Meta plans to deploy its in-house Arke AI chip during the first half of 2027, aiming to reduce dependency on Nvidia GPUs for inference workloads.

Meta Platforms plans to begin deploying its next-generation in-house AI inference chips to data centers in the first half of 2027. Co-designed with Broadcom and manufactured by TSMC, the chips are intended to reduce reliance on Nvidia processors, lower costs, and improve energy efficiency.

Meta announced its in-house AI chip initiative in 2023 and is currently testing the third-generation product, MTIA 450, internally codenamed Arke. The next generation, MTIA 500 (codenamed Astrid), is expected to complete design work within about one month, with data center deployment anticipated by the end of 2027. Meta plans to deploy this chip more broadly than previous generations.

According to Yee Jiun Song, Meta's Vice President of Engineering, each generation accepts slightly higher technical risk in exchange for better performance—measured by performance gains per unit of energy and per unit of investment. The chips are co-designed by Meta and Broadcom, with TSMC as manufacturer. TSMC delivered the first batch of twelve chips on September 1. Actual performance deviated within 2 to 3 percent from prior simulation results, and Meta's engineering team ran the company's own AI models on the chips immediately. Initial testing revealed no design flaws, though several months of debugging and optimization remain necessary while foundry yield ramps up.

Meta's Superintelligence Lab is participating in chip tuning, providing the hardware team with insights into future AI models and their inference-stage requirements. Because Meta completed substantial engineering work in-house, its chips can run its own AI models more efficiently than existing Nvidia offerings, according to Song.

Strategically, Meta had previously planned a chip codenamed Olympus capable of handling both training and inference, but that project has been canceled to focus exclusively on inference. Song stated that at multi-gigawatt deployment scales, cost becomes critical; a chip handling both workloads at approximately 30 percent higher expense is "completely unacceptable" under large-scale deployment. Meta committed to deploying over one gigawatt of in-house chips within twelve months, expected to accelerate further thereafter, assuming no sharp downturn in AI market demand.

Despite this hardware shift, Meta will continue purchasing substantial volumes of GPUs from Nvidia and AMD to support its broader AI infrastructure needs.