Microsoft's In-House AI Model Outperforms OpenAI Using Only a Quarter of the Computing Power
nashnova research
Microsoft AI chief Mustafa Suleyman told staff at an all-hands meeting that in-house foundation model MAI-Base-2 outperforms the OpenAI model behind GPT-5.6 Sol in pre-training benchmarks — using just one-quarter of the compute. The disclosure marks Microsoft's sharpest public signal yet that it is shifting from OpenAI's biggest backer to its most direct technical rival.
Where exactly does MAI-Base-2 beat OpenAI?
Suleyman said MAI-Base-2 scored higher than OpenAI's internal model, codenamed "Spud," on benchmarks measuring a model's ability to understand and predict information patterns — during pre-training (the first and most expensive phase, where a model learns from massive datasets).
This means → Microsoft is not claiming an edge on one narrow task. It is claiming a lead at the foundational-capability level.
Suleyman said the comparison drew on Microsoft's privileged access to OpenAI's model performance data.
What does one-quarter compute really mean?
MAI-Base-2 was trained on one-quarter of the compute used for the equivalent OpenAI model. In plain terms = for the same capability level, Microsoft spent a fraction of the electricity and chip-hours.
The gap with DeepSeek is even wider: Suleyman said MAI-Base-2 is 12× more compute-efficient than DeepSeek-V4-Pro-Base.
This reflects a broader shift in AI competition — from "who has more GPUs" to "who can train a stronger model with fewer resources." Efficiency is the new front line.
Why is Microsoft willing to publicly undercut OpenAI?
Suleyman directly benchmarked Microsoft's model against OpenAI's at an internal all-hands — a rare move given Microsoft has invested billions of dollars in OpenAI and holds a stake valued at potentially over $100 billion.
The two companies are deeply intertwined: Microsoft resells OpenAI models through Azure and uses them in its own Copilot products. This means → Suleyman's remarks are not casual. They signal that Microsoft's leadership is willing to absorb relationship friction.
The relationship has visibly frayed over the past year. In March, Microsoft reassigned Suleyman to a new superintelligence division focused on in-house model development.
Can the in-house model actually replace OpenAI?
MAI-Base-2 is Microsoft's strongest efficiency signal to date, but pre-training is only step one — turning a foundation model into a shipping product involves substantial additional engineering.
In plain terms = Microsoft has shown it can build an equally capable "rough blank" more cheaply. Whether that blank can replace the OpenAI models inside Copilot remains unproven.
That question is the crux of the relationship's future: if Microsoft's in-house models prove viable at the product level, the partnership's underlying logic changes fundamentally.
市场有风险，内容仅供研究参考，不构成投资建议。