Microsoft and OpenAI have been close partners for years, with Microsoft investing billions into OpenAI and integrating its AI models into products like Copilot. But things are shifting. Microsoft is now developing its own in-house AI, called “MAAI,” possibly to cut down on OpenAI’s steep costs—especially with OpenAI’s new AI agents rumored to cost up to $20,000 per month.
In this article, we’ll break down:
- Why Microsoft is building its own AI
- The skyrocketing costs of AI agents
- The game-changing Belief State Transformer (BST)
The Growing Rift Between Microsoft and OpenAI
Microsoft initially invested around $13–14 billion into OpenAI, making it the exclusive cloud provider for OpenAI’s workloads. But now, reports suggest that Microsoft is training its own large language models (LLMs), known as MAAI, to potentially replace OpenAI’s technology in Microsoft products like Copilot.
Currently, Copilot relies on OpenAI’s models for tasks like writing, coding, and spreadsheet management. However, Microsoft seems eager to break free from its reliance on OpenAI. Some sources claim that MAAI performs just as well as OpenAI’s models, prompting Microsoft to explore integrating its own AI into its products. The move could save Microsoft a fortune in AI costs while giving them more control over their technology.
AI Is Getting Expensive—Really Expensive
Training and running AI models isn’t cheap. The specialized AI chips needed to power large language models can cost anywhere from $10,000 to $30,000 per chip. And entire server racks for AI operations can exceed half a million dollars—before even considering power and maintenance costs.
OpenAI is rumored to be rolling out AI agents with prices as high as $220,000 per month for top-tier research models. Some companies, like SoftBank, are reportedly spending billions to access OpenAI’s cutting-edge agents. But not everyone is convinced these prices are justified. While some startups offer AI-powered coding assistants for as low as $500 a month, OpenAI’s premium pricing suggests they’re targeting high-end enterprise clients rather than everyday users.
Microsoft’s Multi-Pronged AI Strategy
Beyond developing MAAI, Microsoft is testing AI models from other companies like Meta, xAI, DeepSeek, and Anthropic. They’ve also changed their agreement with OpenAI—shifting from being OpenAI’s exclusive cloud provider to having the “right of first refusal.” This means Microsoft can still host OpenAI’s workloads, but OpenAI is free to use other cloud providers, such as Oracle.
Meanwhile, massive investments are flowing into AI infrastructure. Microsoft alone plans to invest $80 billion into AI data centers in 2025, including a significant portion in India. Other major players, including SoftBank, Oracle, and Nvidia, are reportedly considering investments as high as $500 billion into next-generation AI hardware.
The Belief State Transformer: A Game Changer in AI Reasoning
One of Microsoft Research’s most exciting developments is the Belief State Transformer (BST). This model represents a major evolution in how AI thinks and plans.
Traditional AI models, like GPT-4, process words from left to right, meaning they only consider past information when generating new text. The BST, however, introduces a bidirectional approach—it considers both what came before and what comes after in a sentence or sequence. This helps AI understand context better, particularly for tasks that require long-term planning and reasoning.
For example, in a problem-solving scenario where AI needs to navigate a complex decision tree, standard AI models often take shortcuts, leading to suboptimal results. The BST, on the other hand, prevents these shortcuts by making AI predict both future and past words, forcing it to consider the entire context before making a decision.
Early experiments show that the BST significantly outperforms traditional models in:
- Planning multi-step tasks
- Generating coherent and well-structured stories
- Navigating complex decision-making problems
While the BST is still in early testing phases (currently using around 80 million parameters compared to GPT-4’s hundreds of billions), the technology could scale up to power the next generation of AI models.
What’s Next?
Microsoft’s AI strategy is clear: they want the best models powering their products, whether that means using OpenAI, their own MAAI, or third-party models. Meanwhile, their research into fundamental AI improvements like BST could give them a major edge in creating AI that’s better at planning, reasoning, and generating high-quality text.
At the same time, Microsoft and OpenAI are transitioning from close allies to frenemies—still collaborating, but also competing. Microsoft is no longer shy about calling OpenAI a competitor in areas like search, advertising, and enterprise AI solutions.
With OpenAI’s pricing soaring and competition heating up, what does this mean for the rest of us? One of two things could happen:
- More competition could drive down AI prices, making advanced AI more accessible.
- AI providers could remain expensive, focusing on deep-pocketed enterprise clients.
Either way, one thing is certain: AI innovation isn’t slowing down anytime soon.