Superpower Daily

Superpower Daily

Your reading space

Sign in
Together AI Adds Three Model Rollout Paths; Canary Pauses on Me...OpenAI Releases GPT-6 Sol and Luna With API Prices Half Their P...Anthropic Releases Claude Opus 5.5, Cutting Prices From Opus 5Nscale’s Filing Leaves $103.4 Billion TCV Without a Disclosed P...Microsoft Takes Down EvilTokens, an AI Service Built for Email...PayPal Brings Its Checkout Service to Meta’s Muse AI AgentTogether AI Adds Three Model Rollout Paths; Canary Pauses on Me...OpenAI Releases GPT-6 Sol and Luna With API Prices Half Their P...Anthropic Releases Claude Opus 5.5, Cutting Prices From Opus 5Nscale’s Filing Leaves $103.4 Billion TCV Without a Disclosed P...Microsoft Takes Down EvilTokens, an AI Service Built for Email...PayPal Brings Its Checkout Service to Meta’s Muse AI Agent
Swipe

Apple Reportedly Considers AI Server for a 2029 Return to the Server Market

The unapproved project would package Apple silicon for private AI inference, but the product and its possible Nvidia networking link remain unsettled.

By 2 min read
Apple Reportedly Considers AI Server for a 2029 Return to the Server Market
Apple Reportedly Considers AI Server for a 2029 Return to the Server Market

Listen to this story

The audio brief

About 1:31
0:001:31
Read transcript
Apple is reportedly considering an enterprise AI server built around two or four future M8 Ultra chips, with a launch no earlier than 2029. The machine would be designed for organizations that want to run already-trained models on their own hardware, rather than sending every inference request to the cloud. But this is still an unapproved project, not a product announcement, and Apple could cancel it. The central engineering question is how those chips would communicate. Apple may be evaluating Nvidia’s NVLink Fusion instead of relying on Thunderbolt 5, the connection used in its current Mac ecosystem. Thunderbolt 5 can transfer up to 10 gigabytes per second in both directions. Nvidia says its sixth-generation NVLink can connect as many as 72 accelerators, with up to 3.6 terabytes per second for each one. Those are published specifications, not performance results from an Apple server that exists today. The idea comes as OpenAI has reportedly bought tens of thousands of Mac minis and Mac Studios for reinforcement-learning work on AI agents, while Anthropic has rented Mac minis through Amazon Web Services. That demand offers context, but it is not evidence that Apple will ship a server. If it does, the system would mark Apple’s return to purpose-built servers after Xserve was retired in 2011. The key constraint is still unresolved: whether Nvidia participates, whether Apple can build the interconnect itself, or whether the whole project disappears before 2029.

Story brief

3 key points

Apple is exploring a possible enterprise inference server built around two or four future M8 Ultra chips, with launch timing no earlier than 2029. The technical decision may hinge on interconnect bandwidth: Apple is reportedly evaluating Nvidia’s NVLink Fusion rather than relying on Thunderbolt 5. The project remains unapproved, the Nvidia relationship is unfinished, and cancellation remains possible. If it ships,...

  1. 01

    The proposed system would run trained AI models for organizations operating inference infrastructure on-premises.

  2. 02

    Apple is considering configurations with two or four M8 Ultra chips; launch is not expected before 2029.

  3. 03

    Thunderbolt 5 offers up to 10GB/s bidirectionally, while Nvidia cites 3.6TB/s per accelerator for sixth-generation NVLink.

Apple is reportedly considering an enterprise AI server with two or four future M8 Ultra chips. It would target organizations running trained AI models on their own hardware and could launch no earlier than 2029. If released, it would be Apple’s first purpose-built server since the company retired Xserve in 2011.

This is a reported plan, not a product announcement. The server has not been approved, its potential Nvidia arrangement is unfinished, and Apple could cancel the project or pursue it without Nvidia technology.

The connection is the hard part

The reported server is intended for inference, meaning it would run already-trained models. Apple is considering Nvidia’s NVLink Fusion to connect the chips, because Thunderbolt-based Mac clustering may not offer enough bandwidth for a data-center system.

The issue is not just fitting more chips into one box. They also need to exchange data quickly. TechRadar says Apple’s Thunderbolt 5 can transfer up to 10GB per second in both directions. Nvidia says its sixth-generation NVLink, available through NVLink Fusion, can connect up to 72 accelerators at 3.6TB per second each. These are product specifications, not a test of an Apple server that does not yet exist.

Demand for Macs set the context

The Information reported that OpenAI bought tens of thousands of Mac minis and Mac Studios for reinforcement-learning work on AI agents. It also said Anthropic rented Mac minis through Amazon Web Services. That demand helps explain the possible server, but does not show that Apple will release it.

A possible return after Xserve

After retiring Xserve, Apple pointed customers toward server configurations of the Mac mini and Mac Pro rather than dedicated rack hardware. The newer effort reportedly began about a year before the September reports and had support from John Ternus while he led Apple’s hardware engineering organization.

Sources

  1. techradar.comApple weighs its first server since the Xserve, and as always, Nvidia may hold the key
  2. arstechnica.comApple reportedly building server packed with M-series Ultra chips for AI
  3. the-decoder.comApple is reportedly building an enterprise AI server with its own M8 Ultra chips

Loading discussion...

YOUR READING SPACE

Notifications