Apple has spent more than a decade watching the AI data centre boom from the sidelines. Nvidia sells the GPUs, hyperscalers build the warehouses of compute, and Apple... sells phones. That might be about to change. According to a report from The Information, relayed by Ars Technica, Apple is developing a proper AI server powered by its own M8 Ultra chips — the company's first server product since it quietly discontinued the Xserve back in early 2011. Honestly, that detail alone made me sit up. We've all joked about Apple sitting out the AI arms race; turns out it was drawing up blueprints.
01.Apple's first server in fifteen years
The Xserve, for those who weren't around for it, was Apple's rack-mounted server that never really caught on. When Apple killed it in 2011, everyone assumed that was the end of Apple and the enterprise. The new project reportedly started roughly a year ago and had the backing of John Ternus — the hardware chief who, in a twist worth its own headline, became Apple's CEO this month. (If you're catching up on Apple's AI year, we covered the big Siri and iOS 27 AI plans in detail earlier.) So the person who greenlit the server is now running the company. That tells you how seriously Cupertino takes this.
02.What we know about the M8 Ultra server
Two configurations are on the table: one with two M8 Ultra chips and one with four. M8 Ultra would be Apple's most powerful planned silicon — though it hasn't been formally announced yet, so everything here comes with a "plans can change" disclaimer. The report says the machine would target AI developers, enterprises and government customers, with a launch not expected before 2029. Apple could still cancel it, and 2029 in chip-roadmap language usually means "don't hold your breath."
03.The Nvidia twist
Here's the part nobody's talking about enough: to link multiple chips inside a single server, Apple has discussed using Nvidia's NVLink Fusion interconnect — the technology Nvidia opened up to third-party chips at Computex Taipei in May 2025. Apple and Nvidia have had a famously prickly relationship for years, so the idea of an Apple server with Nvidia plumbing inside is genuinely wild. No decision has been made, and the server could ship without Nvidia technology at all. But the fact that the conversation is happening tells you how hard the multi-chip problem is.
04.AI labs already run fleets of Macs
What surprised me most in this story isn't the server — it's the reason it makes sense. AI companies, including OpenAI, have already purchased tens of thousands of Mac minis to train AI agents through reinforcement learning, where models learn by repeatedly attempting tasks. Anthropic rents Mac minis through Amazon Web Services. The Mac mini and Mac Studio have quietly become real workhorses for this kind of workload, thanks to their large unified memory and power efficiency. So when we broke down Microsoft and Nvidia's local AI PC push earlier this week, this is the same trend from the other direction — Apple wants a bigger slice of it, in rack form.
05.The memory problem
Every silver lining has a cloud, and this one is made of DRAM. The AI data centre boom has driven memory prices up sharply, and Micron executives have cautioned that meaningful new memory supply won't ramp until 2028. Memory costs have already pushed up prices across several Apple products. A server stuffed with multiple Ultra chips needs big memory pools, and whether those chips are available at sane prices could decide whether the 2029 launch happens at all.
06.Why this matters in India
Two reasons. First, if Apple enters the AI server market, the "compute cost" conversation changes — more competition in inference hardware eventually means cheaper AI everywhere, including for Indian startups renting GPU capacity by the hour. Second, and more immediately: the Mac fleet story is a reminder that how you run AI workloads matters as much as which model you pick. Picking the right hardware and cost setup is exactly where most beginners lose money without realising it — it's the reason we keep telling people to read the cost-optimization chapter in our GPT-6 Astra — Zero to Hero Guide before they scale anything beyond a hobby project.
Key Takeaways
- Apple's first server since the Xserve (2011) is in development, with two or four M8 Ultra chips
- Target launch: 2029 — and the project could still be cancelled
- Apple is in talks with Nvidia to use NVLink Fusion for chip-to-chip connectivity
- OpenAI already owns tens of thousands of Mac minis for agent training; Anthropic rents them via AWS
- Memory supply, not chip design, is the biggest risk to the timeline
Frequently Asked Questions
What is Apple's rumoured AI server?
According to a report from The Information relayed by Ars Technica, Apple is developing an AI server powered by its own M8 Ultra chips — its first server product since the Xserve was discontinued in early 2011. Two configurations are on the table, one with two M8 Ultra chips and one with four, aimed at AI developers, enterprises and government customers. The chips have not been formally announced, and a launch is not expected before 2029.
Will Apple's AI server use Nvidia technology?
Apple has discussed using Nvidia's NVLink Fusion interconnect — the technology Nvidia opened to third-party chips at Computex Taipei in May 2025 — to link multiple chips inside the server. No decision has been made, and the machine could ship without Nvidia technology at all.
Why are AI labs already running fleets of Macs?
AI companies, including OpenAI, have purchased tens of thousands of Mac minis to train AI agents through reinforcement learning, and Anthropic rents Mac minis through Amazon Web Services. The Mac mini and Mac Studio's large unified memory and power efficiency have made them real workhorses for these workloads.
What This Means for You
If you're a developer or an entrepreneur in India, don't wait for 2029 to act on this. The signal to act on is today's one: AI workloads are migrating toward efficient, specialised hardware, and the winners will be people who understand how to run AI cheaply on whatever machines they can already afford. If you're just getting started with AI tools, learning to prompt well, automate real work, and control costs is worth far more than any hardware roadmap — and our GPT-6 Astra — Zero to Hero Guide walks you through exactly that, step by step.