Meta tests its first in-house AI training chip to wean off Nvidia

Meta is reportedly testing its first in-house AI training chip as it aims to cut massive data center costs and reduce its reliance on Nvidia.

According to Reuters, Meta has kicked off a small-scale deployment of the custom silicon inside its data centers. The milestone follows a successful first tape-out, meaning the final design has been manufactured into real, physical silicon for testing. If internal trials pan out, Meta plans to use the chip for live AI training workloads by 2026.

Moving from recommendations to LLMs

The training chip is part of the Meta Training and Inference Accelerator, or MTIA family. Unlike Nvidia's general-purpose GPUs, custom silicon is built specifically for targeted AI tasks, offering superior power efficiency and lower per-token compute costs.

Meta is not starting from absolute scratch. Last year, the company deployed its first MTIA inference chip to power recommendation algorithms across Facebook and Instagram feeds.

The game plan for the new training chip looks identical. Meta intends to start training its ranking and recommendation models on the new hardware first. If stability holds, the ultimate goal is to shift training for heavyweight generative AI tools like the Meta AI chatbot onto its own silicon.

A comeback after a multi-billion detour

Custom hardware is notoriously brutal. Meta originally tried to deploy an earlier in-house inference chip, but scrapped the project after a failed small-scale test deployment.

That misstep forced Mark Zuckerberg to pivot hard in 2022, ordering billions of dollars worth of Nvidia GPUs and turning Meta into one of Nvidia's most lucrative customers. With this new training accelerator, Meta is officially taking a second crack at taking control of its own silicon destiny.

Why it matters

Nvidia's real long-term rivals aren't just chipmakers like AMD; they are its own whale customers. Google has TPUs, Amazon has Trainium, Microsoft has Maia, and Meta is now stepping into the custom training ring. If hyperscalers shift even a fraction of their internal training runs to custom ASICs, Nvidia's pricing power takes a direct hit.

We think Meta's rollout strategy is smartly calibrated. Rather than betting its next flagship Llama foundation model on untested silicon, Meta is pointing the chip at feed recommendations first. It is high-volume, highly predictable, and directly tied to ad revenue.

Still, building custom hardware is only half the battle; the software and compiler stack is where giants often stumble. Meta has until 2026 to prove this chip is ready for prime time—otherwise, it will be writing multi-billion-dollar checks to Jensen Huang for a long time to come.