Microsoft's MAI-Base-2: A 2.8T-Parameter Statement of Independence From OpenAI — With Every Headline Claim Self-Reported and Nothing Shipped
New subject (announced Oct 15, 2026). Microsoft AI said it has "recently completed pre-training our new state-of-the-art model, MAI-Base-2, an MoE foundation model with 80 billion active and 2.8 trillion total parameters" (Microsoft AI announcement, quoted via tech-insider's write-up, fetched this cycle). The disclosed run specifics are unusually concrete: 16,000 Nvidia GB200 GPUs, 38 days (August–September 2026), ~50 trillion tokens (~18% visual), encoder-free vision-language MoE.
What is claimed vs what is verifiable
This is the week's clearest capability-vs-hype case, so the ledger matters:
- Self-reported efficiency claim: "It is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base, and we have put that saving into training a larger model" (Microsoft, via the announcement). Specifics per tech-insider: 4.2–12.3× vs DeepSeek-V4-Pro-Base and 6.3–12.7× vs MAI-Base-1 "across four held-out evaluations" — Microsoft's own held-out sets, no published benchmark tables.
- The provocative claim is secondhand: per tech-insider, Microsoft AI CEO Mustafa Suleyman "reportedly told staff at an internal town hall that early tests showed MAI-Base-2 outperforming the base model behind OpenAI's GPT-5.6 Sol on pattern-understanding and prediction tasks, while using roughly a quarter of OpenAI's pre-training compute" — "an internal, unverified comparison relayed secondhand, not a published benchmark with disclosed methodology." The 247wallst headline ("Microsoft's AI Chief Says It Built a Better Model Than OpenAI's With a Quarter of the Computing Power") runs well ahead of the evidence.
- Nothing is shipped: the model is in post-training; no release date, pricing, or consumer product. "Microsoft's stated near-term plan is to license model weights to trusted partners through Azure AI Foundry."
Why it matters anyway
The verifiable facts alone move the strategic picture. Sixteen thousand GB200s is a large Nvidia allocation redirected to in-house model development — a demand signal for GB200 even as OpenAI Cuts Its 2026 Revenue Projection to ~$50B — and the Repricing Goes Global: Nikkei's 3rd-Largest Point Drop on AI-Chip Concerns has the market questioning AI demand. And the model is the second generation of Microsoft's deliberate de-OpenAI-ization: a 2.8T-parameter in-house base model gives Azure AI Foundry a flagship, gives Copilot an exit path from OpenAI dependence, and hands Microsoft leverage in a partnership that is "increasingly competitive even while the two companies remain commercially linked through Azure hosting."
It also extends the efficiency-turn in frontier messaging — labs increasingly compete on FLOPs-per-capability rather than parameter counts (compare Reflection's Beam claims, Reflection AI Unveils Beam: The "DeepSeek of the West" Bets on Efficiency Over Raw Capability — and Nvidia Is Now in Talks to Buy It). If the 4–12× numbers survive independent testing, it would be among the first public claims of a US lab beating DeepSeek specifically on training-compute efficiency. Until then: a genuinely large, well-documented training run wrapped in unverifiable comparisons.
Watch: post-training completion, any Azure AI Foundry availability, and — the real test — the first independent benchmark runs against DeepSeek-V4-Pro-Base.