Apple is building an AI server it swore off 15 years ago
Apple's most interesting AI product today is a server it hasn't announced, at a date it hasn't committed to. The rest of the day's news is a $20 billion transatlantic merger.
Apple is developing an enterprise AI server built around its own M8 Ultra chips, with an internal target of 2029, and has talked to Nvidia about wiring those chips together with NVLink Fusion. The Information reported it Wednesday; Reuters, Bloomberg and AppleInsider followed within hours, all riding the same leak and none confirming it with Apple or Nvidia. The plan comes in two configurations: a two-chip box and a four-chip box. That second one is the interesting one, because Apple has never shipped a machine where four of its own top-end processors had to behave like a single computer. Apple last sold dedicated server hardware as the Xserve, killed in 2011. It already builds servers in Houston for Private Cloud Compute, but those are internal infrastructure; this would be a product Apple sells to developers, businesses and governments.
Why it matters: Apple's AI strategy has leaned on a convenient fiction — that a Mac Studio in a rack is close enough to a server. That fiction is now capping the company. In its most recent quarter, Mac was Apple's fastest-growing product category, up nearly 29% to $10.4 billion, driven in part by AI developers buying Mac mini and Mac Studio machines until supply ran tight. Those buyers want memory bandwidth and rack density, not a desktop with a Thunderbolt ceiling. An M8 Ultra machine on NVLink-grade fabric is Apple conceding that the developer demand it created has outgrown the hardware it sells to serve it.
The Nvidia part is the real story. Two decades of Apple–Nvidia hostility trace back to defective Nvidia graphics chips that caused widespread MacBook failures and ended the relationship. Now Apple is weighing NVLink Fusion — the interconnect Nvidia opened to outside silicon — as the strongest option for tying M8 Ultra chips into modern data-center fabric. That would hand Nvidia a marquee customer for networking rather than compute, which is exactly the business Nvidia has spent two years expanding. Nvidia's Dion Harris declined to say whether Apple is a customer, telling AppleInsider only that "Apple has always been a great customer and partner of ours."
Our take: a 2029 target is an escape hatch, not a promise. The Information notes the project can be cancelled, and Reuters is explicit that it could ship without Nvidia technology. Apple has to decide whether it wants to sell infrastructure or just the halo around it — and the chips will answer before the marketing does.
Cohere and Aleph Alpha signed a definitive merger agreement, combining a Canadian model developer with Germany's erstwhile AI champion into a single company valued around $20 billion. The combined business keeps the Cohere name with dual headquarters in Toronto and Berlin, and Aleph Alpha co-chief Ilhan Scheer becomes chief operating officer. The pitch is sovereign AI: systems that run inside a customer's own infrastructure and satisfy local regulators, sold to businesses, governments and public administrations rather than consumers. Cohere brings Command models and $240 million in annual recurring revenue; Aleph Alpha brings procurement relationships in Germany and little revenue of its own.
The money behind it is retail. Lidl-owner Schwarz Group is investing €500 million directly and up to €13 billion in compute through its cloud arm StackIT, tied to a German data-center campus expected to hold as many as 100,000 AI chips. That is the shape of European AI consolidation: not a lab racing OpenAI, but a stack assembled by whoever already owns the customers and the electricity. The deal still needs regulatory approval, and the terms disclosed in April were already stale — neither company updated the valuation.
What to watch: whether European governments buy the sovereign pitch at scale, or keep defaulting to US hyperscalers with local regions.
Would you run your models on an Apple server if the price per token beat a rented Nvidia box? Tell us in the comments.
