SHELFCRITTER AI Digest

MODEL

MTIA 450

modeltopic-notemetaaccelerator

Overview

MTIA 450 (code-named “Arke”) is Meta‘s third-generation in-house AI accelerator, positioned for inference and serving workloads rather than training. Announced in September 2026 for H1 2027 deployment, MTIA 450 is fabbed by TSMC with Broadcom as design partner. First 12 chips were delivered on Sept 1, 2026. MTIA 450 is Meta’s cost/energy play against NVIDIA H200/B200 spend inside its own serving stack.

Timeline

  • 2026-09-17-AI-Digest — Meta confirmed MTIA 450 “Arke” is in testing and will deploy on H1 2027 racks, with the first 12 chips delivered by TSMC on Sept 1, 2026 and Broadcom as design partner. The load-bearing corrections are: (1) MTIA is inference-first, not training-parity — Meta is explicit that MTIA is for serving workloads, so nothing shifts on the training side of Meta’s stack; (2) MTIA 450 makes Meta the fourth hyperscaler (behind AWS Trainium 3, Microsoft Maia 200, and Google TPU v7 Ironwood, all in production ahead of MTIA 450’s H1 2027 window), not the second.

Key Developments

  1. Inference-first, not training-parity: Meta framed MTIA 450 as an inference / serving accelerator against a hypothetical NVIDIA H200/B200 spend — the training-side calculus is unchanged for Meta.
  2. Fourth-hyperscaler positioning: MTIA 450 lags AWS Trainium 3, Microsoft Maia 200, and Google TPU v7 Ironwood in deployed-production timing; “second hyperscaler with a credible custom-silicon path off Nvidia” framing that some coverage carried is incorrect.
  3. TSMC + Broadcom supply chain: TSMC handles fabrication (12 chips delivered Sept 1, 2026); Broadcom is co-design partner. Same partner-shape as MTIA 500 “Astrid” downstream.