Apple weighs enterprise AI servers built around Ultra chips

Apple is considering a return to the enterprise server business with machines designed for artificial-intelligence workloads, according to reporting cited by eu.36kr.com. The proposed systems could use two or four M8 Ultra chips and may incorporate Nvidia’s NVLink technology to connect them.

The project has reportedly been under development for almost a year. John Ternus approved it at an early stage, when he led Apple’s hardware engineering group; he is now the company’s chief executive. Any commercial release remains distant, however. The machines are not expected before 2029, and the project could still be abandoned.

If Apple proceeds, it would be the company’s most significant move into enterprise servers since the Xserve line was discontinued in 2011.

Macs have found an unexpected role in AI

Apple’s interest is emerging as laboratories, cloud providers and software developers look for large numbers of Macs to support AI development. Mac minis and Mac Studios, originally designed as desktop computers, are increasingly being installed in racks and used as hosted macOS machines.

These systems can run agent software, gather behavioural data and provide the real Mac environment required for applications that interact with a graphical desktop. OpenAI is reported to have acquired tens of thousands of Apple computers, while Anthropic is said to rent substantial Mac capacity through Amazon Web Services. Specialist hosting companies have also created racks intended for large Mac deployments.

One reason for this demand is the growing importance of post-training and agent capabilities in AI development. Systems that can operate a computer and complete practical tasks often need access to a genuine macOS desktop and its applications. Apple’s licensing rules require macOS to run on Apple hardware, so conventional servers cannot simply substitute for Macs through ordinary virtual machines or unofficial installations.

For some workloads, buying or renting Mac minis directly can therefore be more economical than trying to reproduce the same environment on different equipment. The demand has also spread beyond laboratories. Following the rise of OpenClaw and other agent products, Mac mini availability has reportedly been stretched, with some customers facing waits of a month or more. The Mac division’s quarterly revenue increased 29% year on year, although higher memory costs also contributed to changes in product pricing.

Apple’s memory design is suited to inference

Apple’s hardware architecture gives the company another possible advantage in AI inference. Since the M1 generation, Apple Silicon has used unified memory, allowing the processor and graphics processor to share a common pool rather than relying on separate system memory and graphics memory.

Large-model inference depends on memory capacity and data-transfer speed as well as raw processing power. In conventional computers and servers, moving information between separate memory pools can add communication delays, particularly in workloads that require heavy data movement.

The latest Mac Studio with an M5 Ultra chip is described as supporting as much as 512GB of unified memory and bandwidth of 1.2TB per second. That capacity can be sufficient for holding model weights that might otherwise need to be distributed across several specialist GPUs. For large models running at relatively low concurrency, a high-memory Apple system could therefore offer a practical alternative.

A dedicated server would let Apple adapt this existing interest into hardware specifically designed for rack deployment and enterprise use. The company would not be starting from zero: current consumer machines are already attracting institutional customers, despite not being built for data centres.

The Xserve experience remains a warning

Apple’s possible return is far from certain because enterprise computing demands more than attractive hardware and strong performance. Server customers expect competitive pricing, customisation, compatibility with established management tools, long-term maintenance and extensive technical support. Sales, installation, operations and contractual service obligations are central to the business.

That model conflicts with Apple’s traditional focus on consumer products, where design, simplicity and the user experience are major selling points. Enterprise purchasing is divided among procurement teams, IT administrators and data-centre operators, each with different priorities. The person approving a server may never see it, while the people maintaining it may care more about power use, density and repairability than industrial design.

Apple’s earlier Xserve illustrates the difficulty. Introduced in 2002, it had an unusually polished aluminium enclosure and carefully organised front-panel hardware. Yet its premium pricing and the closed nature of Mac OS X Server limited its appeal. It did not become a broadly adopted enterprise platform, and Apple eventually ended the product as its strategy moved elsewhere.

The current AI market gives Apple a stronger commercial reason to reconsider servers than it had in the past. Demand for Mac-based infrastructure already exists, and unified memory could make Apple Silicon useful for particular inference workloads. Still, the company would need to build the sales, deployment and support structure that its consumer business does not require. For now, the reported 2029 target is a possibility rather than a commitment.