It's expensive, but that's because 72 GB200s are faster than 144 networked M2 Ultras. You're tackling bigger workloads with better-standardized software, genuin

It's expensive, but that's because 72 GB200s are faster than 144 networked M2 Ultras. You're tackling bigger workloads with better-standardized software, genuine first-party support and proprietary integration that goes even further than Apple's memory coherency. Datacenter customers aren't drooling over the advantages of Apple Silicon because most of them are table-stakes at their scale; Apple has to go the extra mile to stand out, specifically for these customers. I won't deny, Nvidia charges the sort of margins that would make Apple blush. But they do that (and are successful) because they're scaling the workloads Apple won't. Apple turned their back on Nvidia, gave up funding Kronos' alternative to CUDA, and now wants a piece of the pie they've spent the past 15 years trying to avoid. It just doesn't seem like they're capable of winning here; their strategy is to dominate or die, and domination isn't an option.