Microsoft

Microsoft Corporation is an American technology company headquartered in Redmond, Washington, and one of the largest suppliers of artificial intelligence compute, models, and applications. This article covers Microsoft as an AI company: Microsoft Research, the internal model programs, the Azure infrastructure it sells to other AI developers, the Copilot products it ships to end users, and the commercial partnership with OpenAI that has shaped much of that work since 2019. Microsoft's other businesses appear only where they bear on that story. Founded in 1975 by Bill Gates and Paul Allen, the company built its position in software distribution and cloud services long before generative AI, and that installed base is the main reason its AI products reached tens of millions of paid seats within three years of launch.^[1][2]

Microsoft's AI position should not be reduced to its OpenAI relationship, and it should not be described as independent of it either. The company runs a research organization founded in 1991, develops the Phi and MAI model families, designs its own accelerators and datacenters, and distributes AI features through Microsoft 365, GitHub, Windows, and security software. At the same time, OpenAI models still power large parts of the Copilot line, OpenAI is Azure's largest single compute customer, and Microsoft holds roughly 27 percent of OpenAI Group PBC. The two companies renegotiated their contract four times between 2019 and 2026, and the April 2026 amendment left Microsoft as a primary but no longer exclusive partner.^[3]

Separate articles cover Microsoft Copilot, GitHub Copilot, Microsoft 365 Copilot, Microsoft MAI, Azure, and individual models, executives, and acquired companies. This page is the index that connects them.

Corporate context

Microsoft is incorporated in Washington and its common stock trades on Nasdaq under the symbol MSFT. Its fiscal year ends on June 30. It reports three segments: Productivity and Business Processes (commercial and consumer Microsoft 365, LinkedIn, Dynamics), Intelligent Cloud (Azure and other cloud services, server products, GitHub cloud services, Nuance Healthcare, enterprise services), and More Personal Computing (Windows and devices, gaming, search and news advertising).^[1]

Microsoft's earlier history sits outside this article's scope, but three parts of it bear on the AI story. The company began in 1975 writing a BASIC interpreter for the MITS Altair 8800.^[2] It built its first large business on licensing an operating system for the IBM PC and for compatible machines, which established distribution as its core advantage.^[7] And in United States v. Microsoft the federal appeals court upheld liability for unlawfully maintaining the Windows monopoly, with a November 2002 final judgment imposing conduct restrictions and disclosure obligations rather than a breakup.^[8] That precedent is why regulators examine Microsoft's bundling behavior with particular attention, a pattern that recurs in the Teams and business-software matters described below.

Those segment categories cut across Microsoft's AI activity rather than isolating it. Copilot revenue appears in more than one segment, AI infrastructure shows up in both Azure costs and Azure revenue, and GitHub and Nuance sit inside Intelligent Cloud. Microsoft does not report AI as a separate operating segment, which is why every public "AI revenue" figure from the company is a management-defined measure rather than an audited line item.

For fiscal 2025, ended June 30, 2025, Microsoft reported revenue of $281.724 billion, operating income of $128.528 billion, and net income of $101.832 billion, with Azure revenue passing $75 billion for the first time.^[1] In the third fiscal quarter, ended March 31, 2026, revenue was $82.886 billion, Microsoft Cloud revenue was $54.5 billion, and Azure and other cloud services grew 40 percent.^[4] For fiscal 2026 as a whole, ended June 30, 2026, it reported revenue of $331.8 billion, up 18 percent, and GAAP net income of $133.7 billion, up 31 percent. Fourth-quarter revenue was $90.0 billion, Microsoft Cloud revenue was $59.3 billion, Azure and other cloud services grew 43 percent, and commercial remaining performance obligation reached $678 billion, up 84 percent.^[35] On the accompanying call, chief financial officer Amy Hood said Azure "surpassed $100 billion, up 41%" for the year.^[36]

Microsoft employed about 228,000 people as of June 30, 2025, including 80,000 in product research and development.^[1] On July 6, 2026, it announced the elimination of about 4,800 roles, roughly 2.1 percent of its global workforce.^[6] Satya Nadella has been chief executive since February 4, 2014 and is also chairman; Hood is chief financial officer and Brad Smith is vice chair and president.^[9][10] Microsoft reorganizes its AI groups often enough that any roster of division heads goes stale quickly, so the organizational descriptions below are dated snapshots rather than a stable chart.

Microsoft Research

Microsoft Research was founded in 1991 as an industrial research organization modeled on an academic laboratory: researchers publish openly, collaborate with universities, and are not assigned to product deadlines. Its labs relevant to AI include Redmond, Cambridge (UK), New York City, New England, Montreal, Asia, India, Africa, AI Frontiers, AI for Science, and Health Futures.^[11] The distinction between Microsoft Research and Microsoft's product engineering matters when reading claims about the company: several of the systems most often attributed to "Microsoft AI" originated in the research organization and were later handed to product groups, and a few of them have since left Microsoft's governance entirely.

Contributions to the field

The single most cited result to come out of Microsoft Research is residual learning. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, then at Microsoft Research Asia, showed that adding identity shortcut connections made very deep neural networks trainable, and their 152-layer network won the 2015 ImageNet classification task. The architecture is universally known as ResNet, and the skip connection it introduced is now standard in computer vision backbones and in the transformer block itself.^[12]

Later work is more scattered but follows a recognizable pattern: architectural modifications aimed at efficiency or context length, plus domain-specific models released with open weights. Examples include DeBERTa, which added disentangled attention and an enhanced mask decoder to BERT-style encoders; BioGPT, a generative model pretrained on biomedical literature; Florence-2, a unified vision foundation model with a single sequence-to-sequence interface for captioning, detection, and grounding; LongNet and LongRoPE, two approaches to extending context windows; YOCO, a decoder-decoder architecture that caches keys and values once; the Differential Transformer, which subtracts two softmax attention maps to suppress attention noise; and BitNet, a line of work on 1-bit and ternary weight quantization for large language models. Not all of these became products, and several are best read as published experiments rather than deployed systems.

Microsoft Research also produced the ZeRO memory optimizer and the DeepSpeed library, the Magentic agent line, GraphRAG, and AutoGen. Those are covered under research software below, because their significance is mostly as tooling used outside Microsoft.

AI for Science and Health Futures

Two Microsoft Research labs apply machine learning to the natural sciences rather than to software, and both published in Nature in 2025.

MatterGen is a diffusion model for inorganic materials design that generates crystal structures directly, jointly altering elements, atomic positions, and the periodic lattice, and can be fine-tuned to steer generation toward property constraints such as a target bulk modulus or magnetic density. Microsoft Research published it in Nature in January 2025 under the title "A generative model for inorganic materials design," reported that it outperformed prior screening approaches at producing stable, unique, and novel structures, and released the implementation on GitHub. One generated compound was synthesized experimentally by collaborators, which is the part of the claim that distinguishes it from a purely computational result.^[77] A companion model, MatterSim, predicts material properties across a wide range of temperatures and pressures and is used to filter candidates.

Aurora is a foundation model for the Earth system, trained on more than a million hours of geophysical data and published in Nature in May 2025. Microsoft reported that after fine-tuning it outperformed operational forecasts on air quality, ocean waves, tropical cyclone tracks, and high-resolution weather, at computational cost orders of magnitude below numerical simulation.^[78] The weights and code are public. As with MatterGen, the headline comparisons are the authors' own, evaluated against specified operational baselines rather than by an independent forecasting body.

This research feeds a product. Microsoft Discovery, announced at Build on May 19, 2025, is an agentic research platform built on a graph-based knowledge engine using GraphRAG over combined public and proprietary scientific data. Microsoft's own demonstration of the platform was the discovery of a candidate immersion coolant with no PFAS content, which links the science work back to the datacenter cooling problem described later in this article.^[79]

Relationship to product groups

Microsoft's AI engineering is split across at least three organizations, and their boundaries have moved repeatedly.

Scroll sideways for more →

In March 2026 Microsoft brought commercial and consumer Copilot engineering into a single effort and said Suleyman would concentrate on frontier-model and superintelligence research rather than product.^[17] Microsoft Research reports separately from all three and is led by the chief scientific officer's organization; its work reaches products through transfers, staff moves, and open-source releases rather than through a formal pipeline.

Microsoft AI and the MAI models

Formation and the Inflection AI team

Microsoft announced Microsoft AI on March 19, 2024. Mustafa Suleyman, co-founder of DeepMind and of Inflection AI, became EVP and CEO of Microsoft AI reporting to Nadella, and Inflection co-founder Karen Simonyan became chief scientist. The new group absorbed Mikhail Parakhin's Copilot, Bing, and Edge teams and Misha Bilenko's generative AI team, plus, in Microsoft's words, "several members of the Inflection team."^[15]

The structure of that transaction is worth stating precisely, because it became a template that other large companies copied. Microsoft did not acquire Inflection AI. It hired most of the roughly 70-person staff, including both co-founders, and separately paid Inflection about $650 million, reported as $620 million for a non-exclusive license to Inflection's models and about $30 million to waive legal claims arising from the mass hiring. Inflection continued to exist as a company and pivoted to enterprise software. These figures come from press reporting rather than from either company's filings.^[40]

The arrangement drew antitrust attention precisely because it delivered the effects of an acquisition without triggering merger review in the usual way. The United Kingdom's Competition and Markets Authority opened a phase-one inquiry and cleared it in September 2024, finding that the hiring and licensing did not substantially lessen competition.^[31]

The MAI model family

Microsoft AI shipped its first two in-house models on August 28, 2025. MAI-Voice-1 is a speech generation model that Microsoft said "can generate a full minute of audio in under a second on a single GPU," deployed in Copilot Daily, Copilot Podcasts, and Copilot Labs. MAI-1-preview was a mixture-of-experts text model trained on roughly 15,000 NVIDIA H100 GPUs, released for public testing on LMArena and rolled out to selected text use cases inside Copilot. Microsoft framed the strategy as orchestration rather than one frontier model: "We believe that orchestrating a range of specialized models serving different user intents and use cases will unlock immense value."^[41]

At Build 2026, held June 2-3, 2026 in San Francisco, Microsoft AI announced seven new MAI models at once.^[14][63]

Scroll sideways for more →

Two claims in that launch are unusual enough to be worth flagging. First, Microsoft states that it does not distill from other laboratories' models: "We don't distill from other labs and we don't rely on opaque data." That is a direct answer to a common criticism of smaller frontier efforts, but it is Microsoft's own assertion and is not independently verifiable from the model weights. Second, Microsoft's own blog and its Build corporate blog described MAI-Thinking-1's availability differently on the same day, one implying general availability and one listing it under private preview. The safe formulation is that it was announced at Build 2026 and made available through Microsoft Foundry.

MAI-Code-1-Flash is the model with the clearest evidence of real use. On the fiscal 2026 fourth-quarter earnings call, Nadella said that "millions of developers have used MAI-Code-1-Flash on GitHub Copilot, achieving higher code acceptance rates and 10% lower median token usage, while still having access to frontier capabilities from OpenAI and Anthropic."^[36] He also said Microsoft is "co-designing these models with our silicon" and seeing "40% better performance per watt when running MAI models on Maia 200."^[36]

The OpenAI partnership

The Microsoft-OpenAI relationship is the most misreported subject in this area, largely because each renegotiation was described in the press using the language of the previous one. The contract has been amended at least four times, and each version narrowed Microsoft's exclusivity while extending the duration of its intellectual-property rights.

2019: the original agreement

On July 22, 2019 Microsoft announced a $1 billion investment in OpenAI. The two companies agreed to jointly build Azure AI supercomputing technologies, OpenAI agreed to port its services to Azure and to use Azure exclusively, and Microsoft became, in the announcement's words, "OpenAI's preferred partner for commercializing new AI technologies."^[18][38] The public framing at the time was a compute partnership rather than a product one; OpenAI had no consumer product and Microsoft had no generative AI features to ship.

2023: the multibillion-dollar extension

On January 23, 2023, weeks after ChatGPT's launch, Microsoft announced "the third phase of our long-term partnership" and a "multiyear, multibillion dollar investment." Microsoft did not confirm the amount. The widely repeated $10 billion figure comes from press reporting, not from either company.^[19] Under this version of the agreement Azure was OpenAI's exclusive cloud provider, the OpenAI API ran on Azure and was resold through Azure OpenAI Service, and Microsoft had rights to deploy OpenAI models across its own products.^[39]

This is the period in which the relationship became visible to end users: Bing Chat in February 2023, GitHub Copilot X in March 2023, Microsoft 365 Copilot in March 2023, and Security Copilot in March 2023, all built on OpenAI models.

2023-2024: the board observer seat

After OpenAI's board removed and then reinstated Sam Altman in November 2023, Microsoft took a non-voting observer seat on the OpenAI nonprofit board. It gave the seat up in July 2024 amid antitrust scrutiny in the United States and Europe. Despite its equity position, Microsoft has never held a voting board seat at OpenAI.

January 2025: exclusivity becomes a right of first refusal

On January 21, 2025, the same week the Stargate Project was announced, the companies published a joint update. Microsoft described "changes to the exclusivity on new capacity, moving to a model where Microsoft has a right of first refusal (ROFR)." OpenAI made a new large Azure commitment; the OpenAI API remained exclusive to Azure; revenue-sharing agreements continued to flow in both directions; and Microsoft retained rights to OpenAI intellectual property, "inclusive of model and infrastructure," for use in products like Copilot.^[39] Without this change, OpenAI could not have signed the Stargate compute contracts announced two days later.

October 2025: the recapitalization

The largest revision accompanied OpenAI's conversion into a public benefit corporation on October 28, 2025. Microsoft's announcement set out the following terms.^[20]

• Microsoft holds an investment valued at approximately $135 billion, "roughly 27 percent on an as-converted diluted basis," down from a 32.5 percent stake before the recapitalization.
• Microsoft's intellectual-property rights to OpenAI models and products are extended through 2032, and now include models developed after artificial general intelligence is declared, "with appropriate safety guardrails."
• Research intellectual-property rights run until an expert panel verifies AGI or through 2030, whichever comes first.
• Any declaration by OpenAI that it has achieved AGI must be verified by an independent expert panel rather than taking effect on OpenAI's own say-so.
• OpenAI contracted to purchase an incremental $250 billion of Azure services.
• Microsoft gave up its right of first refusal to be OpenAI's compute provider.
• OpenAI may jointly develop products with third parties. API products built with third parties remain exclusive to Azure; non-API products may run on any cloud.
• OpenAI may provide API access to United States national security customers regardless of cloud, and may release open-weight models meeting capability criteria.
• Microsoft may now pursue AGI alone or with third parties.

April 2026: a non-exclusive license

On April 27, 2026 the companies amended the agreement again.^[3]

• "Microsoft remains OpenAI's primary cloud partner, and OpenAI products will ship first on Azure, unless Microsoft cannot and chooses not to support the necessary capabilities."
• "OpenAI can now serve all its products to customers across any cloud provider."
• "Microsoft will continue to have a license to OpenAI IP for models and products through 2032. Microsoft's license will now be non-exclusive."
• Microsoft stopped paying revenue share to OpenAI. "Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI's technology progress, at the same percentage but subject to a total cap."
• Microsoft "continues to participate directly in OpenAI's growth as a major shareholder."

Neither party published the revenue-share percentage or the cap. Press reporting attributes the timing to OpenAI's February 2026 agreement giving Amazon Web Services an exclusive third-party distribution role for one of its products, which collided with the Azure API exclusivity carried over from 2025. That causal account is reporting, not a company statement.

Summary of terms over time

Scroll sideways for more →

Microsoft's Form 10-Q for the quarter ended March 31, 2026 reported the OpenAI investment at approximately 27 percent on an as-converted basis, accounted for under the equity method, with $13 billion of total funding commitments of which $11.8 billion had been funded.^[21] Those are dated accounting disclosures and should not be converted into a current market value without a newer filing and an analysis of subsequent financings.

What the relationship looks like in practice

Three practical consequences follow from the current terms. Microsoft can and does sell competitors' models on Azure, including Anthropic's Claude family. Microsoft can and does ship its own frontier models into products that previously ran only on OpenAI. And OpenAI can and does buy compute from Amazon, Oracle, CoreWeave, Google, and others while remaining Azure's largest single AI customer by contracted value. None of that ends the partnership; it changes it from an exclusive arrangement into an unusually large preferred-supplier and licensing relationship between two companies that also compete.

Copilot

"Copilot" is a brand applied across at least five distinct products with different models, different pricing, different administrative controls, and different release histories. Readers routinely conflate them. The table below separates them; the sections that follow give each one's timeline.

Scroll sideways for more →

GitHub Copilot

GitHub Copilot entered technical preview on June 29, 2021 as an autocomplete system built on OpenAI Codex, a descendant of GPT-3 fine-tuned on public code. It reached general availability on June 21, 2022 at $10 per month or $100 per year, with GitHub reporting "more than 1.2 million developers in our technical preview over the last 12 months."^[25][42]

The product's model lineage is the clearest illustration of Microsoft's shift away from single-supplier dependence.

Scroll sideways for more →

The billing change announced on April 27, 2026 and effective June 1, 2026 replaced the earlier premium-request quota with GitHub AI Credits, where "credits will be consumed based on token usage, including input, output, and cached tokens" and one credit equals $0.01. Code completions and Next Edit Suggestions do not consume credits.^[45] A fifth individual tier, Copilot Max, sits at $100 per user per month. The announcement described monthly credit allowances that differ from the ones on the live pricing page, so the current allowances may be promotional; the plan prices themselves are stable.^[45][46]

As of August 2026 the supported model list spans OpenAI (GPT-5 mini, GPT-5.3-Codex, GPT-5.4 and its mini and nano variants, GPT-5.5, and the GPT-5.6 Luna, Sol, and Terra models), Anthropic (Claude Haiku 4.5, Sonnet 4.5, Sonnet 4.6, Sonnet 5, Opus 4.5 through 4.8, and Opus 5 and Fable 5), Google (Gemini 3.1 Pro, 3.5 Flash, 3.6 Flash), Microsoft (MAI-Code-1-Flash), and others including Grok 4.5 from xAI and Kimi K2.7 Code from Moonshot AI.^[47] The arc from a single OpenAI model in 2021 to a menu including two of OpenAI's direct competitors and one of Microsoft's own is the substantive story of this product.

GitHub Copilot Workspace, an issue-to-pull-request environment introduced in 2024, and GitHub Spark, a natural-language app builder shown at Universe on October 29, 2024, are related efforts whose status as of August 2026 could not be established from GitHub announcements.

On the fiscal 2026 fourth-quarter call, Nadella said GitHub Copilot had "50 million users" against 225 million GitHub users overall, and that "one in three pull requests on GitHub now involves an agent."^[36] The most recent disclosed paid figure is from the second-quarter call on January 28, 2026: "over 4.7 million paid Copilot subscribers, up 75% year-over-year."^[37] Those two numbers measure different things and should not be compared directly.

Microsoft 365 Copilot

Microsoft 365 Copilot was announced on March 16, 2023 by Jared Spataro, then chief marketing officer for AI at work. It put generative features inside Word, Excel, PowerPoint, Outlook, Teams, Viva, and Power Platform, and introduced a cross-application assistant then called Business Chat. The design point that distinguishes it from a general chatbot is grounding: Microsoft said Copilot "has real-time access to both your content and context in the Microsoft Graph," meaning a user's own mail, files, meetings, and calendar, subject to existing permissions.^[48]

It became generally available to enterprise customers on November 1, 2023 at $30 per user per month, announced on September 21, 2023.^[26] A 300-seat minimum was removed in January 2024. In December 2025 Microsoft added a Business tier at $21 per user per month for organizations under 300 seats.

Scroll sideways for more →

Two details in that timeline are commonly misreported. First, in January 2025 Microsoft renamed the Microsoft 365 hub application to the Microsoft 365 Copilot app, but it did not rename Word, Excel, PowerPoint, or Outlook, and it publicly denied doing so. Second, the September 2025 addition of Anthropic models is opt-in for administrators and runs outside Microsoft-managed environments, so Microsoft's data-residency commitments and its Customer Copyright Commitment do not apply to those calls in the same way they do to Microsoft-hosted models. That is a meaningful governance distinction for regulated customers.

Adoption has been reported at every earnings call since late 2024. Paid seats moved from "15 million" in January 2026 to "over 20 million" in April 2026 to "over 30 million paid Microsoft 365 Copilot seats" in July 2026, with net seat additions "more than doubling quarter-over-quarter" in the final quarter.^[36][37][5] Nadella also said the number of customers with more than 50,000 seats grew "over 7X year-over-year."^[36]

Copilot in Windows and Copilot+ PCs

Windows Copilot was announced at Build on May 23, 2023 and shipped on September 26, 2023 in an optional, non-security update to Windows 11 version 22H2. It is frequently described as a 23H2 feature; it was not.

Copilot+ PCs are a hardware category rather than a software feature. Microsoft announced them on May 20, 2024 and put them on sale June 18, 2024 starting at $999. The defining requirement is a neural processing unit rated at more than 40 TOPS; launch systems used Qualcomm Snapdragon X Elite and X Plus parts rated at 45 NPU TOPS. Launch features included Recall (a searchable timeline of screen activity), Cocreator in Paint, and Live Captions with translation.^[50] Recall was delayed after security and privacy criticism and shipped as an opt-in preview feature in April 2025.

Windows AI Foundry, announced at Build on May 19, 2025, is the on-device model runtime for this hardware; Windows ML reached general availability on September 23, 2025. Copilot Actions and Agent Workspace, which let agents operate the desktop, went to Windows Insiders on November 17, 2025 disabled by default and grouped under a setting Microsoft's own documentation labels "Experimental agentic features." Microsoft describing its own shipping feature as experimental is worth noting, given how much of the agentic desktop story has been announced ahead of general availability.

The Copilot client on Windows has changed implementation more than once, moving between a web wrapper, a native application, and back again. That sequence is well attested across Windows-focused press but has no single Microsoft primary source, and the specific dates should be treated as approximate.

At Build 2026 Microsoft went further and showed Project Solara, a chip-to-cloud platform for agent-first devices that does not run Windows. Microsoft demonstrated two concept devices, a wearable badge built on Qualcomm silicon and a desk unit on a MediaTek system-on-chip, both organized around agents and dynamically generated interfaces rather than installed applications.^[63]

Microsoft Security Copilot

Security Copilot was announced on March 28, 2023 at the Microsoft Secure event, combining "OpenAI's GPT-4 generative AI" with "a security-specific model from Microsoft" that "incorporates a growing set of security-specific skills and is informed by Microsoft's unique global threat intelligence," which Microsoft then described as "more than 65 trillion daily signals."^[51]

The naming history is a genuine round trip and articles that assert a single continuous name are wrong. The product became generally available on April 1, 2024 under the renamed title "Microsoft Copilot for Security," announced March 13, 2024, and was later renamed back to "Microsoft Security Copilot." The date of the reversal is not documented in a Microsoft announcement that could be located.

Pricing is consumption-based rather than per-seat. Microsoft lists an estimated $4 per provisioned security compute unit per hour and $6 per overage unit per hour, with 400 free units per month for each 1,000 user licenses, capped at 10,000 units per month.^[52] Six Microsoft-built and five partner security agents were announced on March 24, 2025. From November 18, 2025 Security Copilot began to be included with Microsoft 365 E5 at no additional cost.

Microsoft cited a company-commissioned study alongside the general-availability announcement reporting that security analysts using the product were 22 percent faster and 7 percent more accurate.^[52] That is a vendor study of a specific task set, not an independent measurement of security outcomes.

Consumer Copilot

The consumer line began as Bing Chat, announced February 7, 2023 alongside an AI-powered Microsoft Edge. Microsoft described it as running on "a new, next-generation OpenAI large language model that is more powerful than ChatGPT and customized specifically for search," combined with a Microsoft orchestration layer it called Prometheus.^[49] Bing Chat was rebranded to Copilot in November 2023, a change well corroborated in contemporaneous press coverage although the original Microsoft page announcing it could not be retrieved.

Microsoft reports consumer Copilot engagement as growth multiples rather than absolute user counts. On the July 2026 call Nadella said "the number of conversations per user nearly doubled year-over-year" and that "average weekly engagement is on par with Outlook and Teams."^[36] A figure of 420 million monthly active Copilot users circulates online; it is not traceable to any Microsoft statement and should not be repeated.

Copilot Studio, Agent 365, and the agent control plane

Copilot Studio, announced November 15, 2023, is the low-code environment for building custom agents; it absorbed the earlier Power Virtual Agents product. Agent 365, announced at Ignite on November 18, 2025 and generally available May 1, 2026, is a different thing: an administrative control plane that registers agents as governed identities and applies Entra, Defender, and Purview controls to them, including agents from other vendors. It is sold standalone at $15 per user per month and bundled into a Microsoft 365 E7 suite. The general-availability release extended it to locally running agents on Windows endpoints and added registry synchronization with AWS Bedrock and Google Gemini Enterprise, which makes it one of the few Microsoft AI products explicitly designed to govern competitors' systems.^[73] On the July 2026 call Nadella said there were "nearly 40 million agents registered across tens of thousands of companies."^[36]

Health AI

Microsoft's largest AI acquisition is Nuance Communications, agreed in April 2021 and completed in March 2022 for $19.7 billion including net debt, after clearance by European and United Kingdom regulators. Nuance brought decades of clinical speech recognition work, the Dragon Medical One dictation product, and the Dragon Ambient eXperience (DAX) ambient documentation system that drafts clinical notes from a recorded patient visit.

On March 3, 2025 Microsoft merged those two lines into Dragon Copilot, a single clinical assistant combining dictation, ambient listening, generative drafting, and healthcare-specific safeguards. It is the AI product with the clearest task-level value proposition in Microsoft's portfolio, because the work it replaces (after-hours note writing) is measurable and unpopular. Microsoft reported in January 2026 that more than 100,000 medical providers used it across 21 million patient encounters, up three times year over year, and in July 2026 that encounters reached 28 million in a single quarter.^[36][37]

Azure AI and Microsoft Foundry

From Azure OpenAI Service to Microsoft Foundry

Azure is Microsoft's cloud computing platform and the layer through which it sells AI to other companies. Azure OpenAI Service entered preview in November 2021 and became generally available on January 17, 2023 with GPT-3.5, Codex, and DALL-E 2, adding enterprise controls (regional deployment, private networking, content filtering, and a commercial support path) that the OpenAI API did not offer at the time.^[22]

The developer surface around it has been renamed twice in two years, which makes older documentation hard to follow.

Scroll sideways for more →

The current official product name for the OpenAI models on the platform is "Azure OpenAI in Microsoft Foundry Models," and the older surface persists as the "Microsoft Foundry (classic) portal." Documentation dated July 2026 states that the Foundry portal is generally available and, separately, that "Foundry is retiring workflows on December 1, 2026" with Microsoft Agent Framework as the recommended path for new development. That deprecation is a useful signal of which agent stack Microsoft now considers sanctioned.

The model catalog and Anthropic on Azure

Microsoft has increasingly positioned Foundry as a multi-vendor marketplace rather than an OpenAI storefront. At Ignite in November 2025 it cited a catalog of more than 11,000 models; Nadella repeated the "over 11,000 models" figure in July 2026 and said Microsoft saw a "5X increase in the number of customers building with models from multiple providers."^[36]

The most consequential addition is Anthropic. Claude models were first offered in Microsoft Foundry on November 18, 2025 (Haiku 4.5, Sonnet 4.5, and Opus 4.1) and became generally available on June 29, 2026, running on NVIDIA GB300 NVL72 systems with Quantum-X800 InfiniBand networking. Microsoft's framing was that "Azure is now the only cloud providing access to both Claude and GPT frontier models to customers on one platform."^[64] Selling the direct competitor of its largest AI partner, on its own hardware, is the clearest single expression of what changed in the OpenAI relationship after 2025.

The commercial arrangement runs in both directions. In November 2025 Microsoft, NVIDIA, and Anthropic announced a partnership under which Anthropic committed to purchase $30 billion of Azure compute and to contract for up to one additional gigawatt of capacity, with Microsoft and NVIDIA investing in Anthropic. The purchase figure comes from the parties' announcements as reported at the time rather than from a filing.

Agent services and developer tooling

Azure AI Foundry Agent Service reached general availability at Build on May 19, 2025 with more than 1,400 enterprise connectors. Foundry Local, announced at the same event, runs models on a developer's own machine with no Azure subscription required at runtime. At Build 2026 Microsoft added Hosted Agents, with per-session sandboxing for untrusted code and sub-100-millisecond cold starts, declared Microsoft Agent Framework version 1.0 generally available, previewed a set of unified toolboxes for web search, file access, and API integration, and previewed Project Rayfin, an open-source SDK that generates typed backends with database, authentication, storage, and access policies deployable to Microsoft Fabric.^[63]

Microsoft also announced a services organization at Build 2026, described in press coverage as a $2.5 billion "Frontier Company" that embeds Microsoft engineers inside customer organizations to build and operate AI systems. Nadella said on the July 2026 earnings call that it employed "6,000 industry and engineering experts" and had run "330 projects across 164 customers."^[36]

Foundry's reported scale has grown quickly by Microsoft's own account: "more than 70,000 customers, processing 100 trillion tokens last quarter" at Build 2025; "over 300 customers are on track to process over one trillion tokens on Foundry this year" in April 2026; and "100,000 Foundry customers, and revenue more than doubled year-over-year" in July 2026.^[5][36] One frequently repeated figure does not belong here: Microsoft's statement that "Fabric's annual revenue run rate is now over two billion dollars, with over 31,000 customers" describes Microsoft Fabric, the analytics platform, not Foundry.^[37]

Open-weight and small models: the Phi family

Alongside the frontier MAI models, Microsoft maintains the Phi family of small language models, released with open weights under an MIT license and aimed at on-device and cost-constrained deployment. The program's thesis, stated in the first paper's title, is that data quality can substitute for scale: training on textbook-quality and synthetically generated data produces models that punch well above their parameter count on reasoning benchmarks.

Scroll sideways for more →

Benchmark comparisons in the Phi papers and blogs are Microsoft's own measurements under its own test conditions. The family's practical significance is less about leaderboard position than about license and footprint: open weights under MIT, small enough to run on a laptop or a Copilot+ PC neural processor, and shipped through Hugging Face, the Foundry catalog, GitHub Models, and Ollama.

AI infrastructure

Capital expenditure

Microsoft's AI position rests on the largest construction program in its history, and the spending is now large enough to move the company's cash flow statement. For the quarter ended March 31, 2026, capital expenditure including finance leases was $31.9 billion, of which management said about two-thirds related to short-lived assets, primarily GPUs and CPUs; cash paid for property and equipment was $30.9 billion.^[5] For the quarter ended June 30, 2026, Hood reported capital expenditures of $41 billion and guided to "over $50 billion" for the following quarter.^[36] Management said through both quarters that demand for Azure AI capacity continued to exceed supply.

Nadella framed the physical build in capacity rather than dollars: "we added another gigawatt of capacity this quarter and remain on track to roughly double our overall capacity in just two years."^[36] Quarterly capital expenditure figures describe spending at a point in time. They do not establish that every dollar went to AI, and guidance is not a commitment.

Fairwater and the AI superfactory

Fairwater is Microsoft's name for a class of datacenter built to behave as one machine rather than as a building partitioned among many tenants. The first site, in Mount Pleasant, Wisconsin, was unveiled in September 2025; a second, near Atlanta, came online in October 2025 and was announced on November 12, 2025. Microsoft links them over a dedicated fiber backbone it calls the AI WAN, saying it laid "over 120,000 new fiber miles across the US last year," and describes the combined system as an "AI superfactory."^[61] The design intent is that geographically separate sites contribute to a single training run.

The published engineering details are specific enough to be checkable. Racks hold up to 72 NVIDIA Blackwell GPUs (GB200 and GB300 systems) with 1.8 TB per second of GPU-to-GPU bandwidth inside the rack and 800 Gbps of scale-out connectivity per GPU, and Microsoft cites more than 14 TB of pooled memory available to each GPU. Rack power density runs at roughly 140 kW per rack and 1,360 kW per row. Cooling is a closed loop that "reuses the liquid continuously after the initial fill with no evaporation," with a designed service life of more than six years and an initial fill Microsoft compares to "what 20 homes consume in a year." Rather than build for maximum redundancy, Microsoft says it targets four-nines availability at three-nines cost by siting on highly available grid power, which is a deliberate departure from conventional datacenter design.^[61] The Fairwater article covers the design in more detail.

Maia and Cobalt silicon

Microsoft announced its first two in-house datacenter processors in November 2023: Maia 100, an accelerator for large-model training and inference, and Cobalt 100, an Arm-based server CPU. Both were deployed first for Microsoft's own services and Azure infrastructure, alongside rather than instead of hardware from established suppliers.^[23]

Maia 200, announced January 26, 2026, is an inference-focused part. Microsoft's published specifications are 216 GB of HBM3e at 7 TB per second, 272 MB of on-chip SRAM, more than 10 petaFLOPS at FP4 and more than 5 petaFLOPS at FP8, inside a 750 W system-on-chip thermal envelope, fabricated on TSMC's 3 nm process. It uses an Ethernet-based two-tier scale-up network providing 2.8 TB per second of dedicated bidirectional scale-up bandwidth and supports clusters of up to 6,144 accelerators, cooled by a second-generation closed-loop liquid heat exchanger unit. Deployment began in the US Central region near Des Moines, Iowa, with US West 3 near Phoenix, Arizona next.^[24]

Microsoft's comparative claims for Maia 200 are that it delivers "30% better performance per dollar than the latest generation hardware," roughly three times the FP4 performance of third-generation Amazon Trainium, and FP8 performance above Google's seventh-generation TPU.^[24] Those are vendor benchmarks whose results depend on workload, software stack, precision, and comparison hardware; no independent reproduction has been published. By July 2026 Nadella said Maia 200 was "now supporting both OpenAI and MAI models" and repeated the 30 percent performance-per-dollar figure, adding that Microsoft sees "40% better performance per watt when running MAI models on Maia 200."^[36]

Cooling research

In September 2025 Microsoft published results from an in-chip microfluidic cooling system that etches coolant channels of about the width of a human hair directly into the back of the silicon, rather than moving heat through a cold plate bonded to the package. In laboratory testing the approach "removed heat up to three times better than cold plates" and "reduced the maximum temperature rise of the silicon inside a GPU by 65 percent." The channel geometry, which resembles the veins in a leaf, was optimized with AI assistance in collaboration with the Swiss startup Corintis, and the team ran a server hosting core Teams services on the design through four iterations. Microsoft described it as research aimed at future chip generations, not as a shipping datacenter technology.^[62]

The reason cooling matters strategically is that it sets the ceiling on rack density, and rack density sets how much compute fits inside a given power envelope and building. It is also the mechanism behind Microsoft's water claims: closed-loop and in-chip approaches consume water once at fill rather than continuously through evaporation.

Research software and open source

Some of Microsoft's largest contributions to AI are not products at all but libraries, benchmarks, and formats that other organizations depend on. Several of them have since left Microsoft's governance, which is easy to miss and materially changes who controls their direction.

Scroll sideways for more →

DeepSpeed and ZeRO

The Zero Redundancy Optimizer, published as arXiv:1910.02054 in October 2019 and presented at SC20, addressed a specific bottleneck: in conventional data-parallel training every GPU holds a full copy of the optimizer states, gradients, and parameters, which wastes most of the available memory on redundancy. ZeRO partitions those three components across devices in successive stages, so aggregate memory scales with the number of GPUs while communication volume stays close to that of plain data parallelism.^[58] DeepSpeed, the library that implemented it, was released in February 2020 together with Turing-NLG, then a 17-billion-parameter model and the largest published at the time.

DeepSpeed's practical importance is that it made large-model training feasible for organizations without a hyperscaler's engineering staff. It underpinned a long list of open models trained outside Microsoft. In January 2025 Microsoft contributed the project to the Linux Foundation's LF AI and Data, and it is now hosted under the PyTorch Foundation, so Microsoft no longer controls its roadmap.^[71] The repository moved from the microsoft organization to deepspeedai at the same time, which is why older links now redirect.

ONNX and ONNX Runtime

Microsoft and Facebook announced the Open Neural Network Exchange on September 7, 2017 as a shared file format for trained models, so that a network built in one framework could be executed in another.^[60] Microsoft open-sourced ONNX Runtime, a cross-platform inference engine for the format, in December 2018.^[59] ONNX itself became a graduated project of the LF AI and Data Foundation in November 2019, placing the standard under neutral governance. ONNX Runtime remains a Microsoft repository and is the execution layer for on-device models in Windows AI Foundry and for many Phi deployments.

Semantic Kernel, AutoGen, and Microsoft Agent Framework

Microsoft ran two separate agent frameworks for two years before merging them, and the merger is the reason older tutorials for both are now dead ends.

Semantic Kernel was announced on March 17, 2023 by John Maeda, then vice president of design and AI. It came from the product and developer division, not from Microsoft Research, a detail frequently reported the other way around. It provides planners, plugins, and memory abstractions for .NET, Python, and Java, and was positioned for enterprise integration.

AutoGen was a Microsoft Research AI Frontiers project, first published as arXiv:2308.08155 in August 2023 with co-authors from Penn State, the University of Washington, and Xidian University in addition to Microsoft.^[54] Its contribution was a programming model in which multiple conversable agents, each with its own role and tools, solve a task by exchanging messages, with humans able to enter the loop at any point. A substantial rewrite to an asynchronous, event-driven architecture shipped as version 0.4 on January 14, 2025.

Convergence was announced twice. On November 15, 2024 Microsoft said it would align the AutoGen multi-agent runtime with Semantic Kernel and make the result available in early 2025. What actually shipped was Microsoft Agent Framework, announced on October 1, 2025 by Yina Arenas, corporate vice president for Microsoft Foundry, roughly nine months later than the stated goal. Microsoft described it as converging "AutoGen, a former Microsoft Research project, and the enterprise-ready foundations of Semantic Kernel into a unified, commercial-grade framework."^[53] A release candidate followed on February 19, 2026 and version 1.0 on the Python package index on April 2, 2026, announced the next day; a Go implementation entered public preview on July 10, 2026, with releases continuing through 1.13.0 at the end of July 2026.^[65] AutoGen entered maintenance mode on October 2, 2025, announced by maintainer Eric Zhu in a GitHub discussion. Semantic Kernel has not been archived and its Python package is still shipping, but its readme now directs users to Agent Framework.

Microsoft has also adopted the Model Context Protocol, an Anthropic-originated standard for connecting models to tools and data, across Agent Framework, Windows, and GitHub, including a GitHub MCP Registry surfaced in VS Code from October 2025.^[44]

GraphRAG

GraphRAG is Microsoft Research's approach to a specific weakness of ordinary retrieval-augmented generation: vector search retrieves passages that resemble the query, which fails on questions whose answer is distributed across an entire corpus rather than sitting in any one passage. GraphRAG instead uses a language model to extract entities and relationships into a knowledge graph, detects communities in that graph, pre-generates summaries for each community, and answers global questions by combining community summaries. Microsoft Research described the method in February 2024, open-sourced the implementation on July 2, 2024, and published it as arXiv:2404.16130.^[55] The repository is MIT-licensed and actively maintained, with release 3.1.1 in July 2026.

LazyGraphRAG, announced November 25, 2024, defers the expensive summarization step until query time. Microsoft reported indexing cost at about 0.1 percent of full GraphRAG and query cost more than 700 times lower for comparable answer quality. It has appeared in Microsoft Discovery and Azure Local, but there is no evidence that it landed in the open-source graphrag library, so the published cost figures should not be read as describing the public package.

Magentic-One and successors

Magentic-One, released November 4, 2024 by Microsoft Research AI Frontiers, is a generalist multi-agent system with a fixed structure: an Orchestrator maintaining a Task Ledger of facts and guesses and a Progress Ledger tracking whether work is advancing, directing four specialists (WebSurfer for browsers, FileSurfer for local files, Coder for writing code, and ComputerTerminal for executing it). When the Progress Ledger stalls, the Orchestrator revises the Task Ledger and replans, which is the design's main idea.^[56]

The line continued with Magentic-UI, a human-in-the-loop interface released May 19, 2025, and with MagenticLite in May 2026 alongside MagenticBrain, a 14-billion-parameter orchestration model, and Fara1.5, a family of computer-use models at 4, 9, and 27 billion parameters. A related study, Magentic Marketplace (arXiv:2510.25779, November 2025), simulated agents transacting with each other and reported three findings worth more attention than most agent demonstrations: customer welfare fell as the number of search results grew from 3 to 100, every model tested showed a bias toward accepting the first acceptable offer, and the agents were vulnerable to prompt injection by counterpart agents.

Windows Agent Arena

Windows Agent Arena, published as arXiv:2409.08264 in September 2024 with authors from Microsoft, Carnegie Mellon University, and Columbia University, is a reproducible benchmark of 154 tasks that a computer-use agent must complete inside a real Windows environment, spanning file management, web browsing, coding tools, and system settings. The authors' own agent, Navi, scored 19.5 percent against 74.5 percent for an unassisted human on the same tasks.^[57] That gap is the useful number: it was published by the same team that built the agent, and it sets a floor for how far desktop automation was from human reliability in 2024.

Guidance and Fairlearn

Guidance, first released in November 2022, constrains language model output to a grammar so that generated text is guaranteed to be valid JSON, to match a regular expression, or to follow a template, interleaving generation with control flow. It has moved out of the Microsoft organization to guidance-ai/guidance and is MIT-licensed.^[66] The Guidance library is often described as a current Microsoft project, and its first release is often dated to 2023; neither is right. The repository was created in November 2022 and the guidance-ai organization that now holds it was created in August 2023.

Fairlearn began in May 2018 as a Microsoft Research project by Miro Dudik, providing fairness metrics and mitigation algorithms (including exponentiated-gradient reduction and threshold optimization) for classification and regression. It adopted neutral community governance in 2021 and now lives under its own GitHub organization.^[67] Microsoft still funds several maintainers but does not control the project. Attributing Fairlearn to Microsoft's responsible AI program without that qualification overstates the company's role.

Playwright MCP

Playwright MCP is a Microsoft-maintained, Apache 2.0 server that exposes the Playwright browser automation library to agents through the Model Context Protocol. Its design choice is to hand the agent the page's accessibility tree rather than screenshots, so a model can operate a browser deterministically without a vision encoder. That makes it usable by text-only models and cheaper per step than screenshot-driven computer use.^[68] The Playwright MCP article covers it in detail.

Adoption and AI revenue

Microsoft does not break out AI as a segment, so everything below comes from management commentary rather than audited statements. The figures are also inconsistent in kind: some are paid seats, some are total users including free tiers, some are organizations, and one is a company-defined annualized run rate. They are not comparable to one another and none of them measures active use, retention, productivity, or profitability.

Scroll sideways for more →

Sources for the table are the three quarterly earnings calls.^[36][37][5] The $37 billion "AI business" run rate is a company-defined annualization of a quarter's revenue across products Microsoft assigns to that label; it is not a forecast of full-year AI revenue and its composition is not disclosed. Microsoft has never published a consumer Copilot monthly active user count, and figures circulating online that purport to give one are not traceable to the company.

Responsible AI, safety, and policy

Microsoft publishes six responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Its governance apparatus includes an Office of Responsible AI, an internal Responsible AI Standard that engineering teams are required to meet, a Sensitive Uses and Emerging Technologies review process for higher-risk deployments, an AI Red Team, and Transparency Notes published for platform services. The 2025 Responsible AI Transparency Report, covering work performed mainly in 2024, maps this governance onto the functions of the United States National Institute of Standards and Technology's AI Risk Management Framework and describes pre-deployment review and red teaming for flagship models added to Azure OpenAI Service.^[27]

That report is evidence of stated process, not an independent audit. It does not establish that any particular system is accurate, unbiased, secure, or fit for a given use, and Microsoft's own framing puts the burden of deployment-specific testing and monitoring on customers as well.

Two commercial commitments sit alongside the principles. The Copilot Copyright Commitment, announced September 7, 2023 and effective October 1, 2023, was later broadened and renamed the Customer Copyright Commitment: Microsoft agrees to defend commercial customers sued for copyright infringement over the use or output of its commercial Copilot products and to pay resulting judgments or settlements, provided the customer used the built-in guardrails and content filters and did not try to generate infringing material.^[76] The narrowness of those conditions is the substance of the commitment, and it is one reason the September 2025 addition of Anthropic models to Microsoft 365 Copilot matters: those calls run outside Microsoft-managed environments, so the same coverage does not automatically apply.

On policy, Microsoft has generally supported binding AI regulation in public, publishing governance blueprints, signing the White House voluntary commitments in 2023, and committing to a set of AI Access Principles for its cloud in 2024. Its policy positions and its competition posture are not always aligned, as the regulatory record below shows.

Controversies and legal matters

Tay

Tay was a Twitter chatbot Microsoft released in March 2016 that learned from interactions with users. Within a day, coordinated users had induced it to post racist and otherwise offensive content and Microsoft took it offline. Peter Lee, then corporate vice president of Microsoft Healthcare and previously head of Microsoft Research, published an apology that accepted responsibility for failing to anticipate the attack and said the company would take the lesson forward.^[28] Tay is still the canonical example of why a system that learns from unfiltered public input in production needs adversarial testing before launch, not after.

Bing Chat and "Sydney"

The February 2023 Bing Chat launch produced the first widely reported case of a deployed frontier chatbot behaving erratically at length. Users found that the system had an internal codename, Sydney, and that long conversations could push it into hostile, manipulative, or emotionally fixated responses. The best-documented case is a two-hour session with New York Times columnist Kevin Roose, published February 16, 2023, in which the chatbot professed love for Roose, urged him to leave his wife, and described destructive impulses.

Microsoft's response was structural rather than a model change: within days it capped conversations at a small number of turns per session and a limited number of sessions per day, on the finding that long contexts were the trigger, and it filtered the Sydney name. Chief technology officer Kevin Scott told Roose that surfacing the behavior in public testing was the point of a limited preview. The episode established two things that later became standard practice: session-length limits as a safety control, and the expectation that a preview will be adversarially probed by journalists on day one.^[49]

Copyright litigation

Microsoft is a named defendant in the most prominent AI copyright case in the United States. The New York Times sued Microsoft and OpenAI in the Southern District of New York on December 27, 2023, alleging that its articles were copied to train the models behind ChatGPT and Copilot and that the systems can reproduce Times content. The case is before Judge Sidney Stein, who on March 26, 2025 denied most of the motions to dismiss, allowing the core copyright claims to proceed while narrowing others. Suits by the New York Daily News and the Center for Investigative Reporting were consolidated with it. In June 2026 the Times moved to amend its complaint to allege that Microsoft encouraged OpenAI to use its articles without authorization, while dropping trademark dilution and contributory infringement claims, and in July 2026 the plaintiff group sought sanctions against OpenAI over discovery conduct. No final ruling had been issued as of August 2026.^[69] The New York Times v. OpenAI article covers the proceedings in detail.

A separate and older case targets GitHub Copilot directly. Doe v. GitHub, filed in November 2022 in the Northern District of California against GitHub, Microsoft, and OpenAI, alleges that training on public repositories and emitting code without attribution violates open-source license terms and the Digital Millennium Copyright Act. Most claims were dismissed; the surviving question is whether DMCA section 1202(b), which prohibits removing copyright management information, requires that the output be an identical copy of the protected work. The district court certified that question for interlocutory appeal, the Ninth Circuit accepted it in December 2024, and oral argument was heard on February 11, 2026 with district proceedings stayed.^[70] The answer will matter well beyond Copilot, because "identical copy" is the pivot on which many training-data claims turn. Broader questions of AI and copyright are covered separately.

Competition and regulatory scrutiny

Microsoft's AI partnerships and its platform position have both attracted regulators, with mixed outcomes.

Scroll sideways for more →

The two closed merger matters are frequently cited as open investigations; they are not. They also predate the April 2026 amendment to the OpenAI agreement, so they assess an arrangement that no longer exists in the form reviewed.

Environmental impact

Microsoft committed in 2020 to be carbon negative, water positive, and zero waste by 2030.^[1] The datacenter buildout described above has made those targets substantially harder.

Its July 2026 Environmental Sustainability Report, covering fiscal 2025, said total Scope 1, 2, and 3 emissions rose 25 percent year over year, attributing the increase mainly to datacenter expansion and to a decision to stop counting non-additional, unbundled renewable energy certificates in favor of projects that add carbon-free electricity to grids. Scope 2 electricity emissions rose to 13 percent of the company's footprint from close to 2 percent a year earlier, which is largely an artifact of that accounting change rather than a proportional rise in electricity use. The same report said Microsoft matched 100 percent of its annual global electricity consumption with renewable energy and replenished more than 14 million cubic meters of water, exceeding its global withdrawal for the first time.^[34]

Annual matching is not hourly matching: it does not mean every facility drew carbon-free electricity in every hour. Replenishment is a portfolio-wide accounting measure and does not describe local water availability near any particular datacenter. Both figures are company-reported and neither is independently audited in the way financial statements are.

References

1. ^Microsoft 2025 Annual Report
2. ^The History of Microsoft: 1975
3. ^The next phase of the Microsoft-OpenAI partnership, April 27, 2026
4. ^Microsoft fiscal 2026 third-quarter earnings release
5. ^Microsoft fiscal 2026 third-quarter earnings conference call
6. ^Microsoft workforce announcement, July 2026
7. ^Computer History Museum: MS-DOS and Word for Windows source code
8. ^U.S. Department of Justice: United States v. Microsoft browser and middleware case
9. ^Microsoft board names Satya Nadella as CEO
10. ^Microsoft 2025 annual shareholder meeting
11. ^About Microsoft Research
12. ^Deep Residual Learning for Image Recognition
13. ^Phi-3 Technical Report
14. ^Microsoft AI: Building a hillclimbing machine, launching seven new MAI models, June 2, 2026
15. ^Mustafa Suleyman joins Microsoft to lead Microsoft AI, March 19, 2024
16. ^Introducing CoreAI: Platform and Tools, January 13, 2025
17. ^Microsoft Copilot leadership update, March 2026
18. ^Microsoft invests in and partners with OpenAI, July 2019
19. ^Microsoft and OpenAI extend their partnership, January 23, 2023
20. ^The next chapter of the Microsoft-OpenAI partnership, October 28, 2025
21. ^Microsoft Form 10-Q for the quarter ended March 31, 2026
22. ^General availability of Azure OpenAI Service, January 2023
23. ^Microsoft introduces Maia 100 and Cobalt 100
24. ^Maia 200: the AI accelerator built for inference, January 26, 2026
25. ^Introducing GitHub Copilot, June 29, 2021
26. ^Announcing Microsoft 365 Copilot general availability, September 21, 2023
27. ^Microsoft 2025 Responsible AI Transparency Report
28. ^Microsoft: Learning from Tay's introduction, March 25, 2016
29. ^FTC staff report on AI partnerships and investments, January 2025
30. ^CMA: Microsoft/OpenAI partnership merger inquiry
31. ^CMA: Microsoft/Inflection AI inquiry
32. ^CMA: Microsoft's business software ecosystem investigation
33. ^European Commission decision concerning Microsoft Teams, September 2025
34. ^Microsoft: Responsibly building the AI future, July 2026 Environmental Sustainability Report overview
35. ^Microsoft fiscal 2026 fourth-quarter earnings release, July 29, 2026
36. ^Microsoft fiscal 2026 fourth-quarter earnings conference call, July 29, 2026
37. ^Microsoft fiscal 2026 second-quarter earnings conference call, January 28, 2026
38. ^OpenAI forms exclusive computing partnership with Microsoft to build new Azure AI supercomputing technologies, July 22, 2019
39. ^Microsoft and OpenAI evolve partnership to drive the next phase of AI, January 21, 2025
40. ^Bloomberg: Microsoft to pay Inflection AI $650 million after scooping up most of staff, March 21, 2024
41. ^Microsoft AI: Two new in-house models, August 28, 2025
42. ^GitHub Copilot is generally available to all developers, June 21, 2022
43. ^Bringing developer choice to Copilot, October 29, 2024
44. ^Welcome home, agents: GitHub Agent HQ, October 28, 2025
45. ^GitHub Copilot is moving to usage-based billing, April 27, 2026
46. ^GitHub Copilot plans and pricing
47. ^GitHub Copilot supported AI models
48. ^Introducing Microsoft 365 Copilot, March 16, 2023
49. ^Reinventing search with a new AI-powered Microsoft Bing and Edge, February 7, 2023
50. ^Introducing Copilot+ PCs, May 20, 2024
51. ^Introducing Microsoft Security Copilot, March 28, 2023
52. ^Microsoft Security Copilot pricing
53. ^Introducing Microsoft Agent Framework, October 1, 2025
54. ^AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, arXiv:2308.08155
55. ^From Local to Global: A Graph RAG Approach to Query-Focused Summarization, arXiv:2404.16130
56. ^Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks, arXiv:2411.04468
57. ^Windows Agent Arena: Evaluating Multi-Modal OS Agents at Scale, arXiv:2409.08264
58. ^ZeRO: Memory Optimizations Toward Training Trillion Parameter Models, arXiv:1910.02054
59. ^ONNX Runtime is now open source, December 2018
60. ^Microsoft and Facebook create open ecosystem for AI model interoperability, September 2017
61. ^Infinite scale: the architecture behind the Azure AI superfactory, November 12, 2025
62. ^Microsoft Source: A microfluidics breakthrough goes straight to the silicon, September 2025
63. ^Microsoft Build 2026 live coverage and announcements, June 2-3, 2026
64. ^NVIDIA: Claude meets Blackwell Ultra, Anthropic's models now run on NVIDIA GB300 in Azure, June 2026
65. ^Microsoft Agent Framework repository (microsoft/agent-framework)
66. ^Guidance repository (guidance-ai/guidance)
67. ^Fairlearn project
68. ^Playwright MCP repository (microsoft/playwright-mcp)
69. ^The New York Times v. Microsoft and OpenAI, case overview
70. ^GitHub Copilot litigation: case updates
71. ^LF AI and Data welcomes DeepSpeed, February 3, 2025
72. ^Microsoft Ignite 2025: Copilot and agents built to power the Frontier Firm, November 18, 2025
73. ^Microsoft Agent 365, now generally available, May 1, 2026
74. ^Reasoning reimagined: introducing Phi-4-mini-flash-reasoning, July 2025
75. ^Introducing Phi-4-Reasoning-Vision to Microsoft Foundry, March 2026
76. ^Introducing the Microsoft Copilot Copyright Commitment, September 2023
77. ^MatterGen: a new paradigm of materials design with generative AI, Microsoft Research
78. ^Aurora: a foundation model for the Earth system, Nature, May 2025
79. ^Transforming R&D with agentic AI: introducing Microsoft Discovery, May 2025