Meta launches new set of AI chips, but its approach is the big lesson

Meta launched a new set of custom AI chips designed for inference, but the biggest takeaway is the company's approach.

The company unveiled the MTIA 300, 400, 450, and 500. These are the latest custom chips that started two years ago with the launch of MTIA 100 and MTIA 200. The chips have been deployed or scheduled for deployment in 2026 and 2027. The workloads cover ranking and recommendation inference and training and genAI workloads and genAI inference.

Like all new processors, Meta's new family of chips make various improvements. The bigger picture is the approach. Meta said:

"AI models are evolving faster than traditional chip development cycles. Chip designs are based on projected workloads, but by the time the hardware reaches production — often two years later — those workloads may have shifted substantially."

Meta added that its approach is to be iterative, use modular chiplets, incorporate AI knowledge and deploy quickly. Meta is on a 6-month development cycle. And since it leverages the same racks and form factors, Meta can continuously refresh.

• Meta to double down on Nvidia's AI stack, including CPUs
• Meta CEO Zuckerberg sees AI returns ahead, new models on deck

The company is also focusing on inference first since those improvements tend to drive Meta's financials.

Meta said:

"At the system level, MTIA 400, 450, and 500 all utilize the same chassis, rack, and network infrastructure. Therefore, each new chip generation can be dropped into the same physical footprint, accelerating the transition from silicon to production deployment. Our modular, reusable designs also minimize the resources needed to develop and deploy multiple chip generations, and the benefits of these highly optimized chips can offset the resources used for development and deployment."

The company added that it is also focused on industry standard software and hardware including PyTorch and Open Compute Project (OCP) designs.

Simply put, the AI cycles are only accelerating and Meta is making sure it doesn't get boxed in.