From First Silicon to Data Centre Scale: How Meta Brings Its MTIA Chips to Life

Key highlights

• Meta's custom AI chip, MTIA, depends on close teamwork across silicon design, servers, racks, software, manufacturing and data centre operations.
• Preparation for a new chip begins many months before the first physical unit arrives.
• Custom silicon only reaches its full value when the entire system around it is built to support it.
• Lessons from every deployment are fed back into the design of the next generation.

Introduction

Meta has shared a behind-the-scenes look at how its engineering teams take the Meta Training and Inference Accelerator (MTIA), a custom chip built for AI workloads, from its first arrival in the lab to full deployment across data centres. With four new generations of MTIA planned by 2028, the article shows that building a successful AI chip is about far more than the chip itself.

Preparation starts long before the chip arrives

The moment a new chip first reaches the lab is an exciting milestone, but it is the result of months of groundwork. Meta's silicon validation team begins work during the design phase, building models, reviewing estimates and preparing tests in advance. Once the chip arrives, they examine how it behaves electrically and thermally, how efficiently it runs and how it performs under real AI workloads.

Because AI workloads can rise and fall sharply, the team plans early for sudden changes in power demand. Even small improvements at chip level can add up to significant gains when multiplied across an entire data centre.

Building the system around the chip

A custom chip cannot work in isolation. It relies on processors, memory, networking, cooling, diagnostics and support from manufacturing and data centre teams. Meta's hardware engineers often begin designing the surrounding server and rack environment six months to a year ahead of the chip's arrival, with an eye on future generations, including how racks connect, how liquid cooling is routed and how systems will be maintained.

Where everything comes together

The bring-up phase is where hardware, software, thermal, manufacturing and operations teams meet. It is often the most challenging stage, as new issues surface for the first time, but it is also where the fastest learning happens. Meta describes this work as building a strong sense of ownership and broadening the skills of the engineers involved.

Teamwork at scale

Meta emphasises that bringing custom silicon to scale relies on trust between teams at every stage, from early testing through to deployment. Insights gathered in the field flow directly into the next round of chip and system design, allowing each generation to improve on the last.