Apple studies returning to the server market with up to four M8 Ultra and NVIDIA technology

Apple is exploring its return to the server market with an enterprise system aimed at running AI inference loads. The team would use own M8 Ultra processors and could incorporate the NVLink Fusion technology from NVIDIA to communicate several chips within the same platform, according to information published by The Information.

The project remains in development and its commercialization is not expected before 2029. Apple could still cancel it, modify its architecture or dispense with NVIDIA technology. The company has also not publicly announced the product, so known configurations should be treated as internal plans subject to change.

Two versions with two or four M8 Ultra processors

Apple works with a configuration made up of two M8 Ultras and a larger one with four units. Ultra chips represent the higher performance variant of Apple Silicon and they employ unified memory, an architecture in which CPUs, GPUs, and other processing blocks access the same pool of memory.

The main purpose would be to execute inference, that is, use already trained AI models to process requests and generate results. The server would be aimed at artificial intelligence developers, companies and public organizations. This would differentiate it from the Mac Studio and Mac mini computers that some organizations currently bundle together for Apple Silicon-based processing power.

Previous information pointed to the possible use of future M7 Ultra in AI servers. The new report instead identifies the M8 Ultra as the basis of the commercial product planned for 2029. The difference may respond to an evolution of the project or to internal calendars that are still open, but Apple has not officially confirmed either of the two generations.

NVLink Fusion would connect Apple silicon

Apple would have held talks with NVIDIA to use NVLink Fusion as an interconnection system. This platform allows you to integrate processors and accelerators designed by other companies with NVIDIA’s network architecture and rack-scale systems. It includes communication technologies between chips, switches, and software components intended to coordinate multiple processors as a joint platform.

In the system studied by Apple, its function would be to provide high-speed communication between the M8 Ultra. Data transfer between processors is one of the factors that limit performance when spreading a model across multiple chips. A high-bandwidth, low-latency interconnect reduces the time each processor spends waiting for information from the others.

The presence of NVIDIA in the possible Apple server would initially be limited to interconnection and associated infrastructure. The main computing would continue to depend on the M8 Ultra. Discussions do not guarantee that Apple has definitively selected NVLink Fusion or that the finished product will support NVIDIA GPUs.

Mac demand for AI would have fueled the project

Apple’s interest would be related to the Demand for Mac mini and Mac Studio among AI developers. Large laboratories have purchased tens of thousands of these computers, while other companies use them for inference and video analysis services. However, Macs are still designed as desktop systems and lack some common server features, including remote management suited to large deployments.

Apple already manufactures its own servers for Private Cloud Computethe infrastructure that processes Apple Intelligence requests that cannot be resolved directly on an iPhone, iPad or Mac. These devices are intended for the internal functioning of the service and are not sold to companies. The M8 Ultra project would turn some of that experience into a product available to external customers.

The company currently uses internally developed interconnection technologies on much of that infrastructure. Reports claim that this solution is slow and expensive to grow on a large scale. Some partners reportedly requested access to Private Cloud Compute’s servers, but Apple reportedly rejected those requests.

Apple stopped selling Xserve on January 31, 2011 and since then it has not maintained its own line of business servers in rack format. The Mac Pro, Mac Studio, and Mac mini can perform certain server functions, but are marketed as computers and workstations.

The new project would recover a dedicated offering eighteen years later, although focused on AI inference and based on Apple Silicon. The format of the device, its memory, consumption, operating system, storage options or price are not known. The 2029 target is also not a confirmed release date. Until Apple announces the product or begins preparing its distribution, it will remain a plan that can change or be canceled.