Meta(META.US) plans to launch the fourth generation of self-developed AI chips before the end of 2027 to reduce dependence on suppliers like (NVDA.US).

Meta Platforms (META.US) announced on Wednesday that the company plans to launch four generations of its self-developed artificial intelligence chips by the end of 2027 to support its rapidly growing AI computing needs and reduce dependence on external chip suppliers. This plan is also an important initiative for Meta to drive hardware self-development and lower long-term costs in the fierce and costly AI race.

Meta stated that in the coming years, it will successively release four chips: MTIA300, MTIA400, MTIA450, and MTIA500, which belong to its “Meta Training and Inference Accelerator (MTIA)” series. These chips are primarily used to support the company’s internal AI training and inference tasks.

Among them, the MTIA300 has already entered mass production and is mainly used for model training in content ranking and recommendation systems; the MTIA400 (codenamed “Iris”) has completed laboratory testing and is gradually advancing toward deployment. More advanced MTIA450 and MTIA500 (codenamed “Arke” and “Astrid,” respectively) are expected to achieve large-scale deployment by 2027.

Meta’s Vice President of Engineering, Yee Jiun Song, stated that these products are being developed concurrently, with MTIA450 expected to launch in early 2027 and MTIA500 approximately six months later.

Song pointed out that the pace of AI development has far exceeded expectations, which has also forced chip R&D to accelerate. “Even just in the past two or three months, the speed of AI development has shocked many people. Chip R&D must keep up with these changes in workload, so we constantly review our roadmap to ensure we are developing the most valuable products.”

In recent years, Meta has invested heavily in the AI field to build competitive large models and AI products, which has led to unprecedented demand for computing power. Currently, Meta continues to purchase a significant amount of external chips, including AI accelerators from Nvidia (NVDA.US) and AMD (AMD.US). The company recently announced that the scale of AI hardware procurement agreements reached with both companies is in the hundreds of billions of dollars.

At the same time, Meta is also accelerating the construction of its self-developed chip capabilities. Last year, due to CEO Mark Zuckerberg’s dissatisfaction with internal R&D progress, the company attempted to acquire the South Korean AI chip startup FuriosaAI for $800 million, but the deal was rejected. Subsequently, Meta turned to acquire the chip startup Rivos Inc., headquartered in Santa Clara, California, and absorbed its more than 400 employees.

The newly added talent resources help Meta’s MTIA team to advance multiple chip projects simultaneously. This team is primarily focused on developing more efficient computing architectures to meet Meta’s internal needs, including Instagram content recommendation systems, ranking algorithms, and large-scale generative AI inference tasks.

Meta executives stated that self-developed chips can be optimized for specific applications within the company, thereby improving efficiency and reducing costs. “We are not designing chips for the general market, so there is no need for excessive general functionality,” Song stated. “Eliminating unnecessary features can significantly reduce costs.”

However, chip R&D itself is a high-cost and long-cycle endeavor. From design to production by a third-party foundry, typically TSMC (TSM.US), it often requires billions of dollars in investment and takes years. Song indicated that the Meta team typically needs about two years to advance a chip from design to mass production.

Meanwhile, Meta’s chip strategy still faces challenges. Reports previously indicated that Meta has canceled its most advanced AI training chip project, codenamed “Olympus,” due to high design difficulty, and the company has turned to developing a simplified version. Meta has not directly commented on this report, only stating that the company will continue to evaluate and adjust its chip roadmap.

Despite this, Meta remains committed to advancing the development of self-developed AI processors. The company’s Chief Financial Officer, Susan Li, recently stated at a Morgan Stanley-hosted conference that Meta still hopes to ultimately develop processors that can be used to train large AI models.