Meta unveils plans for batch of in-house AI chips

SAN FRANCISCO (Reuters) - Meta Platforms on Wednesday unveiled a roadmap of four new chips that the company is making in-house, as it rapidly expands its data centers.

Like many big tech companies such as Alphabet and Microsoft, Meta has invested heavily in building a team that can design chips in-house in addition to purchasing off-the-shelf products made by Nvidia and Advanced Micro Devices.

The new chips are part of the company's Meta Training and Inference Accelerator (MTIA) program and the first of the new chips called the MTIA 300 is in use powering the company's ranking and recommendation systems. The other three will be rolled out this year and in 2027, with the final two chips, the MTIA 450 and 500 being designed to perform inference, the process when an AI model such as the one that powers the ChatGPT app responds to customer queries and requests.

"We see inference demand exploding at the moment and that's what we're currently focused on," Yee Jiun Song, Meta's vice president of engineering, said in an interview.

Meta has had some success with inference chips but has struggled with its long-time ambitions to make a generative AI training chip, capable of building the large models that power AI apps.

Beginning with the MTIA 400, which the company says is on the path to being used in its data centers, Meta has designed an entire system around the chips, which is roughly the size of several server racks and includes a version of liquid cooling.

The company plans to release the new chips at six-month intervals because it is rapidly expanding the number of data centers it uses to run apps like Instagram and Facebook, Song said.

"That is the reality of how quickly our infrastructure is being built out," Song said.

The company said in January it expects capital spending of between $115 billion and $135 billion this year.