The brief

Analyst view

• Atlas systems in production at Oracle Cloud with over 50 racks deployed.
• Asimov taping out end of 2026; production H2 2027. Supports 16+ trillion parameter models.

Technical approach

As reported

Architecture

Memory-first design

Uses commodity LPDDR5X for inference, prioritizing bandwidth.

System

Titan scales with chiplets

4-8 Asimov chips deliver 18.4TB memory for large model inference.

Software

OpenAI-compatible API

Standard endpoints with Hugging Face support.