Apple Weighs a Return to the Server Business
Apple is reportedly developing an AI server built around the same high-performance M-series Ultra silicon that powers its Mac desktops, a move that would place the company back in the enterprise server market for the first time in roughly twenty years. According to reporting from The Information, the machine is expected to reach the market in 2029, making it the first Apple server to ship in nearly two decades.
The project is said to have gotten backing from Apple's new chief executive, John Ternus, when work began roughly a year ago — a period when Ternus was still leading Apple's hardware engineering organization. That detail matters, because it suggests the effort grew out of the chip and hardware groups rather than a software or services initiative. Ternus officially replaced Tim Cook as CEO on September 1, following Cook's 15 years in the role, so the server project could become an early defining item of his tenure.
Apple has not confirmed the plan, and the reporting notes that the effort could still be shelved before any product reaches customers.
Two Configurations Built on M8 Ultra Silicon
As described, the enterprise server would ship in two different configurations:
- A dual-chip version using two of Apple's future M8 Ultra processors.
- A four-chip version using four of the same M8 Ultra processors.
Both variants would lean on Apple's established strategy of scaling performance through Ultra-tier chips, which fuse multiple high-end cores into a single package for demanding workloads such as video production, 3D rendering, and increasingly, local AI inference and training. Stretching that architecture across a rack-mounted enterprise system would be a significant step beyond anything Apple currently sells.
The report also indicates that Apple is exploring whether to link its M8 chips together using data center networking hardware from Nvidia. Specifically, Nvidia's NVLink Fusion technology has been raised as a possibility, with sources telling The Information that the two companies have held discussions about how Apple might draw on Nvidia technology for its AI ambitions.
Those conversations are not guaranteed to produce anything concrete. The Information cautioned that Apple's server project could be canceled outright, or could proceed without any Nvidia technology inside it.
Why AI Developers Keep Buying Macs
The timing of the reported server project lines up with an unexpected surge in demand for Apple's compact desktop computers. Mac mini and Mac Studio systems have become popular choices among AI developers and companies looking to run AI workloads without building out conventional GPU clusters.
According to The Information, AI companies including OpenAI have purchased tens of thousands of Mac minis and Mac Studios to train AI agents through trial-and-error reinforcement learning. Anthropic has likewise rented Mac minis from Amazon Web Services rather than buying the hardware outright.
That pattern of adoption is unusual for Apple, whose computers have historically been associated with creative and productivity work rather than data center operations. The appeal appears to rest on the combination of unified memory architecture, strong performance per watt, and a form factor that can be stacked in large numbers without the power and cooling demands of traditional accelerator servers.
If developers are already comfortable deploying fleets of Macs for AI work, Apple may see a natural opening to sell them a purpose-built system instead. That would let the company capture more of the value from a market it is currently serving almost by accident.
A Long Road Back From the Xserve Era
Apple's last serious enterprise server was the Xserve, a rack-mounted line that the company retired in January 2011. For years afterward, Apple's official guidance for organizations needing server functionality was to use Mac minis or Mac Pros running server software — a workaround rather than a dedicated product family.
A return would therefore be more than a product launch; it would signal a renewed willingness to sell directly into data centers, a market where Apple has little current presence and where customers expect long support cycles, enterprise procurement terms, and deep integration with existing infrastructure.
Partnering with Nvidia on interconnect technology, if that happens, would be a pragmatic way to bridge the gap. Nvidia's networking fabric is widely used in AI clusters, and adopting it would make an Apple server easier to slot into environments already designed around Nvidia hardware.
The Stakes and the Open Questions
Several important details remain unresolved. Pricing, power specifications, memory capacity, software support, and how the system would be managed at scale have not been reported. It is also unclear whether Apple would sell the server directly to enterprises, offer it through cloud partners, or position it primarily as internal infrastructure to support its own AI services.
The competitive landscape is equally uncertain. The enterprise AI server market is dominated by systems built around Nvidia accelerators and increasingly by custom silicon from major cloud providers. Apple would be entering as a challenger with a different architectural philosophy, one that emphasizes efficiency and memory bandwidth rather than raw floating-point throughput.
There is also the question of commitment. Apple has canceled or quietly abandoned ambitious hardware projects before, and a 2029 target leaves ample time for priorities to shift. The reported support from Ternus at the project's inception is one indicator of internal seriousness, but it is not a guarantee of a shipping product.
What the reporting does make clear is that Apple's M-series Ultra chips have found a real, if unplanned, audience among AI researchers. Whether that grassroots demand translates into a formal enterprise server will depend on decisions Apple has yet to make — and on whether the company wants to compete in a market it left behind more than a decade ago.
This article is based on reporting by Ars Technica. Read the original article.
Originally published on arstechnica.com
