Meta Platforms Plans to Deploy New In-House Arke Chip in First Half, Aiming to Reduce Reliance on Nvidia

TradingKey - Meta Platforms (META) plans to begin deploying its next-generation in-house AI chips to data centers in the first half of 2027, a move the company says will help save money and energy when running AI models while reducing reliance on Nvidia processors.

Meta reportedly announced its in-house AI chip plan as early as 2023 and is currently testing the third-generation product in the series, MTIA 450, internally codenamed Arke. Design work for the next-generation MTIA 500 (codenamed Astrid) is expected to be completed in about a month, with deployment to data centers anticipated by the end of 2027. Meta plans to deploy this chip more broadly than previous generations.

Yee Jiun Song, Meta's Vice President of Engineering, stated in an interview that each generation takes on slightly higher technical risk in exchange for better performance, including performance gains per unit of energy and per unit of investment. According to Song, the chips are co-designed by Meta and Broadcom and manufactured by TSMC. TSMC delivered the first batch of 12 new chips on September 1, with actual performance deviating within 2% to 3% from prior simulation results. The engineering team used these chips to run Meta's own models on the very same day. Initial testing revealed no design flaws, but several months of debugging and optimization are still required, while yield on the foundry side is ramping up.

Based on power consumption—a core metric in the data center industry—Meta committed to deploying over 1 gigawatt of its in-house chips within 12 months, after which it is "expected to accelerate further." Song also noted that this forecast assumes no sharp downturn in the AI market and related demand.

Meta's Superintelligence Lab is participating in chip tuning, providing the hardware team with insights into future AI models and their inference-stage requirements. Song said that because Meta completed a large amount of engineering work itself, its in-house chips can run its own AI models more efficiently than "anything Nvidia currently ships."

In terms of strategic direction, Meta had previously planned a chip codenamed Olympus capable of handling both training and inference tasks, but that project has been canceled to focus on inference chips. Song stated that when deployment scales to multi-gigawatt levels, cost becomes crucial; if a chip handling both training and inference is about 30% more expensive, it is "completely unacceptable" under large-scale deployment. Meta will still continue to purchase large quantities of GPUs from Nvidia and AMD.