Behind the Testing of Meta’s First AI Training Chip

In the race to dominate AI, technology giants are now increasingly looking to control the software as well as the hardware that powers it.

The global semiconductor shortage of recent years, coupled with soaring prices for specialised AI chips and uncertain regulations, has pushed companies to reconsider their supply chains and explore vertical integration strategies.

Now, Meta has begun testing its first in-house chip designed for training AI systems, according to two sources who spoke to Reuters.

This development signals the company's strategy to design more of its own custom silicon and decrease its dependence on external suppliers such as

Nvidia, who is a major manufacturer of graphics processing units (GPUs) used for AI computing tasks within Meta and across the world.

Meta’s move comes amid fierce competition in the AI space, where companies like Google, Amazon and Microsoft have already invested in developing their own custom chips.

However, these tech giants are innovating chips independently for a multitude of reasons – including a growing recognition that control over semiconductor design and supply can provide strategic advantages in both cost management and technical capabilities as AI workloads continue to grow exponentially.

How Meta's custom silicon strategy aims to control high AI infrastructure costs

Meta spent an estimated US$10bn on Nvidia GPUs alone in 2023, according to industry analysts – and the company has further projected total expenses for 2025 between US$114bn and US$119bn, including up to US$65bn in capital expenditure, predominantly driven by spending on AI infrastructure.

Yet by developing its own chips tailored to its specific AI needs, the company could potentially reduce these costs significantly over time while gaining greater control over its technological roadmap.

The company has initiated a small-scale deployment of the chip and intends to increase production for widespread use if the test proves successful, the sources said.

According to Reuters, the development of in-house chips is part of a long-term plan at Meta to reduce its substantial infrastructure costs as the company invests heavily in AI tools to drive growth.

One source explained that Meta's new training chip functions as a dedicated accelerator, meaning it is specifically designed to handle AI-related tasks – which can result in greater power efficiency compared to the integrated GPUs typically used for AI workloads.

Meta is also collaborating with Taiwan Semiconductor Manufacturing Company (TSMC) to manufacture the chip, this person added.

The test deployment commenced after Meta completed its first “tape-out” of the chip, a crucial indicator of success in silicon development that involves sending an initial design through a chip factory, according to the second source.

A typical tape-out costs tens of millions of pounds and requires approximately three to six months to complete, with no guarantee of success.

This means that if the test fails, Meta would need to identify the problem and repeat the tape-out process.

Meta training and inference accelerator programme shows progress despite setbacks

The chip is the latest addition to the company's Meta Training and Inference Accelerator (MTIA) series – despite the programme having experienced an uneven start for several years and at one point abandoned a chip at a similar development stage.

However, Meta began using an MTIA chip last year for inference – the process involved in running an AI system during user interaction – for the recommendation systems that determine which content appears on Facebook and Instagram news feeds.

Now, Meta executives have stated their intention to begin using their own chips by 2026 for training, the compute-intensive process of feeding an AI system large amounts of data to “teach” it how to perform.

Similar to the inference chip, the goal for the training chip is to begin with recommendation systems and later expand its use to Gen AI products like the chatbot Meta AI, according to executives.

“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,” says Chris Cox, Meta's Chief Product Officer, at the Morgan Stanley technology, media and telecom conference last week.

Chris described Meta's chip development efforts as “kind of a walk, crawl, run situation” thus far, but said executives considered the first-generation inference chip for recommendations to be a “big success.”

Explore the latest edition of Technology 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.

Technology Magazine is a BizClik brand