Meta details MTIA chip strategy and multi‑GW data centre buildout

Meta’s MTIA custom chips are central to its AI infrastructure; the vendor relationship with Broadcom and a multi‑GW cluster roadmap are highlighted as part of Meta’s compute advantage for running Llama and Muse models.

In this brief: 3 sections 2 min read

- Meta uses its Meta Training and Inference Accelerator (MTIA) family for large‑scale model training and inference.
- Article states hundreds of thousands of MTIA chips are deployed for Facebook, Instagram, WhatsApp and Llama models.
- Four additional MTIA generations are planned within two years; Broadcom is cited as a co‑developer.
- Prometheus described as a 1GW cluster in New Albany, Ohio, due online in 2026.
- Hyperion project in Richland Parish aims for up to 5GW; a joint venture with Blue Owl Capital covers about $27bn of development costs.
- Meta operates a global fleet of 33 datacentres supporting its apps and AI models.
- Reports water usage effectiveness and volumetric water restoration benefits for 2024.
- Notes Meta’s July venture with BlackRock to develop a 1GW campus in El Paso where BlackRock funds hold 80% and Meta 20%.
- States MTIA racks align with Open Compute Project standards.

Read full analysis on datacentremagazine.com

Useful?