Microsoft

Microsoft is a software and cloud company whose AI work spans research, training systems, models and applications. DeepSpeed distributes training state across GPUs, GitHub Copilot brings model-generated code into programming, and Microsoft Foundry provides managed infrastructure for agents. Microsoft develops its own models alongside its OpenAI partnership, and in January 2025 consolidated developer and AI-platform teams into CoreAI to build its Copilot and AI stack.[7][6][15][17][14][13]

These AI products sit alongside earlier generations of Microsoft software and services. The company began with Altair BASIC in 1975, subsequently released Windows and Office, and expanded into cloud services with Windows Azure and Office 365. That earlier history supplies context for its present work across programming tools, infrastructure and workplace applications.[1][20][13]

Bill Gates and Paul Allen founded Microsoft in 1975 around a BASIC interpreter for the Altair 8800. In his anniversary recollection, Gates recalls promising MITS software before it existed, then dividing the development work: Allen simulated the Intel 8080 processor on Harvard’s PDP-10, Gates wrote the main interpreter, and Monte Davidoff developed the math package. They targeted four kilobytes because extra memory was expensive. Gates explains that interpreting instructions as a program ran gave novice programmers immediate feedback; he recalls a successful demonstration followed by licensing to MITS.[2][1]

Gates and Allen signed the BASIC licensing agreement on July 22, 1975, before the Microsoft name had been chosen or an official partnership existed. A July 29 letter from Gates contains the earliest known written “Micro-soft” reference. Microsoft’s 1975 timeline records year-end sales of $16,005.[1]

Microsoft’s product chronology places MS-DOS on IBM PCs in August 1981, the first Windows release on November 20, 1985, and the first Office release, for Macintosh, on June 19, 1989. Between the Windows and Office launches, Microsoft went public on March 13, 1986, at an offering price of $21 per share.[20]

Microsoft Research was founded in 1991 and conducts fundamental research through regional labs, academic and industry collaborations, and work with Microsoft product teams. Steve Ballmer succeeded Gates as CEO in January 2000. Microsoft introduced Windows Azure in October 2008, made it generally available in 2010, and launched Office 365 in June 2011. Satya Nadella succeeded Ballmer in February 2014.[4][20]

Microsoft completed its GitHub acquisition on October 26, 2018. Nat Friedman, formerly CEO of Xamarin, became GitHub’s CEO, reporting to Microsoft Cloud + AI executive Scott Guthrie. At closing, Microsoft committed to GitHub operating independently and remaining an open platform.[5]

In July 2019, Microsoft announced a $1 billion investment in OpenAI and a multiyear partnership to jointly build Azure AI supercomputing technologies. OpenAI was to move its services to Azure, and Microsoft became its preferred commercialization partner.[21]

Microsoft released DeepSpeed, an open-source library compatible with PyTorch, in February 2020. Its Zero Redundancy Optimizer (ZeRO) addresses the memory consumed by replicated training state. Ordinary data-parallel workers process different data while retaining copies of parameters, gradients and optimizer state. ZeRO partitions that state across workers and communicates the pieces needed during computation. Its stages are cumulative: stage 1 partitions optimizer state; stage 2 also partitions gradients; stage 3 also partitions parameters.[7][8]

The initial release implemented stage 1. The DeepSpeed team—led by Yuxiong He, with researchers and engineers including Samyam Rajbhandari, Jeff Rasley and Olatunji Ruwase—reported using it alongside NVIDIA Megatron-LM to train the 17-billion-parameter Turing-NLG model. The launch’s trillion-parameter discussion was a memory-capacity argument, rather than an achieved training result.[7]

GitHub Copilot made editing context an interface to a model. Its June 29, 2021 technical preview drew on surrounding code to suggest whole lines or functions inside the editor. Developed with OpenAI and powered by OpenAI Codex, the original Copilot brought generation into programmers’ existing workflow. GitHub proposed writing tests and exploring APIs as uses for the tool.[6]

Microsoft extended its OpenAI partnership in January 2023 through an announced multiyear, multibillion-dollar investment, following investments in 2019 and 2021. The extension covered expanded supercomputing infrastructure, deployment of OpenAI models across Microsoft products, and Azure OpenAI Service. Azure’s exclusive-cloud role belonged to those historical terms; the agreement changed again in 2026.[22][14]

Phi-1 explored educational training data as an efficiency axis. In the June 2023 research project Textbooks Are All You Need, Microsoft Research authors including Suriya Gunasekar, Yi Zhang, Jyoti Aneja and Caio César Teodoro Mendes combined filtered code with GPT-3.5-generated Python textbooks. GPT-4 annotated educational quality on approximately 100,000 samples; a random forest classifier using pretrained CodeGen embeddings then filtered the larger code collection. This CodeTextbook corpus supplied pretraining, followed by fine-tuning on approximately 180 million tokens of synthetic Python exercises and solutions, called CodeExercises.[9][10]

The authors reported 50.6% HumanEval pass@1 and 55.5% MBPP pass@1 for the 1.3-billion-parameter model: code-generation benchmark results using one generated solution. They trained the base model with DeepSpeed on eight A100 GPUs in under four days, followed by seven hours of fine-tuning. Phi-1 specialized in Python and retained package and API knowledge gaps, sensitivity to prompt errors and length, and errors in its synthetic training data. The authors released the model but withheld some synthetic-data generation details for proprietary reasons.[10]ⁿ¹

In John Maeda’s March 2023 account of the Schillace Laws, Microsoft’s then Deputy CTO Sam Schillace offered the heuristic: “Code is for syntax and process; models are for semantics and intent.” Maeda recounts Schillace suggesting an output prefix for generated JSON, followed by feeding parser errors and the previous output back to the model for another attempt. This proposed repair loop uses an executable syntax check; it does not guarantee convergence or correct meaning.[11]

In a March 2024 conversation, Schillace recalled building Semantic Kernel with a small team to combine code and inference. His “textbook factory” example placed the larger workflow in code: identify a topic and educational level, establish a curriculum, divide it into chapters, and produce lessons, a teacher’s guide and a table of contents. Models filled in the pieces. He described this division as a way to keep longer tasks from drifting, while acknowledging that the workflow was limited to its defined domain.[12]

Schillace also reconsidered “Text is the universal wire protocol” as multimodal models developed. He said, “Like that particular law is not going to hold up super well.” This was a qualification of his own engineering principle, rather than a Microsoft policy change.[11][12]ⁿ²

Nadella’s January 13, 2025 CoreAI announcement made the product-platform relationship an organizational objective. CoreAI — Platform and Tools, led by Jay Parikh, combined Dev Div, AI Platform and selected Office of the CTO teams, including AI Supercomputer and AI Agentic Runtimes. Its mission covered the Copilot and AI stack for Microsoft and external customers, including GitHub Copilot. Nadella proposed a tight feedback loop between that product and the platform’s roadmap, spanning Azure AI Foundry, GitHub and VS Code.[13]

Following OpenAI’s October 2025 recapitalization, Microsoft reported an investment in OpenAI Group PBC valued at approximately $135 billion, representing roughly 27% on an as-converted diluted basis, including employees, investors and the OpenAI Foundation. That was the stake’s reported value and ownership basis at the event, not cash invested. The announcement also said OpenAI contracted to purchase an incremental $250 billion in Azure services and Microsoft could independently pursue AGI, alone or with other partners. The service commitment was not recognized revenue or cash collected.[23]

The April 27, 2026 amendment retained Microsoft as OpenAI’s primary cloud partner, while permitting OpenAI to serve all its products across cloud providers. Azure-first shipping retained the exception “unless Microsoft cannot and chooses not to support the necessary capabilities.” Microsoft’s model and product IP license became non-exclusive through 2032. Microsoft would stop paying revenue share to OpenAI; OpenAI’s payments to Microsoft would continue through 2030, at the same percentage but subject to a total cap. Microsoft remained a major shareholder.[14]

Alongside the partnership, Mustafa Suleyman’s June 2026 MAI announcement presented seven internally developed models across reasoning, coding, images, transcription and voice. Suleyman described MAI-Code-1-Flash as deeply integrated into GitHub Copilot, VS Code and the Microsoft stack, and announced broader developer distribution through Foundry, OpenRouter, Fireworks and Baseten. Microsoft AI described training its reasoning models from scratch as part of a goal of long-term self-sufficiency.[17]

The same announcement introduced a Mayo Clinic collaboration to develop a healthcare model using Mayo’s clinical expertise, de-identified data and longitudinal insights. The proposed model would be owned by Mayo Clinic, deployed first within Mayo’s environment, and made available through Foundry after validation. Earlier diagnosis and improved treatment planning were expected benefits, rather than established clinical results.[17]

By September 2026, Microsoft Foundry Agent Service offered three build paths. Prompt agents specify instructions, a model and tools through configuration. Hosted agents run custom code supplied as a container or source archive, with managed hosting, scaling and identity. Developers can also keep code running elsewhere and call the Responses API through a Foundry project endpoint without creating a persisted agent resource. Prompt and hosted agents are the two managed agent types.[15]

Hosted agents run in per-session VM-isolated sandboxes with a dedicated Microsoft Entra identity. Session files survive turns and idle compute periods, but sessions are permanently deleted after 30 days of inactivity. Conversation history is separate: the Responses protocol manages it, while Invocations leaves it to application code. A durable key-value state store supplies another mechanism; its item idle window defaults to 30 days, writes renew the window, and a store can be configured for non-expiring items. Background execution and recovery after process interruption are also distinct features.[16]

For the fiscal year ended June 30, 2026, Microsoft’s unaudited earnings release reported consolidated revenue of $331.839 billion, operating income of $155.237 billion, and GAAP net income of $133.749 billion. Nadella, chairman and CEO, reported annual Azure revenue exceeding $100 billion and Microsoft 365 Copilot reaching more than 30 million paid seats. Paid seats establish subscription scale, rather than active usage or productivity.[18]

The full-year statements reported $35.562 billion in research and development expense, $182.935 billion in operating cash flow, and $115.948 billion in cash additions to property and equipment. The property-and-equipment figure covers that consolidated accounting line; it is not exclusively AI spending or capital expenditure including all leases. At June 30, Microsoft reported $76.843 billion in cash, cash equivalents and short-term investments.[18]