Microsoft AI's MAI-Base-2 finishes pre-training, matching DeepSeek V4 Pro with 12x less compute

Microsoft says scaling-law predictions held precisely for MAI-Base-2. The team used a ladder of smaller runs to forecast the final loss before scaling up to almost 100x more training compute. The model also trained smoothly, without the loss spikes that hit MAI-Base-1.

The announcement pairs with Microsoft's silicon push. Nadella said Maia will scale to multi-gigawatt deployments, and Maia 300 claims 40% more tokens per watt than leading GPU systems.

It also strengthens Microsoft's hedge against its dependence on OpenAI.