Mustafa Suleyman says his move away from Microsoft Copilot was never a demotion—it was a plan nine months in the making. The Microsoft AI CEO, in an interview with The Verge, said building frontier AI models is exactly where he wants to be. His team has already delivered MAI-Transcribe-1, a speech-to-text model supporting 25 languages that costs half as much in GPU compute as rival state-of-the-art models.

Mustafa Suleyman is pushing back on the narrative. When Microsoft reshuffled its AI leadership in mid-March, it didn’t look great for him—a former Snap executive was elevated above him on the Copilot side, and the sprawling consumer AI division he once helmed was handed to someone else. Call it a transfer, call it a demotion—the reads weren’t flattering either way. Suleyman’s answer, in an interview with The Verge, was simple: he saw this coming nine months ago, and it was always the plan. Don’t be surprised, he says.“This has been a long-held plan,” he said. Pursuing superintelligence, he added, is “purely my focus” now.

Why observers called Mustafa Suleyman’s move a demotion in the first place

It’s not hard to see where the framing comes from. Suleyman arrived at Microsoft in 2024 as the marquee hire—a DeepMind co-founder brought in to anchor the company’s AI ambitions and give it an identity beyond OpenAI’s shadow. He was the face of Microsoft AI. Then in March, Jacob Andreou, who had been a corporate vice president reporting to Suleyman, was promoted to EVP of Copilot and given a direct line to CEO Satya Nadella. The consumer and commercial Copilot teams Suleyman oversaw now sit under Andreou. Suleyman still reports to Nadella, but his grip on the product side loosened considerably.The broader context makes it sting more. Copilot has had a rough run. The app clocked just 6 million daily active users in February—a rounding error next to ChatGPT’s 440 million and Gemini’s 82 million, according to Sensor Tower. Only 3% of Microsoft 365’s 450 million-plus paying subscribers have the Copilot add-on. Microsoft’s own models have consistently underperformed on benchmarks, hobbled by computing capacity shortages. Pulling Suleyman away from the product side—precisely when Copilot needs a turnaround—is what gave the demotion reading its traction.

MAI-Transcribe-1 is Suleyman’s first answer to the skeptics

Suleyman’s reframe is that the model layer is where the real value will accumulate—and that being handed that mandate isn’t a consolation prize. “The model is the product,” he told CNBC. His job for the next three to five years, as he describes it, is building cost-efficient, enterprise-grade model lineages that wean Microsoft off its dependence on OpenAI, whose IP it only holds rights to through 2032.The first tangible output of that work landed on April 2. Microsoft unveiled MAI-Transcribe-1, a speech-to-text model built for the kind of audio conditions that break lesser systems—overlapping voices, street noise, low-quality recordings. It covers 25 languages, tops the FLEURS benchmark in 11 of them, and processes audio 2.5x faster than Microsoft’s existing Azure transcription service. Suleyman told The Verge it runs at half the GPU cost of comparable state-of-the-art models—a “huge cost-saving” in his words—and starts at $0.36 per hour on Microsoft Foundry.It launched alongside two other in-house models: MAI-Voice-1, which generates 60 seconds of audio in a single second, and MAI-Image-2, which entered the Arena.ai leaderboard as a top-3 model family. All three are now available commercially via Foundry and the new MAI Playground. Suleyman credits a lean 10-person team—deliberately insulated from internal bureaucracy—for moving this fast.Whether this reads as a strategic repositioning or a quiet reassignment still depends on who you ask. Suleyman has clearly decided which story he’s telling.