Meta Plans In-House AI Chip Deployment in 2027

Meta Platforms plans to begin deploying its latest internally designed artificial intelligence chip in data centers during the first half of 2027, according to the Los Angeles Times.

The chip, known internally as MTIA 450 or Arke, is being developed to improve the cost and energy efficiency of running AI models.

Meta has committed to deploying more than 1 gigawatt of computing capacity using its custom chips during a 12-month period. The company expects adoption to accelerate after the initial rollout if performance and AI demand remain strong.

The project is part of Meta’s effort to reduce its dependence on Nvidia, whose graphics processors dominate the artificial intelligence data-center market.

Meta is working with Broadcom on chip design and Taiwan Semiconductor Manufacturing Company on production. Twelve Arke processors reportedly arrived from TSMC on September 1, and early testing produced results within approximately 2% to 3% of Meta’s simulations.

Engineers were able to run Meta’s own models and systems developed by DeepSeek and Alibaba on the chips. Additional testing and tuning will continue as manufacturing volume increases.

Arke is primarily designed for inference, the process of using a trained AI model to respond to requests. It is not intended to replace Nvidia’s most powerful chips for training frontier models.

Focusing on inference allows Meta to optimize the processors for the workloads used across Facebook, Instagram, WhatsApp and its consumer AI services. Even relatively small improvements in performance per watt can create substantial savings when applied across gigawatts of data-center capacity.

Meta is also developing a fourth-generation chip called MTIA 500 or Astrid. The company expects to complete the design soon and deploy the processor in data centers toward the end of 2027.

The company previously considered building a chip that could handle both training and inference. Meta canceled that project after determining that the broader design could cost approximately 30% more, making it less attractive for deployments at enormous scale.

Custom processors give Meta greater control over its infrastructure and allow chip development to be coordinated with future AI models. Engineers from Meta Superintelligence Labs are providing information about expected workloads so the hardware can be optimized for the company’s software.

The strategy does not mean Meta will stop purchasing Nvidia processors. The company continues to require leading-edge accelerators for model training and other demanding workloads.

Instead, Meta is attempting to reserve expensive general-purpose GPUs for tasks where they provide the greatest advantage while shifting predictable inference workloads to internally designed chips.

If the rollout succeeds, Meta could reduce both operating expenses and energy requirements as its AI infrastructure expands. The program also illustrates how the largest technology companies are increasingly designing their own silicon to control costs, secure supply and differentiate their computing platforms.