chip and equipment supply needed for AI workloads.
Susan Li emphasized that Meta adopts a differentiated chip procurement strategy, selecting the most suitable processors based on different application scenarios. She said:
Based on the information we currently have and the current demand, which chip do we think is best suited for each application scenario? Custom chips are a key part of this.
Analysts believe that Meta is not replacing externally sourced chips with self-developed ones, but rather building a hybrid supply system. External procurement meets current scaling needs, while self-developed chips are optimized for highly customized internal workloads to improve efficiency and control long-term costs.
