Meta Tests First In-House Chip for Training AI Systems

Meta is testing its first in-house chip for AI training, according to sources speaking to Reuters. 

This move comes as part of a broader push by technology companies to gain control over both software and hardware amid rising chip costs and regulatory uncertainties. Market challenges and complexities have pushed companies to consider proprietary alternatives and explore vertical integration strategies.

Meta’s move indicates a change in direction towards designing more of its own in-house solutions and slowing its dependence on third-party suppliers like Nvidia, which provides  graphics processing units (GPUs) that support AI innovation.

Competitors to Meta, including Google, Amazon and Microsoft are also pursuing the development of their own custom silicon solutions.

Meta looks to reduce AI infrastructure costs

Meta’s decision to invest in its own chip design is driven by the high cost of AI infrastructure. Industry analysts estimate that the company spent around US$10bn on Nvidia GPUs in 2023 alone. 

Looking ahead, Meta has projected total expenses for 2025 to be between US$114bn and US$119bn, with up to US$65bn allocated to capital expenditure, primarily for AI-related projects.

Developing proprietary chips allows Meta to tailor hardware to its specific needs, potentially improving efficiency and lowering long-term costs. The company has started small-scale deployment of the chip and, if tests are successful, plans to scale up production.

According to Reuters, Meta sees in-house chip development as a long-term strategy to manage infrastructure expenses while continuing to invest in AI tools. 

One source stated that the chip serves as a dedicated accelerator, optimised for AI processing, which can be more power-efficient than conventional GPUs. Meta is also working with Taiwan Semiconductor Manufacturing Company (TSMC) for production, said the source.

The testing phase follows the completion of the chip’s first “tape-out” – a crucial step in chip manufacturing where the initial design is sent to fabrication.

A tape-out can take between three and six months and cost tens of millions of dollars, with no certainty of success. If the test fails, Meta will need to refine the design and repeat the process.

Meta's AI chip programme progresses despite challenges

The newly developed chip is part of Meta’s Meta Training and Inference Accelerator (MTIA) series. The programme has faced setbacks in the past, including the abandonment of a chip at a similar stage of development. 

However, progress has been made, and last year, Meta deployed an MTIA chip for inference tasks—where AI models process and generate responses—within its recommendation systems on Facebook and Instagram.

Now, Meta aims to extend its custom silicon strategy to AI training. AI training involves feeding large datasets into a model to enhance its capabilities, a process requiring significant computing power. 

According to Meta executives, the company plans to integrate its own training chips into its systems by 2026.

Initially, these chips will support recommendation systems, with plans to expand into generative AI (Gen AI) applications such as Meta AI, the company’s chatbot.

Meta’s long-term vision for AI hardware

Meta Chief Product Officer Chris Cox discussed the company’s approach at the Morgan Stanley technology, media and telecom conference last week. 

He said: “We're working on how would we do training for recommender systems and then eventually how do we think about training and inference for gen AI.”

Chris framed Meta’s AI chip development as “kind of a walk, crawl, run situation” but noted that the first-generation inference chip had been a “big success.”

Explore the latest edition of AI Magazine and be part of the conversation at our global conference series, Tech & AI LIVE. 

Discover all our upcoming events and secure your tickets today.

AI Magazine is a BizClik brand