Meta Plans More Than 1 GW of In-House AI Chip Deployments Over 12 Months
Редакция ЦОД:Главное
Meta Platforms plans to expand the use of internally developed AI processors in its data centers, targeting lower model energy consumption and operating costs. The company says it intends to deploy more than 1 GW of such chips over a 12-month period, while MTIA 450 (Arke) remains in testing and MTIA 500 (Astrid) is targeted for data-center deployment by the end of 2027.
- Meta plans to expand internally developed AI processors across its data-center fleet to reduce model energy use and operating costs.
- The company has stated an intention to deploy more than 1 GW of these chips over a 12-month period; this is a planned deployment target, not operational capacity.
- MTIA 450, codenamed Arke, is in testing, while MTIA 500, codenamed Astrid, is planned for data-center deployment by the end of 2027.
- The program could increase demand for specialized AI compute, power delivery and cooling infrastructure while changing Meta’s accelerator sourcing mix.
Meta Platforms is expanding its in-house AI accelerator program for data-center deployment, with the stated objective of reducing AI-model energy consumption and operating costs.
The company is testing its third-generation MTIA 450 processor, codenamed Arke. Its next-generation MTIA 500 processor, codenamed Astrid, is expected to complete design work within a month, with deployment in Meta data centers planned by the end of 2027.
Meta has also stated an intention to deploy these chips with aggregate power exceeding 1 GW within 12 months. This is a deployment target rather than commissioned capacity. The source does not specify the starting date for that 12-month period or the methodology used to define aggregate chip power.
According to a statement attributed to Meta senior vice president of engineering Yee Jiun Song, TSMC delivered 12 Arke chips on September 1; the year was not specified. Song said observed Arke performance was within 2% to 3% of simulation results. The chips were used to run models from Meta, DeepSeek and Alibaba.
For data-center operators and infrastructure suppliers, the stated scale points to potential demand for specialized compute platforms and the supporting power and cooling systems required for high-density AI deployments. If Meta executes the plan, it could become a substantial user of proprietary AI infrastructure and partially shift demand away from externally sourced general-purpose accelerators.
The reported schedule is not fully consistent at the product level. The source separately refers to deployment of Meta’s next-generation AI chips in the first half of 2027 and to MTIA 500 deployment by the end of 2027, without clarifying whether both timelines apply to the same processor.
A planned deployment exceeding 1 GW would represent potentially significant demand for specialized AI compute and associated data-center power infrastructure. If executed, Meta’s shift toward proprietary accelerators could alter its accelerator procurement mix, while increasing the importance of compute energy efficiency, power delivery and cooling in AI infrastructure design.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.