Meta to Deploy Arke AI Chip in First Half of 2027, Astrid by Year-End

Meta, the parent company of Facebook, said it plans to begin deploying a new in-house artificial intelligence chip, codenamed Arke, across its data centres in the first half of 2027 in order to save money and energy when running AI models. The chip belongs to the MTIA 450 series and is the third generation of Meta's proprietary AI silicon. It is currently in testing. Meta first disclosed plans to develop its own AI chips in 2023.

TSMC delivered 12 of the new chips to Meta on 1 September 2026, with measured performance deviating from the company's simulation results by no more than 2% to 3%. On the first day of delivery, the technical team used the processors to run Meta's own models as well as models from DeepSeek and Alibaba Group. The tests showed no design flaws in the semiconductor, though the chip will still require months of extensive testing and tuning as the foundry gradually scales up production. Meta partnered with Broadcom on chip design, while TSMC handles manufacturing. Yee Jiun Song said that by energy consumption, a core metric for the data centre industry, the company has committed to deploying more than 1 gigawatt of the chips within 12 months.

The chips primarily use high-bandwidth memory and are positioned for general-purpose inference rather than the ultra-high-speed inference market, where response times are critical. Meta had previously planned a chip codenamed Olympus that could handle both the training and inference stages of AI models, originally slated for launch in 2028 or 2029, but the project was cancelled and the company shifted its focus to inference, partly for cost reasons. Yee Jiun Song said that when building computing capacity of several gigawatts, cost becomes paramount; if a chip supporting both training and inference costs about 30% more, that becomes unacceptable at large-scale deployment. Meta's AI division, Meta Superintelligence Labs, helps fine-tune the chip by anticipating new generations of AI models and their operating requirements.

After completing development of the Astrid chip, Meta will focus on improving speed and throughput, meaning the total volume of AI tasks a chip can handle, in subsequent iterations, with optical fibre technology also to be used to further boost performance. Meta said it plans to continue developing chips over the coming years. The in-house chip programme is intended to reduce reliance on Nvidia processors.