Microsoft's core AI strategy revealed: Let OpenAI charge forward as cannon fodder, adopt a follower strategy to reap the benefits of technology

Image | Mustafa Suleyman (Source: MIT Technology Review)

In an exclusive interview with CNBC aired last Friday, Suleiman made it clear: "Given the huge capital investment required for these models , we adopt a very cautious follower strategy ( Second mover ), maintaining a 3-6 month technology gap with cutting-edge model builders such as OpenAI, and developing on the basis of their success, which is more cost-effective than direct competition." He added that this strategy not only reduces R&D costs, but also allows Microsoft to use the time difference to optimize specific customer use cases.

This strategy seems abnormal - as a core participant in the generative AI revolution, Microsoft has not yet launched its own cutting-edge basic model. Its core competitiveness is actually tied to OpenAI: Microsoft provides a large amount of Azure cloud computing resources in exchange for the right to integrate the GPT series of models into Copilot brand AI services. At present, Microsoft's Copilot ecosystem has deeply integrated GPT technology, covering the Windows operating system and various cloud services.

Technology dependence and economic accounts Suleiman's strategic logic is rooted in business reality: when OpenAI CEO Sam Altman takes the risk of developing cutting-edge models, Microsoft does not need to invest billions of dollars to bet on a technology route that may fail. But this dependence is not absolute. Microsoft is also developing a series of open source small language models code-named Phi, with parameter scale controlled at the single and double billion level (such as Phi-4 has 14 billion parameters). Compared with GPT-4.5, which requires millions of dollars in GPU clusters, the Phi series can run on a single high-end GPU, significantly reducing the cost of inference.

Although the Phi series lags behind OpenAI's flagship products in cutting-edge areas such as multimodal support and expert mixture architecture (MoE), its technical performance is still competitive. Actual tests show that this series of models has outstanding performance in parameter efficiency - although Phi-4 only has about 1/10 of the parameters of GPT-4, it can reach 80% of the latter's accuracy in specific tasks such as code generation.

Autonomy Roadmap Suleiman

The background of this statement is the "Stargate" supercomputing project announced by OpenAI last year, a new generation of AI infrastructure built in cooperation with Oracle and SoftBank, which broke Microsoft's position as OpenAI's exclusive cloud service provider. However, Microsoft is not the only technology giant to adopt a follower strategy:

• Amazon AWS provides massive computing power to Anthropic through the "Project Rainier" cluster, while secretly developing a proprietary model series code-named Nova (unlike Microsoft's Phi series, Nova is not open source)
• Alibaba’s Qwen team launched the QwQ 32B preview version in just two months after OpenAI released the “Thinking Reasoning” model o1 preview version (September 2025), and officially released it after three months of optimization.
• Chinese AI startup DeepSeek uses the proven technology route of inference models to focus on reducing the computing power requirements for model training/inference
• Meta recently released the first expert mixture model based on the Llama 4 series

Suleiman pointed out that the follower strategy enables Microsoft to focus on the construction of the AI ​​application ecosystem. Compared with the breakthrough of the model itself, how to effectively integrate the large language model into the enterprise system is the key challenge. To this end, Microsoft has built a complete technology stack:

1. Autogen framework : realizes the collaborative work of multiple AI agents and has been applied to Microsoft Teams intelligent conference system
2. KBLaM architecture : By expanding the model knowledge base through structured data, Phi-4 Medical Edition improves diagnostic accuracy by 23% while reducing computing power consumption by 40%.
3. VidTok Toolkit : an open source video tokenization system that can improve the efficiency of 1 minute video processing to real-time analysis level (latency < 200ms)
4. Edge computing optimization : Phi series supports running on devices such as Surface Pro, reducing energy consumption by 75% compared to cloud-based reasoning

Microsoft CEO Satya Nadella added in a separate interview: "AI has yet to find its killer app, the same as email and Excel combined. Our systems-level innovation is preparing for this moment."

The strategy reveals Microsoft's three-level understanding of the AI ​​industry:

1. There is diminishing marginal returns in basic model research and development, and the cost of performance improvement from GPT-4 to GPT-5 is increasing exponentially
2. The enterprise market is more concerned with total cost of ownership (TCO) rather than simply model parameter size
3. In the future, the focus of competition will shift from model performance to:

• Computing efficiency (computing output per dollar)
• System integration (reducing API latency by 32% can increase customer retention by 19%)
• Vertical field optimization (Phi-4 financial version has an F1 value of 0.91 in fraud detection)

Industry data shows that the adoption of a follower strategy has reduced the R&D costs of Microsoft's AI department by 38%, while increasing the speed of customer solution delivery by 41%. This "second-mover advantage" is reshaping the AI ​​industry landscape - when pioneers such as OpenAI take technical risks, Microsoft captures commercial value through system-level innovation.