TL;DR
Microsoft has released a new speech transcription model that its AI chief Mustafa Suleyman calls the most advanced of its kind, but acknowledged the company still lacks the compute capacity to build frontier-scale large language models. Suleyman said Microsoft is competing in the “mid-class range” while scaling up infrastructure.
What Microsoft announced
The new model focuses on speech transcription rather than the text generation and coding tasks where Anthropic, Google, and OpenAI currently lead. Suleyman told the Financial Times that the mid-class positioning is “optimal” for balancing cost, performance, and large-scale deployment, and that the transcription model would undercut competitors on pricing.
It is a candid admission of where Microsoft stands. Despite being one of the world’s largest cloud computing businesses, the company has not yet released large language models that compete at the frontier. Its first foundation model, MAI-1, was announced last year but remains in preview and is not generally available.
The push for AI self-sufficiency
Suleyman, who co-founded Google DeepMind before joining Microsoft in 2024, was speaking from an off-site meeting in Miami for Microsoft’s new Superintelligence team. He and CEO Satya Nadella addressed the 350-strong group about the company’s compute roadmap.
The mission, Suleyman said, is “to deliver AI self-sufficiency for Microsoft over the next two or three years” by building frontier-scale chip clusters and investing in data. This drive comes after Microsoft restructured its exclusive relationship with OpenAI, freeing both parties to compete more directly.
Microsoft faces the same capacity constraints affecting other hyperscalers: local opposition to data centres, equipment shortages, and limited power and labour supply. Allocating data centre capacity between internal AI work, customers like OpenAI, and traditional enterprise software adds another layer of complexity.
Looking forward
Suleyman’s role was narrowed last month to focus purely on model development, with former Snap executive Jacob Andreou taking charge of Copilot products. The reorganisation suggests Microsoft is treating model-building as a distinct priority. Whether the compute ramp arrives on schedule will determine how quickly the company can move from mid-class models to genuine frontier competition.