Meta Platforms has launched a trial deployment of its first in-house chip for training artificial intelligence (AI) models in an effort to reduce the massive costs associated with AI infrastructure. Developed in cooperation with TSMC, Meta’s custom silicon is part of a larger industry push to develop alternatives to NVIDIA’s GPUs, which have been a dominant—but pricey—building block of advanced AI deployments.
Meta has invested deeply in AI development, and appears willing to continue the major expenditure needed to be an AI leader going forward. The company has forecast total 2025 expenses of $114 to $119 billion, which includes an impressive $65 billion to bolster its AI build-out.
But while the social media company is willing to bear great cost, its focus on developing custom in-house silicon demonstrates it’s also adopting strategies to reduce the gargantuan costs of AI. Among these strategies: the new in-house training chip is an AI accelerator, a design that can offer energy efficiency because it is geared specifically for AI workloads. Additionally, Meta, unlike many AI leaders, uses open source to build its Llama AI models.
“Meta’s move to develop custom AI training silicon represents a pivotal strategic shift for the company as it looks to reduce its dependence on NVIDIA’s supply-constrained GPUs and accelerate development of the Llama family of open source models,” said Nick Patience, vice president and practice lead for AI, The Futurum Group.
“While the upfront investment is substantial, the long-term economics make sense given Meta’s massive compute requirements and AI ambitions. The question isn’t whether this is the right move—it’s whether Meta can execute quickly enough to narrow the hardware advantage that Google and others have already established with their custom AI accelerators,” he added.
Meta, which owns Facebook, Instagram and WhatsApp, initially struggled to keep pace with the surge of interest in AI in the wake of ChatGPT’s debut in the November 2022. The company was slow to deploy NVIDIA GPUs, and has acknowledged that it had to play catchup on adopting AI hardware. As it shifted course, Meta has become a major customer for NVIDIA.
As part of its course correction in AI, Meta launched its MTIA (Meta Training and Inference Accelerator) series, of which this newest custom chip is a result. While this new chip is designed for AI training, the company has also produced a MTIA chip for inference, an AI technology which drives the front-end interface for user queries.
AI model training is more computationally intensive, while inference requires greater speed and efficiency. The fact that Meta has produced chips for both training and inference is a notable marker in its push to be more self-reliant in creating AI infrastructure.
Pushing its in-house chip development effort even further, the company’s plan for this new training chip is to eventually use it in the Meta AI chatbot, which is a generative AI tool. It’s also been reported that Meta is in talks to acquire South Korean chip firm FuriosaAI, though those discussions remain inconclusive.
Meta is far from alone in its desire to push NVIDIA off its pedestal as the must-have provider of industry-leading GPUs. Advanced Micro Devices (AMD) debuted an AI chip, MI300, that has enjoyed mammoth sales success, including large purchases by the likes of Samsung Electronics and Microsoft. Amazon has upgraded its AI chip Tranium and boasts a customer list that includes Databricks and Ricoh.
