Meta Platforms is deploying a new home-grown chip to help power its artificial intelligence (AI) services, aiming to decrease its reliance on semiconductors from Nvidia and other outside companies.
The chip, announced Wednesday, is the latest version of the Meta Training and Inference Accelerator, or MTIA, which helps rank and recommend content across Facebook and Instagram. Meta released the first MTIA product last year.
Meta’s pivot to AI services has brought increased demand for computing power. Last year, the social media giant released its own version of an AI model to compete with OpenAI’s ChatGPT. It also added new generative AI features to its social apps, including customised stickers and celebrity-faced chatbot characters.
Meta’s annual sales boosted by Chinese brands seeking global exposure
Meta’s annual sales boosted by Chinese brands seeking global exposure
In October, the company said it would spend as much as US$35 billion on infrastructure to support AI, including data centres and hardware. “AI will be our biggest investment area in 2024,” Chief Executive Officer Mark Zuckerberg told investors that month.
A significant amount of that spending is likely to still flow to Nvidia, which builds the popular H100 graphics cards that power AI models. Earlier this year, Zuckerberg said the company would acquire 350,000 of those chips, which cost tens of thousands of dollars each.
But there is a growing movement among tech giants to develop chips in-house. Meta is joining rivals Amazon’s AWS, Microsoft and Alphabet’s Google in trying to wean themselves off a very expensive dependency. It will not be a quick fix, though. So far, the efforts have not made a dent in the industry’s insatiable need for Nvidia’s AI accelerators.
The AI boom has helped turn Nvidia into the world’s third-most-valuable tech company, behind only Microsoft and Apple. Its sales to data centre operators totalled US$47.5 billion in fiscal 2024, up from just US$15 billion the year before. Analysts predict that the sum will more than double again in fiscal 2025.
