Meta is preparing to release its own AI chip Iris: the company is accelerating the expansion of computing power

Meta plans to begin mass production of a specialized chip for artificial intelligence in September. The project should reduce dependence on external suppliers and ensure the growth of the company's computing power in the coming years. For the Russian market, this is an indicator of how global players are restructuring the infrastructure to accommodate the growing loads of AI.

Meta, the owner of Facebook and Instagram, is preparing to launch its own artificial intelligence chip, writes xrust. According to sources, production of the first generation of Iris accelerators will start in September. The project is part of the MTIA (Meta Training and Inference Accelerators) line and should provide the company with control over key infrastructure elements that were previously entirely dependent on the supply of Nvidia and AMD GPUs.

The Iris chip is optimized for the internal tasks of Meta — content ranking, recommendations and the work of generative models. Testing the prototypes took about six weeks, which was a noticeable improvement over the company's previous attempts to develop its own hardware business. The accelerators are designed jointly with Broadcom, and production is located at TSMC facilities in Taiwan.

Meta expects that the transition to its own solutions will reduce the cost of inference — executing requests from already trained models that form the main load in social networks. At the same time, the company continues to purchase GPUs from Nvidia and AMD, concluding multi-billion dollar contracts for training large models.

According to sources, Meta plans to increase its total computing power to 7 GW this year and to 14 GW in 2027. For comparison, one gigawatt corresponds to the energy consumption of about 800 thousand homes. In the Russian context, such figures demonstrate the scale of global investment: the total capacity of domestic data centers is much lower, although projects in Siberia and the Far East are gradually increasing the available resources for AI services.

Meta's infrastructure investments in 2024-2025 are estimated at tens of billions of dollars. Rising costs for equipment and memory are forming a trend that analysts call “chipflation”—an increase in the cost of components against the backdrop of increased demand from large technology companies.

Experts note that the transition to their own chips repeats the strategy of Amazon and Google, which have been developing the Graviton and TPU lines for several years. This approach allows you to optimize costs and speed up the implementation of new functions. Meta plans to update its line of accelerators every six months, which is significantly faster than the industry's standard annual cycle.

For users, this means increasing the speed of services and improving the quality of recommendations. In the Russian context, despite limited access to Meta products, the trend remains indicative: the development of in-house hardware solutions reduces vulnerability to external supplies. For domestic companies, this confirms the importance of projects to create energy-efficient data centers and local processors.

The launch of Iris intensifies competition in the AI ​​infrastructure market, where the ability to quickly scale computing at controlled costs is key. Global experience shows that vertical integration — from chips to user services — is becoming one of the main advantages in the technology race.

Links to sources: https://www.reuters.com , https://ai.meta.com/blog/

Xrust Meta is preparing to release its own AI chip Iris: the company is accelerating the expansion of computing power