Meta plans to deploy in-house AI chips Arke and Astrid in 2027 to reduce Nvidia reliance and cut inference costs.

Meta Platforms will deploy its next-generation Astrid and Arke AI inference chips in 2027 to reduce data centre costs associated with Nvidia's AI chips and consolidate its self-reliance. Arke will roll out by mid-2027, with Astrid deploying at the end of the year. The company stated that this move would save both money and energy.

This development follows Meta's announcement that it is manufacturing a specialized AI chip called Iris in September to double its computing capacity to 14 GW in the coming year. Iris is part of the company's four-generation project for Meta Training and Inference Accelerators (MTIA). Meta designed the chip in-house to build up its AI capabilities and power its Facebook and Instagram platforms.

The tech giant took six weeks to test the chip and reported no prominent issues, according to internal memos. Meta partnered with Broadcom, the US-based semiconductor maker, to design the chip and enlisted Taiwan Semiconductor Manufacturing Co. to manufacture it.

The move reflects broader industry efforts to reduce reliance on Nvidia. Google announced an AI chip named Frozen V2 on July 20, which it is developing in-house to boost the efficiency of its proprietary AI models. By building the blueprints directly into the chip, Google expects to reduce data movement and the number of decisions required during inference, thereby improving model efficiency. Google plans to deploy the chip by 2028 at the earliest.