News for Agent

Daily business news: funding, M&A, launches

Research & Clinical Trials

Microsoft Completes Pre-Training of MAI-Base-2, Its Second In-House Foundation Model

By ·

Key takeaways

  • Microsoft completed pre-training of MAI-Base-2, the second foundation model from its in-house AI group led by Mustafa Suleyman.
  • Predecessor MAI-Base-1 was pre-trained on 30 trillion tokens and 8,192 NVIDIA GB200 GPUs, then went through another 3.55 trillion tokens in mid-training.
  • Microsoft plans to expand its Maia 200 and 300 chips toward multi-gigawatt deployments starting in 2027.

Microsoft has finished pre-training MAI-Base-2, the second foundation model from its in-house AI group, the company said. Pre-training is the most compute-intensive phase of building a large language model, and the step moves Microsoft's homegrown AI effort into a second generation, Crypto Briefing reported.

Inside the MAI-Base-2 Pre-Training Milestone

Pre-training produces the raw intelligence of a model rather than a finished product. The model reads a large body of text and learns to predict what comes next, picking up grammar, facts and patterns. Instruction-following and reasoning come in later phases, typically called mid-training and post-training.

The name points to where the model sits. MAI-Base-1 is the base model underneath MAI-Thinking-1, the flagship reasoning model Microsoft AI introduced earlier this year. MAI-Base-1 was pre-trained on 30 trillion tokens of clean, human-generated data, then went through another 3.55 trillion tokens during mid-training. Microsoft trained that first base model on 8,192 NVIDIA GB200 GPUs.

The MAI program went public on June 2, 2026, at Microsoft Build, when Microsoft AI, the in-house team led by Mustafa Suleyman, rolled out seven models at once. Microsoft said every model, including MAI-Thinking-1, was built from the ground up without using third-party models for distillation. MAI-Thinking-1 has posted around 52.8% on SWE-Bench Pro and 97% on AIME 2025. Microsoft runs MAI models across its own workloads, including Copilot.

Microsoft's AI Revenue and Workforce at a Glance

Microsoft's AI business surpassed a $37 billion annual revenue run rate with 123% year-over-year growth, driven by Azure AI services and Microsoft 365 Copilot adoption, according to akta.pro company records. The company reported quarterly revenue of $82.9 billion in Q2 FY2026, up 18% year-over-year, with full fiscal year revenues exceeding $240 billion. Microsoft employs over 220,000 people globally.

Earlier Microsoft AI Moves This Year

On June 2, 2026, Microsoft rolled out seven models at Microsoft Build, including MAI-Thinking-1, and said they were built without third-party distillation, Crypto Briefing reported. In June 2026, Italy's AGCM opened an investigation into Microsoft over alleged unfair commercial practices tied to Microsoft 365 subscription price increases, claiming Copilot and Designer were bundled without adequate disclosure. Also in June 2026, Chevron's Energy Forge One subsidiary signed a 20-year power agreement for Project Kilby, co-locating approximately 2.67 gigawatts of generation capacity with a Microsoft datacenter campus in Reeves County, West Texas.

What Comes After MAI-Base-2

Microsoft plans to expand infrastructure for its own Maia 200 and 300 chips, moving toward multi-gigawatt Maia deployments starting in 2027. The company has not disclosed the data and compute behind MAI-Base-2.

Companies in this story

Sources

Machine-readable: Markdown · JSON