Meta is DIYing chips

Just like your new year's resolution to make coffee at home instead of buying it everyday, Meta is trying to wean itself off outside vendors.

Today, Meta announced it’s rolling out four of its own custom AI chips to fuel its rapid data center expansion. The first, MTIA 300, is meant to help train smaller AI models that run Meta’s recommendations on Facebook and Instagram. The next three—MTIA 400, MTIA 450, and MTIA 500—are going to be used for more advanced tasks like creating images and videos based on prompts. A new chip will be released every six months, a shockingly rapid rate that underscores Meta’s growing need for new tech.

Zoom out: By designing its own chips, which are then manufactured by Taiwan Semiconductor, the social media giant can lower its very expensive capex tab, Meta Vice President of Engineering Yee Jiun Song told CNBC.

Meta expects to spend up to $135 billion on AI this year, including opening a massive data center in Louisiana, as well as building more in Ohio, Indiana, and Texas. But part of that spending spree also includes tens of billions of dollars on GPUs, mostly from Nvidia. Homegrown chips can not only reduce costs, but also let Meta optimize chips for specific tasks.

Tech wants to be vertically integrated

As we all know, Nvidia has been dominating the world of AI chips for years. The company controls 92% of the global data center GPU market—and as the reigning king of AI hardware, Nvidia can afford to make its chips expensive, because demand is so strong and Nvidia has all the bargaining power.

Now, big tech companies are trying to gain some leverage by not only creating chips of their own, but building more specialized chips—application-specific integrated circuits designed for narrower tasks that don’t need energy-intensive GPUs:

• Google was early to the party, creating its first Tensor Processing Unit (TSU) in 2015. The Mag 7 company struck a $10 billion deal with Anthropic in December to sell the startup its TPUs, and some analysts think they could grow to become a $900 billion business.
• Amazon created Trainium, its first custom chip, in 2018. Anthropic has since bought millions of dollars worth of the newest version, Trainium 2, which isn’t as powerful as Nvidia’s offerings but is more cost effective.

Rolling out its own chips doesn’t mean Meta has weaned off Nvidia completely though: Just last month Meta inked a series of major deals with Nvidia and AMD to get its hands on new GPUs.

Look, we all need that store-bought caramel macchiato sometimes.—LB