Meta Plans 1GW Deployment of Next-Gen Custom MTIA AI Inference Chips

Meta plans to deploy its next-generation custom AI chip, MTIA 450 (codenamed Arke), in its data centers, scaling to over 1GW within 12 months to lower AI inference costs. Co-designed with Broadcom and manufactured by TSMC, initial samples of the chip successfully executed Meta's internal models alongside third-party models from DeepSeek and Alibaba. The massive deployment aims to significantly reduce operational energy and hardware spending as Meta scales its AI infrastructure, while lessening its reliance on costly Nvidia GPUs. By focusing purely on inference rather than combining training and inference, Meta avoids an estimated 30% penalty in chip costs at gigawatt scale. Meta canceled a dual-purpose chip project codenamed Olympus to concentrate fully on main-stream inference workloads utilizing High Bandwidth Memory (HBM). Meanwhile, the fourth-generation chip, MTIA 500 (codenamed Astrid), is completing its design phase and is scheduled for deployment around 2027.