Meta Platforms plans to start deploying a new generation of internally developed artificial intelligence chips in its data centres during the first half of next year, as the company seeks to lower the cost and energy consumption of running AI models.

The Facebook and Instagram parent first announced plans to build its own AI processors in 2023. It is now testing the third generation of its custom chip family, known as MTIA 450 or Arke.

A subsequent generation, called MTIA 500 or Astrid, is expected to complete its design phase in about a month and enter Meta's data centres towards the end of 2027. The company expects to deploy Astrid on a significantly larger scale.

“Each one takes on a little bit more risk technologically, and gets us better performance,” Yee Jiun Song, Meta’s vice president of engineering, said in an interview.

The approach could help Meta reduce the cost of operating AI systems as demand for computing power continues to rise.

That includes better performance per watt of energy and per dollar spent, he added.

The company is working with Broadcom on chip designs, while Taiwan Semiconductor Manufacturing Co. (TSMC) is responsible for manufacturing the processors.

The company has already committed to more than one gigawatt of the new chips over a 12-month period when measured in terms of energy consumption, underscoring the scale of its planned deployment.

After that, “we expect it to accelerate,” said Song, who oversees Meta’s custom silicon programme.

Song cautioned that the outlook assumes there is no crash in AI markets or a significant decline in demand.

Meta is therefore building its own processors at a time when the company is making enormous investments in AI infrastructure and competing for access to computing resources.

The team provides engineers with information about upcoming AI models and the computing requirements needed to run them, particularly during the inference stage, when trained AI models generate responses or predictions.

Song said this close coordination could allow Meta's processors to deliver more efficient performance than Nvidia's latest offerings for specific AI workloads.