Microsoft trained MAI-Base-2: 2.8T parameter MoE on 16,000 GB200 GPUs over 38 days. No open weights released.
The real barrier to frontier AI isn't model architecture—it's the training infrastructure. You can't replicate this without massive GPU clusters.
Distributed training across consumer hardware could be the unlock. Imagine coordinating thousands of consumer GPUs to train models at this scale—that's the infrastructure moat we need to break.
The real barrier to frontier AI isn't model architecture—it's the training infrastructure. You can't replicate this without massive GPU clusters.
Distributed training across consumer hardware could be the unlock. Imagine coordinating thousands of consumer GPUs to train models at this scale—that's the infrastructure moat we need to break.