MTIA 500

Overview

MTIA 500 (code-named “Astrid”) is Meta‘s fourth-generation in-house AI accelerator, the successor to MTIA 450 “Arke”, targeted for late-2027 racks. Meta stated in September 2026 that MTIA 500 is finishing design work within ~a month — a design-done milestone, not tape-out. Like the 450, MTIA 500 is scoped for inference / serving workloads rather than training.

Timeline

• 2026-09-17-AI-Digest — Meta said MTIA 500 “Astrid” is completing design work within ~a month for late-2027 racks. Bloomberg’s headline compressed this to a tape-out timing claim, but Meta’s phrasing is design-done, not taped out — tape-out is the point of no return on a silicon spin and comes later. The MTIA 500 sits behind MTIA 450 in Meta’s roadmap and shares the inference / serving positioning; nothing changes on Meta’s training stack.

Key Developments

1. Design-done ≠ tape-out: The Bloomberg compression flattened a milestone-earlier signal into “tape-out in about a month”. Design-done is a real milestone but does not commit the physical silicon spin.
2. Late-2027 racks, inference-first: MTIA 500 follows MTIA 450 in Meta’s serving stack; the training / open-weights side of Meta’s roadmap is unaffected.