Meta to start production of its new generation AI chips in September

01:59 / 10.07.2026·41·Technology

Meta, led by Mark Zuckerberg, is starting mass production of its new generation AI processors to reduce dependence on Nvidia chips and optimize costs. According to Reuters, citing internal company documents, the new chips are expected to hit the assembly line starting this September. This move provides a strategic advantage for the company amid the global GPU shortage. This is reported by Techcrunch.com reports .

It is reported that at least one of the new chips has successfully passed a six-week testing phase. Meta is collaborating with Broadcom on the design of this project, but the actual manufacturing process has been entrusted to Taiwan's TSMC, the world's largest semiconductor foundry. Additionally, industry giants such as Samsung (RAM), Sandisk (storage devices), and Sumitomo Electric (fiber optic equipment) are involved in supplying components.

MTIA program and technological approach

The new chips were developed under Meta's MTIA (Meta Training and Inference Accelerator) program, primarily serving to improve model training and algorithm efficiency. Considering the rapid changes in the field of AI, company engineers created the chips based on a modular architecture. This approach allows the technology to adapt quickly to new types of tasks without becoming obsolete.

Meta intends to use these chips to make recommendation algorithms on Facebook and Instagram, as well as advertising systems, even smarter. Although the company will continue to purchase products from Nvidia and AMD, having its own developments will help save billions of dollars in capital expenditures. For reference, Meta plans to spend between $125 billion and $145 billion on AI infrastructure this year.

Competition and global trend

Creating its own chips in the AI race is not limited to Meta. Today, major technology corporations are striving to break free from Nvidia's dominance:

OpenAI is working on creating its own processor in collaboration with Broadcom;
Google and Amazon are already using their own training and inference chips;
Anthropic is in negotiations with Samsung regarding chip manufacturing.

Meta is mobilizing massive computing power to train its new Muse Spark series AI models. The company aims to reach 7 GW this year and double that figure next year. Such massive projects demonstrate how important it is to achieve independence not only in software but also in hardware.