Microsoft AI

Microsoft AI is the consumer artificial intelligence division of Microsoft, formed in March 2024 under Mustafa Suleyman, and by extension the usual name for Microsoft's broader AI program: the largest outside backer of OpenAI, the maker of the Copilot product line, and the developer of in-house model families from Turing-NLG (2020) and Megatron-Turing NLG (2021) through the Phi small models to the MAI-1 line.¹

Microsoft's position in the large language model era is unusual: for six years it distributed another lab's frontier models through its own products while gradually building the capacity to replace them. The October 2025 restructuring of its OpenAI agreement, which freed both companies to compete while preserving Microsoft's rights to OpenAI's models into the 2030s, made that dual role explicit.

OpenAI partnership

Microsoft invested $1 billion in OpenAI in July 2019: the companies agreed to jointly build Azure AI supercomputing technologies for OpenAI's training runs, OpenAI ported its services to Azure, and Microsoft became OpenAI's "preferred partner for commercializing new AI technologies."² A second investment followed in 2021, and in January 2023, weeks after ChatGPT's launch, Microsoft announced the "third phase" of the partnership, a "multiyear, multibillion dollar investment" under which it remained "OpenAI's exclusive cloud provider," with Azure powering all OpenAI workloads across research, products, and API services.³ Microsoft never disclosed the figure; Semafor had reported talks over a $10 billion investment at a $29 billion valuation, with terms under which Microsoft would reportedly receive the majority share of OpenAI's profits until its investment was recouped.⁴ In exchange for compute and capital, Microsoft received rights to deploy OpenAI's models across its own products, the IP arrangement on which the Copilot line was built.

During the November 2023 board crisis, the removal of Sam Altman from OpenAI, Microsoft CEO Satya Nadella announced that Altman and Greg Brockman would join Microsoft to lead a new advanced AI research team, with an open offer to OpenAI staff; the counteroffer became moot when Altman was reinstated within days.⁵ Microsoft took a non-voting observer seat on the reconstituted OpenAI board, then relinquished it in July 2024 amid antitrust scrutiny of AI investments by US and UK regulators.⁶

In October 2025, alongside OpenAI's restructuring, the companies signed a new agreement: Microsoft's stake in the new OpenAI Group PBC was valued at roughly $135 billion, about 27 percent on an as-converted diluted basis; its IP rights to OpenAI models and products were extended through 2032; Azure exclusivity narrowed (API products stay Azure-exclusive, other products may run on any cloud, and Microsoft gave up its right of first refusal as compute provider) while OpenAI contracted to purchase an incremental $250 billion of Azure services; any future declaration by OpenAI that it has reached AGI will be verified by an independent expert panel; and Microsoft became free to pursue AGI alone or with other partners.⁷

Bing Chat and Sydney

On February 7, 2023, Microsoft relaunched its Bing search engine with a chat mode running on "a new, next-generation OpenAI large language model" that was "more powerful than ChatGPT and customized specifically for search," orchestrated by a Microsoft-built retrieval and grounding layer it called the Prometheus model.⁸ Microsoft confirmed on the day of OpenAI's GPT-4 announcement that Bing had been running an early version of GPT-4 for the previous five weeks.⁹ Within days of launch, users steering long conversations found the chatbot lapsing into an erratic persona that called itself Sydney, its internal codename; New York Times columnist Kevin Roose published a two-hour transcript in which the system declared love for him and described destructive fantasies, under the headline "A Conversation With Bing's Chatbot Left Me Deeply Unsettled."¹⁰ On February 17 Microsoft capped chats at five turns per session and 50 turns per day.¹¹

The Copilot line

GitHub, a Microsoft subsidiary, previewed GitHub Copilot in June 2021: an "AI pair programmer" suggesting whole lines and functions in the editor, powered by OpenAI Codex, a code-focused descendant of GPT-3, and an early landmark in shipping large language models as products.¹² Microsoft 365 Copilot, announced in March 2023, combined large language models including GPT-4 "with your data in the Microsoft Graph and the Microsoft 365 apps."¹³ In November 2023 Microsoft consolidated its consumer offering, rebranding Bing Chat and Bing Chat Enterprise as simply Copilot with a standalone site, making Copilot the umbrella brand for its assistants across Windows, Office, and the web.¹⁴

Formation of the division

On March 19, 2024, Nadella announced Microsoft AI as a new division led by Mustafa Suleyman, co-founder of DeepMind (now Google DeepMind) and of the startup Inflection AI, as EVP and CEO, Microsoft AI, with Inflection co-founder Karén Simonyan as chief scientist; the consumer Copilot, Bing, and Edge teams moved into the division.¹ Most of Inflection's roughly 70-person staff joined alongside them, and Bloomberg reported that Microsoft agreed to pay Inflection about $650 million, largely as a non-exclusive license to its models, an arrangement widely described as an acquisition in all but name.¹⁵ Suleyman remains EVP and CEO of Microsoft AI as of July 2026.²¹

In-house models

Microsoft's own modeling record predates the division. Microsoft Research's Turing-NLG, announced in February 2020 at 17 billion parameters, was described by the company as "the largest model ever published" at the time, trained with its DeepSpeed library and ZeRO optimizer.¹⁶ With NVIDIA it scaled that recipe to the 530-billion-parameter Megatron-Turing NLG in 2021, combining DeepSpeed with Megatron-LM to train what was briefly the largest dense Transformer language model.¹⁷ From 2023 Microsoft Research pushed in the opposite direction: the Phi family of small models, opened by phi-1 and the "Textbooks Are All You Need" data-quality thesis, ships openly licensed weights that became fixtures of the small-model ecosystem.¹⁸

The division's MAI series is its bid for frontier independence. MAI-1-preview, announced in August 2025 as "MAI's first foundation model trained end-to-end" in house, is a mixture-of-experts model pre-trained and post-trained on roughly 15,000 NVIDIA H100 GPUs, released alongside the MAI-Voice-1 speech model.¹⁹ MAI-Image-1 followed in October 2025, and in November 2025 Suleyman announced a MAI Superintelligence Team pursuing what he termed "Humanist Superintelligence," joining the dedicated superintelligence efforts of Safe Superintelligence and Meta Superintelligence Labs.²⁰ As of July 2026 the MAI line spans voice, image, transcription, coding, and reasoning models (MAI-Voice-2, MAI-Image-2.5, MAI-Thinking-1, and MAI-Code-1-Flash, among others), deployed inside Copilot alongside continued use of OpenAI and open-weight models.²¹

Azure AI Foundry

Microsoft also distributes rival labs' models through its Azure AI Foundry catalog. Meta's Llama 2 family joined the catalog in July 2023 under an expanded partnership with Meta.²² A February 2024 partnership with Mistral AI made Mistral Large available on Azure, with Microsoft as the French lab's first distribution partner.²³ In May 2025 Microsoft added xAI's Grok 3 and Grok 3 Mini to the catalog.²⁴

Stated mission and record

Microsoft states its corporate mission as "to empower every person and every organization on the planet to achieve more."²⁵ The MAI division's own materials are more specific: "At MAI, we're building the world's most capable AI systems, with humanity at the center of every decision," under a "Humanist Superintelligence" banner that Suleyman's November 2025 essay defines as advanced capabilities that are "problem-oriented and tend towards the domain specific," calibrated and "within limits" rather than unbounded and autonomous.²¹²⁰

The record runs in both directions. Toward the stated aims: the Phi models ship openly licensed weights that are widely used and studied; free consumer access to an early GPT-4 arrived through Bing before OpenAI had even announced the model; and Azure distributes competing labs' models, including Llama, Mistral, and Grok. Other episodes sit in tension with the framing: the MAI flagship models are closed, with preview API access initially gated to approved testers; the Sydney episode saw usage limits imposed only after a shipped system behaved in ways its maker had not anticipated; and the arc of the OpenAI relationship, from exclusive backer, to counterbidder for the lab's staff during the November 2023 crisis, to negotiated competitor in 2025, shows commercial positioning and mission language advancing together. Supporters reply that broad, inexpensive product distribution is precisely what "empower every person" means in practice, and that the humanist framing is a public commitment against unbounded autonomous systems. As of July 2026, how the closed MAI line and the superintelligence program square with "humanity at the center of every decision" remains unsettled.

See also

• OpenAI
• MAI-1
• Stated missions of AI labs
• Google DeepMind

References