Microsoft Builds Its Own Models as OpenAI Ties Fray

Microsoft just released two new models: MAI-1-preview, a large language model, and MAI-Voice-1, a speech model. The company described MAI-Voice-1 as efficient enough to generate a minute of audio in under a second on a single GPU, while MAI-1-preview was trained on 15,000 Nvidia H100 GPUs and is already being tested for Copilot features. The launch came just weeks after Microsoft announced it was incorporating OpenAI’s GPT-5 into its consumer, developer, and enterprise products. On paper, Microsoft is both doubling down on OpenAI and setting up its own alternatives. The moves are not contradictory. They are a correction of risk that the company built into its strategy. For years, Microsoft has leaned almost entirely on OpenAI to supply the foundation models running its Copilot assistant, Bing AI features, and developer tools. Microsoft invested more than $13 billion, won exclusivity to host OpenAI models on Azure, and oriented major products around GPT technology. That approach gave Microsoft a lead over peers in bringing AI to end users. But it also created a single point of failure. Other hyperscalers (Google, Amazon, Meta) developed their own foundation models while still experimenting with external collaborations. Microsoft was the outlier in placing its bet almost entirely on one startup. This dependence now sits uneasily against signs of strain. Microsoft and OpenAI are renegotiating their contract, which runs until 2030. Among the sticking points are whether OpenAI can sell API access through clouds beyond Azure and whether Microsoft should retain rights to OpenAI’s intellectual property. A separate clause gives OpenAI the option to cut Microsoft’s access if it develops what it defines as artificial general intelligence. Tensions have surfaced at the leadership level too. Reports describe reduced communication between Satya Nadella and Sam Altman, who just two years ago jointly unveiled Copilot as the centerpiece of Microsoft’s AI push. At the same time, OpenAI has broadened its infrastructure partners, turning to CoreWeave, Google, and Oracle to handle surging demand. Its ChatGPT assistant now reaches hundreds of millions of people weekly, making OpenAI less dependent on Microsoft’s distribution channels (CNBC). Seen against this backdrop, Microsoft’s new models are less about leapfrogging GPT-5 and more about filling a gap in its own capabilities. Mustafa Suleyman, the former DeepMind and Inflection co-founder hired last year to run Microsoft AI, put it in an interview with Semafor: “We are one of the largest companies in the world. We have to be able to have the in-house expertise to create the strongest models in the world”. Until now, the company’s internal model program consisted mainly of smaller, open-source projects like Phi-3. With MAI-1-preview and MAI-Voice-1, Microsoft has demonstrated it can train large-scale models on its own clusters. Suleyman described this as proof that efficiency in data selection can rival brute-force scale, saying, “This is a model that is punching way above its weight.” MAI-1-preview is already being tested publicly on LMArena, a benchmarking platform. On launch it ranked 13th for text tasks, below models from Anthropic, Google, Mistral, and OpenAI. These are decent results given its smaller training footprint. MAI-Voice-1 has been deployed in Copilot Daily, which reads top news stories aloud, and in experimental podcast-style features. The lesson is straightforward: relying on a single outside lab for core technology exposed Microsoft to risks beyond its control. Negotiations over contract terms, decisions about cloud exclusivity, or disagreements over what counts as “general intelligence” could all leave Microsoft scrambling. Building internal models doesn’t eliminate those risks, but it reduces the concentration. That explains why the company is pursuing a two-track strategy. GPT-5 is being rolled out across Microsoft 365, GitHub Copilot, and Azure AI Foundry. MAI-1-preview and MAI-Voice-1 are being woven into Copilot features and tested with developers. If one track falters, the other exists. If they both thrive, Microsoft benefits from scale and redundancy. The models themselves may not upend the competitive order. What they do is give Microsoft its own end-to-end training pipeline, something it has not had before at scale. Microsoft now has a baseline of technical competence. It can attract talent with experience training frontier systems, refine models to suit its own products, and prepare for hardware optimized for its workloads. Those are assets it did not have two years ago. The launches mark the point where Microsoft shifts from being a distributor of someone else’s AI to being a builder. For a company that runs the largest enterprise software franchise in the world, that change matters.

Microsoft Builds Its Own Models as OpenAI Ties Fray

Keep reading. It's free.

Related stories