Meta testing next‑gen custom AI inference chips, plans data‑center deployment starting 1H 2027

Meta is advancing in‑house AI silicon for inference, citing partner relationships with Broadcom and TSMC and a deployment timetable that could materially affect its AI cost curve if successful.

In this brief: 3 sections 2 min read

- Meta testing MTIA 450 'Arke' third‑generation processors; successor MTIA 500 'Astrid' expected to complete design soon.
- GuruFocus reports chips set to enter data centers in first half of 2027, with broader MTIA 500 deployment by end of 2027.
- Early Arke samples reportedly performed within 2–3% of Meta simulations and have run Meta models alongside third‑party models.
- Article states Meta is working with Broadcom on chip design and TSMC on production.
- Meta canceled a prior Olympus plan that combined training and inference; current emphasis is on inference processors to improve performance per watt and per dollar.
- Reported plan to deploy more than a gigawatt of chips over a 12‑month period, accelerating with AI demand.
- If custom silicon reduces inference energy and dollar costs at scale, it could materially improve margins for compute‑intensive AI services.
- Execution, partner coordination and production scale are cited as primary risks in realizing cost benefits.

Read full analysis on gurufocus.com

Useful?