The MTIA 450 chip, code-named Arke, will begin deployment across Meta data centers in the first half of 2027 to lower the energy and financial cost of running AI models. A second next-generation chip, the MTIA 500 code-named Astrid, will enter data centers by the end of 2027. Meta has committed to more than 1 gigawatt of custom silicon capacity over a 12 month period to provide better performance per watt than current Nvidia systems.
Meta partnered with Broadcom for design and TSMC for manufacturing, receiving the first 12 Arke units on Sept. 1 with testing results within 2% to 3% of simulations. The company canceled a separate chip project called Olympus after determining that a design for both training and inference would cost 30% more. Later generations of custom silicon will focus on fiber optic technologies to increase throughput and speed.
Key sources
- SOURCE@wallstengine“canceled a previously planned chip called Olympus, which was intended to handle both AI training and inference, after determining that such a design could cost roughly 30% more”x.com
- SUPPORT@financialjuice“deploy new in-house Arke chips in first half of 2027”x.com
- SUPPORT@financialjuice“roll out next-generation Astrid Chips at end of 2027”x.com
- SUPPORT@financialjuice“Chips will save money and energy compared with Nvidia”x.com
- SOURCEmarketbrief.now