Forward-looking: Microsoft is investing in AI through a five-year plan and making steady progress each quarter. With ongoing investments, new internal technologies, and recent strategic hires, the company is positioning itself as a leader in generative AI.
Microsoft is developing and testing proprietary large language models, signaling a shift from its previous heavy reliance on OpenAI technology.
This week, the company announced the release of two internally developed AI models: MAI-1-preview, a text-based foundational model, and MAI-Voice-1, its first natural speech generation model. Both are intended to enhance Microsoft's consumer-facing Copilot assistant and other offerings.
MAI-1-preview, now available for public evaluation on the LMArena website, is Microsoft's first foundational AI model built entirely in-house from start to finish. Public testing is designed to drive improvements based on user feedback, and the company expects to deploy the model in certain Copilot text use cases in the near future. Microsoft is also inviting developers to apply for early access to MAI-1-preview as it works to refine the model's capabilities.
MAI-Voice-1 has already been integrated into Copilot's Daily and Podcast features.
MAI-1-preview was trained on approximately 15,000 Nvidia H100 graphics processing units, an approach described as cost-effective by AI division chief Mustafa Suleyman in an interview with Semafor. By comparison, competing AI models such as xAI's Grok have required upwards of 100,000 H100 GPUs for training.
Suleyman emphasized that advances in model performance now depend not only on computing power but also on the efficient selection of training data. "Increasingly, the art and craft of training models is selecting the perfect data and not wasting any of your flops on unnecessary tokens that didn't actually teach your model very much," he said.
Despite these developments, Microsoft maintains a close business relationship with OpenAI. The company has invested over $13 billion in the AI startup, whose GPT models continue to power major features in Microsoft products, including Bing and Windows 11. OpenAI, valued at around $500 billion, also relies on Microsoft's cloud infrastructure for its operations.
Lots more to come! We have big ambitions for where we go next - model advancements, an exciting roadmap of compute, and the chance to reach billions of people through Microsoft's products. We're building AI for everyone. If that resonates, come build it with us. My DMs are open. https://t.co/FeBOAOGI4s
– Mustafa Suleyman (@mustafasuleyman) August 28, 2025
Both companies, however, have signaled growing competition. Microsoft formally listed OpenAI as an industry competitor in its most recent annual report, alongside technology giants such as Amazon, Apple, Google, and Meta.
Meanwhile, OpenAI has begun diversifying its cloud partnerships, leveraging providers like CoreWeave, Google, and Oracle to handle surging demand for its services.
On the LMArena benchmark, Microsoft's new MAI-1-preview ranked 13th for text-processing capabilities as of Thursday, trailing models from Anthropic, DeepSeek, Google, Mistral, OpenAI, and xAI, based on publicly available data.
Nevertheless, Microsoft's decision to train foundational models internally is part of a longer-term effort to gain greater independence in AI development. Suleyman's leadership plays a central role in this strategy. The AI executive, who previously led ventures at DeepMind and co-founded Inflection, joined Microsoft last year with several colleagues from both organizations.