Apple is exploring the development of a dedicated enterprise server for AI inference, potentially utilizing its own M8 Ultra chips for a release no earlier than 2029.
Key facts
- •The proposed server would be Apple's first entry into the server market in nearly two decades.
- •A potential market release for the server is not expected before 2029.
- •The project was initiated about a year ago when John Ternus served as the head of hardware engineering.
- •Apple is evaluating Nvidia's NVLink Fusion to manage chip connectivity.
- •OpenAI and Anthropic are currently using Mac hardware for AI-related tasks.
Apple is reportedly working on an enterprise server project designed for AI inference tasks. The proposed hardware would feature either two or four of the company's future M8 Ultra chips. The initiative, which began roughly a year ago under the leadership of current CEO John Ternus, aims to target AI developers, businesses, and government entities.
By the numbers
Technical Specifications and Potential Partnerships
The server project is currently in development with two distinct configurations planned. To facilitate high-speed communication between the chips within data center environments, Apple is reportedly considering the use of Nvidia's NVLink Fusion technology. While the technology was originally designed for Nvidia's own hardware, it has since been opened to third-party developers. The project remains in the development phase and could potentially be canceled or proceed without the integration of Nvidia's technology.
Market Momentum and AI Demand
The development effort follows a surge in demand for Apple's existing hardware among AI researchers. Companies such as OpenAI have reportedly purchased tens of thousands of Mac mini and Mac Studio units for AI workloads, including training agents through reinforcement learning. Additionally, Anthropic has utilized Mac mini hardware via Amazon Web Services. Apple's Mac revenue saw a significant increase of nearly 29 percent in the most recent quarter, reaching $10.4 billion.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder, Ars Technica.
