Microsoft says it has completed pretraining of 3-trillion-parameter MAI-BASE-2 model

The claim matters less as a capability statement than as a signal about vertical integration: a model built entirely in-house, rather than sourced or partnered, points to Microsoft treating frontier pretraining as internal infrastructure. That has downstream implications for the compute supply chain and for the RWA and crypto sectors that increasingly rent AI capacity rather than owning it. With no benchmarks, no release date and post-training results still pending, the open question is whether the in-house route translates into deployable, differentiated models or remains a scale claim awaiting validation.