Meta is preparing to deploy two new generations of internally developed AI chips in its data centres in 2027, as the company looks to reduce its reliance on Nvidia processors and lower the cost of running artificial intelligence workloads.
The chips, known as Arke and Astridare part of Meta’s broader Meta Training and Inference Accelerator (MTIA) programme.
Arke to Arrive in First Half of 2027
Meta is currently testing MTIA 450, internally known as Arkewhich represents the third generation of its custom AI chip family. The company plans to begin deploying Arke in its data centres during the first half of 2027.
The next-generation MTIA 500, known as Astridis expected to complete its design phase shortly and enter data centres towards the end of 2027. Meta plans to use Astrid on a significantly larger scale.
The company has already committed to deploying more than 1 gigawatt of computing capacity using the new chips over a 12-month period.
Designed to Cut AI Infrastructure Costs
Meta’s custom silicon strategy is primarily focused on improving performance per watt and performance per dollar. As the company builds increasingly large AI data centres, even relatively small improvements in chip efficiency can translate into substantial savings.
Meta’s custom-chip programme is also intended to give the company greater control over its AI infrastructure rather than depending entirely on commercially available processors from Nvidia and other suppliers.
Broadcom and TSMC Are Involved
Meta is working with Broadcom on the chip designs, while Taiwan Semiconductor Manufacturing Co. (TSMC) is manufacturing the processors.
Early testing of the latest chips has reportedly been encouraging. Twelve processors reached Meta from TSMC on September 1, with their performance reportedly within 2–3% of the company’s simulations. The chips were able to run Meta’s own AI models as well as models from DeepSeek and Alibaba.
Meta Focuses on AI Inference
The new processors are designed mainly for general-purpose AI inferencerather than extremely low-latency inference workloads.
Meta has also changed its chip strategy. The company previously planned a processor called Olympus that would handle both AI training and inference. That project was cancelled as Meta concluded that a dual-purpose chip could be significantly more expensive at the massive scale of computing capacity it is building.
Meta’s custom silicon programme, first announced in 2023, is therefore becoming an increasingly important part of its AI infrastructure strategy as computing demand and data-centre costs continue to rise.
Summary
Meta will deploy its in-house Arke and Astrid AI chips in 2027with Arke expected in the first half and Astrid towards the end of the year. Developed with Broadcom and manufactured by TSMC, the chips are designed to improve AI inference efficiency, reduce energy consumption and lower Meta’s dependence on Nvidia processors as its AI infrastructure expands.
Image Source