Microsoft AI

Microsoft AI is the artificial intelligence business of Microsoft Corporation. It spans the Copilot assistants sold with Windows and Microsoft 365, the Azure AI platform on which customers build and run models, a family of in-house models, and a stake of about 27 per cent in OpenAI, whose technology underpins much of the product line 1 2.

Microsoft is not a frontier laboratory in the way OpenAI or Anthropic is. It is a distribution company that has become one of the largest sellers of AI, partly by reselling another company's models and partly by building its own. Its advantage is reach: Windows, Office, Teams, GitHub, Dynamics and Azure together touch almost every large organisation in the world 2.

The strategy has been consistent since 2019: invest in a leading model developer, host its models exclusively for long enough to build a business, then diversify. That strategy produced Copilot, the fastest-growing product line in Microsoft's history by some measures, and it produced a dependency that the company has spent years unwinding 3.

Microsoft Research, founded in 1991, is one of the largest industrial computer-science laboratories in the world and has contributed to speech recognition, machine translation, computer vision and programming languages. For most of its history its work reached customers indirectly, through features in Office, Windows and Bing rather than through products labelled as AI 1.

That changed in stages. In 2016 Microsoft launched Tay, a chatbot on X designed to learn from conversation; it was taken offline within a day after producing racist and offensive messages. The company apologised and described the episode as a lesson in adversarial behaviour 4.

The decisive move came in 2019, when Microsoft invested a billion dollars in OpenAI and agreed to host its models on Azure. A further investment, reported at around ten billion dollars, followed in January 2023, alongside an extension of the cloud agreement and a plan to deploy OpenAI's models across Microsoft's consumer and enterprise products 3.

Bing Chat, launched in February 2023, was the first public result. It answered questions with citations, could hold a conversation, and was built on GPT-4. Within a year the company had renamed the product Copilot and begun inserting it into every application it sold 5.

Copilot is now a family rather than a product. The consumer version is a chat assistant with web grounding, image generation and voice, available through a website, Windows, Edge and the Microsoft 365 mobile applications. The enterprise version is embedded in Word, Excel, PowerPoint, Outlook and Teams, where it drafts documents, summarises threads and answers questions about a company's own files 6.

The enterprise product is the commercially important one. Microsoft 365 Copilot was announced in March 2023 and became generally available in November, priced per user per month on top of an existing subscription. It uses a retrieval system that indexes a tenant's documents, mail and chats, and it inherits the organisation's permissions, which is both its selling point and its main technical difficulty 5.

GitHub Copilot, launched in 2021, predates the current wave and remains the most clearly successful AI product Microsoft sells. It completes code in the editor, answers questions about a repository, and has been extended into an agent that can open pull requests. Its adoption among developers is the strongest evidence that AI tools can change how work is done rather than merely assist at the margins 6.

Copilot+ PCs, announced in May 2024, moved AI onto the device. The category requires a neural processing unit capable of tens of trillions of operations per second, and its flagship feature, Recall, was intended to let users search everything they had seen on screen. Recall became the company's most criticised AI feature within a week of announcement 7.

Azure is where Microsoft's AI business actually scales. The Azure OpenAI Service, launched in 2021, gave enterprises access to OpenAI's models with Microsoft's data-handling terms, private networking and regional controls, which many regulated customers required. It grew quickly and became the standard route for regulated enterprises to use OpenAI's models inside their own subscriptions 3.

Microsoft has since positioned Azure as a marketplace rather than a single-model platform. Azure AI Foundry offers models from OpenAI, Meta, Mistral, Cohere and Microsoft itself, with tools for evaluation, fine-tuning and deployment. The strategy accepts that customers will want more than one model and tries to make Azure the place where the choice is made 1.

The company has also designed its own silicon. The Maia accelerator, announced in November 2023, is aimed at inference workloads, and the Cobalt processor is an Arm-based server chip. Neither replaces NVIDIA GPUs for training, but both reduce Microsoft's exposure to a single supplier for the less demanding part of the workload 2.

Azure's scale is the foundation of the business. Microsoft operates data centres in more than sixty regions, and its capital spending has risen sharply as it builds capacity for AI. In the fiscal year to June 2025 the company reported capital expenditure of more than 80 billion dollars, most of it for AI infrastructure 2.

Microsoft AI, the division, was created in March 2024 under Mustafa Suleyman, a co-founder of DeepMind who had led Inflection AI. The division consolidated consumer AI products — Copilot, Bing, Edge — under one leader and was given a mandate to build models as well as products 8.

Its first significant hire came through a transaction: Microsoft licensed Inflection's technology for a reported 650 million dollars and hired much of its staff, including Suleyman. The deal was structured as a licensing arrangement rather than an acquisition, which avoided an antitrust filing and was noted by regulators as an example of how AI consolidation can occur without a merger 5.

The company's model family began with Phi, a series of small models designed to show that carefully curated training data could produce capable systems at a fraction of frontier scale. Phi-1, in June 2023, focused on code; Phi-3 and Phi-4 extended the approach to general reasoning and were released with weights and technical reports 8.

In 2025 the division began releasing models under the MAI name: MAI-1-preview for general chat, MAI-Voice-1 for speech, and MAI-Image-1 for images. The models are trained in-house and deployed in Copilot, which reduces — without eliminating — the company's reliance on OpenAI for the core capability 8.

The partnership has been the defining fact of Microsoft's AI business. Microsoft provided capital and compute; OpenAI provided models; Azure gained an exclusive franchise that no competitor could match. The arrangement made Microsoft the most valuable company in the world for a period in 2024, and it made OpenAI dependent on a single cloud 3.

The relationship has been renegotiated repeatedly as OpenAI's ambitions grew. The October 2025 recapitalisation converted Microsoft's stake into about 27 per cent of OpenAI Group PBC, worth roughly 135 billion dollars, extended Microsoft's rights to OpenAI's intellectual property, and loosened the exclusivity that had bound OpenAI to Azure. OpenAI also committed to purchase large volumes of Azure capacity, reported at 250 billion dollars over several years 2.

Both sides have diversified. OpenAI signed computing agreements with Oracle and others and announced data-centre projects that do not involve Microsoft as the lead. Microsoft, meanwhile, added models from Meta, Mistral and others to Azure and built its own MAI family. The partnership remains central to both companies, but it is no longer exclusive in either direction 1.

The commercial logic still holds. Microsoft sells more OpenAI models than anyone else, and OpenAI reaches enterprises through a sales force it could not build quickly. The tension is strategic: each company would prefer not to depend on the other.

Microsoft is a public company listed on Nasdaq and one of a small number whose market value has exceeded three trillion dollars. Its AI business is not reported as a separate segment; it appears inside Intelligent Cloud, Productivity and Business Processes, and More Personal Computing, which makes the returns on AI spending difficult to isolate 2.

The company's financial position is unusually strong. It funds capital expenditure from operating cash flow, carries little debt relative to its earnings, and has continued to return capital to shareholders while spending more than 80 billion dollars a year on infrastructure. That combination is rare among companies building AI at scale, and it is the main reason Microsoft can afford to be a fast follower rather than a frontier leader 2.

Investors have treated the spending as justified. The share price has risen with the AI cycle, and the company's guidance has repeatedly pointed to AI as the driver of growth in Azure. The risk is that the returns arrive more slowly than the spending, a concern that has surfaced in every quarter since 2024 5.

The OpenAI stake adds a large, illiquid asset to the balance sheet. Its valuation is marked by the price of OpenAI's most recent funding round rather than by a market, and it is accounted for in a way that gives limited visibility into the underlying business 2.

Tay, in March 2016, remains the canonical example of an AI product released without adequate safeguards. Microsoft built a chatbot designed to learn from its conversations on X, and users deliberately taught it to produce racist and antisemitic messages. The company took it offline within sixteen hours and apologised, describing the episode as a coordinated attack as well as a design failure 4.

Recall, in 2024, was the more instructive failure because it was foreseeable. The feature took screenshots of a user's activity every few seconds and stored them in a local database so that the user could search their own history. Security researchers found that the database was unencrypted and accessible to other accounts on the same device, and that it could capture passwords and financial details. Microsoft delayed the launch, made the feature opt-in, encrypted the store and required Windows Hello authentication before reopening it to testers 7.

The episode did lasting damage to the company's reputation for security, which had already been strained by a series of breaches. It also illustrated a structural problem: AI features that collect data to be useful are in tension with the security model of a general-purpose operating system, and Microsoft owns both sides of that tension 5.

A third area of criticism concerns competition. Microsoft's investment in OpenAI, its licensing of Inflection's technology and its bundling of Copilot into Office have all attracted regulatory attention in the United States, the European Union and Britain. No case has yet forced a change in the arrangements, but the inquiries have made Microsoft's dealmaking more cautious 2.

Whether Copilot makes workers more productive is disputed. Microsoft has published studies and customer testimonials reporting time savings and higher satisfaction; independent trials have found effects that are smaller, unevenly distributed and dependent on the task. The gap is not surprising: vendor studies select willing customers, and productivity is difficult to measure in knowledge work 5.

The value of the OpenAI stake is contested in a different way. Microsoft treats it as a strategic asset and a source of model access; sceptics note that the stake is a minority position in a company that is now free to buy computing elsewhere, and that the intellectual property rights have a horizon. The October 2025 agreement was presented by both companies as a strengthening of the partnership, a characterisation that is consistent with the terms and also with the need to reassure investors 2.

A third disagreement concerns Microsoft's role in the field. It is either the company that made AI useful to ordinary businesses, or a distributor that captured the value of another company's research and now competes with it using licensed technology. Both descriptions are defensible, and the company's behaviour supports each of them 1.

Microsoft is the largest seller of enterprise AI and the second-largest provider of cloud infrastructure. Its models are not the best in the world by most measures, but they are embedded in the software that organisations already use, and that distribution is worth more than a benchmark lead for most customers 1.

Its position depends on two things it does not fully control: the health of its partnership with OpenAI, and the willingness of customers to pay a per-seat premium for AI features. The first has been renegotiated successfully twice; the second is still being tested, and the early evidence suggests that adoption is slower and more uneven than the initial projections 5.

For the encyclopedia, Microsoft AI is the strongest case for distribution beating invention. The company was late to large language models, bought its way into the leading laboratory, and turned the resulting access into a product line that reaches more knowledge workers than any competitor's. Whether that position survives the next generation of models is the open question 2.