MTIA

MTIA is a series of computer chips developed in-house by Meta – the company behind Facebook, Instagram, and WhatsApp. The abbreviation stands for “Meta Training and Inference Accelerator.” An accelerator is a specialized chip that performs one single type of computing task extremely fast, but can barely do anything else. In the case of MTIA, that task is computation for artificial intelligence, meaning programs that learn patterns from data. Meta doesn’t build the chips for sale, but only for its own data centers. The first version was publicly introduced in 2023, followed by a second version in 2024.

Why Meta builds its own chips instead of buying them

So far, most of the world’s AI systems run on graphics chips made by Nvidia. These chips are scarce and expensive. A single top-tier model temporarily cost more than $30,000, and Meta buys hundreds of thousands of them. Anyone ordering that much from a single supplier has little negotiating power and is dependent on that supplier’s delivery times.

An in-house chip partially solves this problem. Meta only needs to make it good for its own tasks, not for all possible customers. This allows components that the company never needs to be left out. The result is a chip that delivers more useful computing work per watt of power consumed. In data centers that together draw as much electricity as a small town, that is a major cost factor.

However, MTIA does not fully replace Nvidia. Meta itself emphasizes that both types of chips run in parallel. For training very large language models like Llama, the purchased graphics chips remain the standard choice. MTIA initially takes over the tasks that are unspectacular but occur in enormous volume.

What the chip actually computes

Meta’s most important AI application is not a chatbot, but the recommendation system. It decides which post appears next in your feed and which ads you see. To do this, the system has to make a selection from millions of possibilities within milliseconds. This calculation doesn’t run once a day — it runs every single time someone, anywhere in the world, opens the app.

Such recommendation models rely heavily on so-called embeddings. These are long lists of numbers stored for each user and each video, describing their characteristics. The chip’s main job is to retrieve many of these lists from memory and compare them with each other. So it needs less extreme raw computing power and more very fast access to large amounts of data. That is exactly what MTIA is tailored for.

Technically, several MTIA chips sit together on a single card, several cards in a server, and many servers in a rack. Meta also supplies the software that translates existing AI programs to run on the new hardware. This software is often the bigger hurdle than the chip itself. A chip without matching programming tools remains unusable for developers.

MTIA in the news and in your own feed

You will never touch MTIA directly. The chip sits inside Meta’s data centers, and you only notice it indirectly: if your Instagram feed shows fitting suggestions, a chip like this probably helped compute that. There is no MTIA available for consumers to buy, unlike, say, graphics cards for gaming.

In business news, the name usually comes up in a specific context. Analysts then ask how strong Nvidia’s market position still is. Google builds its own AI chips with TPUs, Amazon with Trainium and Inferentia, and Meta with MTIA. Every one of these chips that goes into operation is an order that Nvidia doesn’t get. That’s exactly why stock prices react to such announcements.

A common misconception is that MTIA is a competing product on the open market. It is not: Meta doesn’t sell the chips and thus doesn’t act as a chip manufacturer. Experts refer to this as a “custom silicon” approach — that is, specialized hardware built for one’s own use. Success is not measured in sales figures, but in saved electricity costs and purchasing expenses.