Consider the cost of GPUs, losing what could be double digit percents on overhead might not make this very competitive. The macOS microkernel can still beat NT in some situations (like not having filters slowing filesystem down to a crawl), but it lags significantly behind the investment in Linux performance over the years by every other major company.
The idea being there's no VMware, kernel or piece of hardware that can have backdoors built into unless someone files off the top of the chip and somehow probes the silicon
> Apple said that it, together with Foxconn Technology Group, will later this year begin producing the servers that power the cloud component of Apple Intelligence — a system called Private Cloud Compute — in Houston. That marks a relocation, at least for some production, from overseas. Next year, it says a 250,000-square-foot facility for such manufacturing will open in the city.
> The Private Cloud Compute servers use advanced M-series chips already found in the company’s Mac computers. Those chips themselves, however, continue to be produced in Taiwan.
Doing a Mac Mini sized version for SOHO would be great, too.
One can dream, I guess...
You can't easily sell "Good / Better / Best" version of a single model, and tell "These are the options, take it or leave it". Servers are customized to the screws they come with and are expanded throughout over the years. So, the logistics are somewhat different for these kinds of devices.
Plus, macOS is not a CLI first operating system for server operations, and macOS Server is not updated for some years. Allowing Linux would be a different offering, and allowing macOS to work with all kinds of hardware from ordinary Ethernet to 100G+ Ethernet and 400gbps Infiniband (plus all the other interconnects) will be a fun exercise in testing flexibility of both macOS and Apple development teams.
So it's quite complicated. All servers are built to order SKUs. Dell keeps configurations "per server" in their databases, for example. If you have a Dell server, enter its service tag to support site, and you'll get the configuration of the device as it left the factory.
I mean places like Github and AWS are painfully racking up Mac Minis for their deployments and this theoretical server model would simplify everything. It also becomes an option for on premises AI inference using MLX, especially if they manage to get ANE support working in conjunction with the GPU for faster prefill.
The support and software stack for the server model would be the exact same as the consumer variant and they certainly wouldn't have special Linux offerings, Infiniband, and all that. If there's networking beyond their existing 10G it's going to be built into the board and they aren't going to support random 3rd party cards. The unit also doesn't need to be upgradable either.