Meta begins testing in-house AI training chip: Reuters
Make us a preferred source
Make us a preferred source on Google
The news: Meta has begun testing its first in-house chip for training artificial intelligence systems as it seeks to become less reliant on external suppliers like Nvidia, according to sources cited by Reuters.
The numbers: Meta plans capital expenditure of up to USD65 billion ($103.4 billion) to expand its AI infrastructure in 2025, with total expenses for the year expected to top out between USD114 billion and USD119 billion.
The context: Meta began a small deployment of the chip, and will increase production for widescale use if the testing goes well. The sources explained that the chip is a “dedicated accelerator” which means it is designed to handle AI-specific tasks, making it more efficient than GPUs that are typically being used for AI workloads.
Meta is reportedly working with Taiwan Semiconductor Manufacturing Co. to produce the chip, and has already completed its ‘tape-out’ phase. The phase costs tens of millions of dollars and involves sending an initial chip design through a chip factory. A failure would require the company to diagnose the problem and start the months-long ‘tape-out’ process again.
Meta has signalled its intent to start using in-house chips by 2026 for training, or for the process of feeding AI models data to teach them how to perform, which is a compute-intensive exercise.
The social media giant has spent billions on GPUs which it currently uses to train its AI models, and is one of Nvidia’s largest customers. However, as seen earlier this year, the long-term value of GPUs has come into question with the emergence of low-cost models by DeepSeek which rely more heavily on inference than incumbent models.
The source: Reuters
