Meta Started Testing Its AI Training Chip

Meta Platforms started testing its first in-house AI training chip. The social media company thus continues its efforts to reduce reliance on more costly semiconductors from Nvidia and other vendors.

The company is currently trialing the chip in a small deployment ahead of plans to increase production if the results are promising. Meta’s new training chip is a dedicated accelerator that only handles AI-specific tasks, which will make it more efficient compared to GPUs that are generally used for AI workloads. It stated it can achieve greater efficiency by controlling its own stack and using its domain-specific silicon than it can with outside GPUs. Meta is developing the chip with TSMC.

In April 2024, Meta announced the second generation of its Meta Training and Inference Accelerator (MTIA) AI chip to support generative AI workloads. The chip also performs inference for recommendation systems across Facebook and Instagram news feeds. On a 4Q24 earnings call, company executives explained that the MTIA chip would be optimized to run across its network elements and cut costs.

“We expect to further ramp adoption of MTIA for these use cases throughout 2025 before extending our custom silicon efforts to training workloads for ranking and recommendations next year,” CFO Susan Li explained on the earnings call. She noted Meta hopes to expand MTIA to support some of its core AI training workloads and some of its generative AI use cases in 2026. In January, CEO Mark Zuckerberg said that planned capex is $60 billion to $65 billion this year as part of a scheme to broaden its AI infrastructure.