MICROSOFT

Institutional Archetype

THE LANDLORD OF THE FRONTIER — Microsoft is the frontier company that holds a stake in or sells compute to several of the other frontier labs. It owns roughly 27 percent of OpenAI, booked a $3.2 billion gain on its Anthropic investment in a single quarter, signed a multibillion-dollar capacity agreement with Mistral AI, offers xAI’s model on its platform, and in 2026 shipped a reasoning model of its own whose technical paper puts dangerous-capability evaluations “out of scope for this release.” Its president is also the industry’s most consistent voice for binding rules: a licensing regime for frontier developers in 2023, and in October 2026 “Right now, your washing machine is subject to more safety rules than AI.” Its chief executive is not among the signatories NPR names to the White House accord. The position is a landlord who asks for the building code while while selling compute to two of the developers the code would reach. This profile scores that position, not the people in it.

Mandate & Origin

The company and its officers, in its own words. “Satya Nadella is Chairman and Chief Executive Officer of Microsoft,” and “As Microsoft’s vice chair and president, Brad Smith is responsible for spearheading the company’s work and representing it publicly on a wide variety of critical issues involving the intersection of technology and society, including artificial intelligence, cybersecurity” (Microsoft, Satya Nadella; Microsoft, Brad Smith). “Kevin Scott is Microsoft’s chief technology officer” (Microsoft, Kevin Scott). Natasha Crampton signs the responsible-AI reports as “Chief Responsible AI Officer” (Microsoft On the Issues, 1 Sep 2026). See Satya Nadella and Brad Smith.

Three layers of AI. Microsoft is a platform (Azure, Foundry, Copilot), a stakeholder in other labs, and since 2025 a lab. On 6 November 2025 Mustafa Suleyman wrote: “we have formed the MAI Superintelligence Team, led by me as part of Microsoft AI,” and “Creating superintelligence is one thing; but creating provable, robust containment and alignment alongside it is the urgent challenge facing humanity in the 21st century” (Microsoft AI, Towards Humanist Superintelligence). The release was made possible by the October 2025 agreement with OpenAI: “Microsoft can now independently pursue AGI alone or in partnership with third parties” (Microsoft, 28 Oct 2025).

The OpenAI terms, from Microsoft’s side. “Microsoft holds an investment in OpenAI Group PBC valued at approximately $135 billion, representing roughly 27 percent on an as-converted diluted basis”; “Microsoft’s IP rights for both models and products are extended through 2032 and now include models post-AGI”; “Once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel”; and “OpenAI has contracted to purchase an incremental $250B of Azure services, and Microsoft will no longer have a right of first refusal to be OpenAI’s compute provider” (Microsoft, 28 Oct 2025). See OpenAI.

Funding & Backers

The AI business is funded by the company’s own results. Fourth-quarter fiscal 2026: “Revenue was $90.0 billion and increased 18%,” “Operating income was $40.6 billion and increased 18%,” “Microsoft Cloud revenue was $59.3 billion and increased 27%,” and “commercial remaining performance obligation increased 84% to $678 billion.” The quarter included “a $3.2 billion gain from our investment in Anthropic,” and Nadella said “Azure revenue surpassed $100 billion for the first time, and Microsoft 365 Copilot reached over 30 million paid seats” (Microsoft, FY2026 Q4 results, 29 Jul 2026).

Who else pays it. Anthropic: “Anthropic has committed to purchase $30 billion of Azure compute capacity,” and “NVIDIA and Microsoft are committing to invest up to $10 billion and up to $5 billion respectively in Anthropic” (Microsoft, 18 Nov 2025). Mistral: “This represents a multibillion dollar commitment from Microsoft,” for capacity Microsoft will use “to support the delivery of MSFT’s cloud and AI services” (Microsoft Source, 21 Jul 2026). See Anthropic and Mistral AI. xAI’s Grok 4.7 was announced live on Microsoft Foundry on 7 October 2026 (SpaceXSI on X); see xAI.

Lobbying on the record. Microsoft’s in-house filings report $2.39 million for Q1, $2.69 million for Q2 and $2.44 million for Q3 2026; the issue lists include “Artificial Intelligence: Focused on DoD policy and adoption of AI” and the bills “S. 3982, AI Fraud Accountability Act,” “S. 1367, NO FAKES Act” and “S. 4098, AI Ready Data Act”; 17, 18 and 18 outside registrants filed for Microsoft in the same quarters (LDA, Q1 2026; LDA, Q2 2026; LDA, Q3 2026).

The money it sends the evaluator and safety field. The Frontier Model Forum’s AI Safety Fund lists “founding members of the Frontier Model Forum, including Anthropic, Google, Microsoft, and OpenAI” among its supporters (Frontier Model Forum, AI Safety Fund); Microsoft’s own report says it “has also supported research addressing the biological risks of advanced AI systems through grantmaking efforts led by the Frontier Model Forum” (Microsoft, 2026 Responsible AI Transparency Report), grants that fund “wetlab” research. The UK institute’s Alignment Project names its funders: “the UK AI Security Institute, Canadian AI Safety Institute (CAISI)… OpenAI, Microsoft, Amazon Web Services (AWS), Schmidt Sciences, Anthropic” among others (Alignment Project). Its External Red Team Alliance “provides unrestricted funding to 18 university labs across six continents” (Microsoft, 2026 Responsible AI Transparency Report). AVERI says “We’ve received API credits as donations or compensation from Amazon, Anthropic, Google DeepMind, Microsoft, OpenAI, and Thinking Machines Lab” (AVERI, About).

Institutional Voice & Intent

Microsoft’s register is the policy paper, the transparency report and the president’s op-ed: rules first, hedges in the conditional, and a framework that says who decides. The Frontier Governance Framework (February 2026) “tracks the following capabilities that we believe could emerge over the short-to-medium term and threaten national security or pose at-scale public safety risks if not appropriately mitigated”: chemical, biological, radiological and nuclear weapons, offensive cyberoperations, advanced autonomy, loss of control and harmful manipulation. “Models assessed as posing low or medium risk may be deployed with appropriate safeguards as outlined below. Models assessed as having a high or critical risk are subject to further review and safety and security mitigations prior to deployment.” The decision: “The Executive Officers (or their delegates) will make the final decision on whether to approve the recommendation for secure and trustworthy deployment.” Outside parties enter at the company’s discretion: “We engage qualified third parties to conduct evaluations in ways that are appropriate to the risk profile of the model,” and “The decision to involve qualified and expert external actors that meet relevant security standards, including those with domain-specific expertise, is made by internal experts who are independent from the model development team.” Disclosure: “Information about the capabilities and limitations of the model, relevant evaluations, and the model’s risk classification will be shared publicly, with care taken to minimize information hazards.” The change log records the February 2026 update as “One year update following finalization of the EU Code of Practice for General Purpose AI (CoP), NY RAISE Act and CA Transparency in Frontier Artificial Intelligence Act” (Microsoft, Frontier Governance Framework, Feb 2026 (PDF)).

Who tests it, in the company’s words. The 2026 Responsible AI Transparency Report (1 September 2026): “Over the last year, we have worked with ten firms that have general AI risk and red teaming expertise,” and “we partnered with eight firms and government research institutes with deep subject matter expertise” in areas “including biological, chemical, and cyber weapons.” It names none of them. It does name the government counterparties: “Microsoft has put in place a Memorandum of Understanding with the US Center for AI Standards and Innovation (CAISI) and a Collaboration Agreement with the UK AI Security Institute (AISI),” through which “qualified external evaluators with national security expertise can help examine how our AI models and systems behave under challenging and unexpected conditions.” Elsewhere in the same report: a CRADA with NIST and an MoU with CAISI “focus on improving methodologies for evaluating the resilience of AI technologies under real-world adversarial pressure,” an MoU with the Australian Government for “human-AI interaction risks,” and an MoU with Singapore’s IMDA (Microsoft, 2026 Responsible AI Transparency Report). NIST’s centre signed testing agreements with “Google DeepMind, Microsoft and xAI” in May 2026, under which “the agreements allow government evaluators to assess frontier AI models before public release” (ExecutiveGov). See US CAISI and UK AI Security Institute.

The rulebook as applied: MAI-Thinking-1 (2026). The technical paper for the company’s first reasoning model: “Additional red-teaming was conducted independently by Microsoft’s AI Red Team (AIRT) and third-party vendors,” scoped “to risk areas where static evaluations are weakest,” and “Structured dangerous-capability / uplift evaluations (Phuong et al., 2024; Li et al., 2024b) are out of scope for this release.” The launch post says: “We don’t distill from other labs and we don’t rely on opaque data,” and states the design stance: “our models must not refuse legitimate requests under the guise of safety and compliance as then they are not truly serving humans,” with “Safety is trained with the same reinforcement learning infrastructure used for capability” (Microsoft AI, technical paper (PDF); Microsoft AI, Introducing MAI-Thinking-1). No third-party evaluator is named in either document. The record is in The Evaluator Alliances.

Stated intent: Build “Humanist Superintelligence” kept “subordinate” and “under human control”; ask governments for binding rules and pre-release evaluation; and run a framework in which executive officers make the final deployment decision.

Observed intent: Hold stakes in and sell compute to every frontier lab, build a model of its own that is not yet claimed to approach the framework’s high-risk thresholds, sign testing agreements with the US and UK institutes, and put the call for binding rules in the president’s voice.

Gap: Narrow on what is promised, wide on who is named. The framework gives the company the choice of which third parties to involve; the transparency report counts “ten firms” and “eight firms and government research institutes” and names none; the model paper names none and puts the evaluations that would test the framework’s own tracked capabilities out of scope. The record does not settle whether the MAI models are in scope of the framework’s high-risk tests, because the paper does not say. A company that writes the framework, picks the testers, and decides on release never needs a public refusal to show it works.

Interest: What each position the company argues would pay it, documented, with its own best case.

• Binding rules, licensed developers, and government evaluators before release. Position: the 2023 blueprint, “The first will require new regulations and licensing for these models themselves,” and, to the Senate Judiciary subcommittee on 12 September 2023, “Let’s require licenses for advanced AI models and uses in high risk scenarios. Let’s have an agency that is independent and can exercise real and effective oversight over this category,” with “you do need a license from the government before you go forward, and that is real accountability” (Microsoft On the Issues, 25 May 2023; Tech Policy Press, hearing transcript). In 2026: “Everyone is reluctant to say there should be regulation, but what we really have right now is regulation without transparent or complete rules,” and “The government doesn’t have the tools it needs” (Fortune, 9 Jul 2026); and to the BBC, a legal requirement for incident reports and an “emergency brake,” with “Right now, your washing machine is subject to more safety rules than AI” (Fox Business, Oct 2026). What it pays: a licence and a pre-release test would apply to developers whose compute Microsoft sells (OpenAI’s “$250B of Azure services,” Anthropic’s “$30 billion”); Microsoft already holds the NIST memorandum and the UK collaboration agreement and publishes a framework written to the same laws. Best case, Smith’s: “We know how to do this for other products. We should build on that knowledge and think about how to apply it to AI.” Convergence: Google DeepMind’s standards-body proposal, Anthropic’s call for rules, and the evaluators who would be the testers.
• Government should buy Copilot, free at first. Position: GSA’s agreement “makes Microsoft 365 Copilot available at no cost for up to 12 months for Microsoft G5 customers,” with “potential savings of $3.1 billion in the first year”; Nadella: “we will help federal agencies use AI and digital technologies to improve citizen services, strengthen security, and save taxpayers more than $3 billion in the first year alone” (GSA, 2 Sep 2025). What it pays: the release’s own terms, “no cost for the first year for Microsoft G5 customers with substantial discounts in subsequent years,” make the free year the installed base for a later paid tier. Best case, Nadella’s, above. Convergence: the same schedule at OpenAI, Anthropic, xAI and Google DeepMind.
• Model choice and open weights. Position: the open-weights letter of 24 July 2026 lists “Microsoft” among its signatories and asks policymakers to avoid “premature restrictions on open models that stifle competition or drive innovation overseas” (NVIDIA, Open Weights and American AI Leadership). What it pays: a platform that sells access to OpenAI’s, Anthropic’s, xAI’s, Mistral’s and its own models gains from a market in which no single lab’s rules decide what may be offered; the Mistral agreement lists “Mistral Medium 3.5 and OCR 4” as “now available in Microsoft Foundry.” Best case, the letter’s: open weights “let every organization match the right model to the right job at the right cost.” Convergence: Meta, Google DeepMind and Mistral AI, all on the same letter.
• European sovereignty on Microsoft’s cloud. Position, Smith: “Europe should have access to the world’s most capable AI without compromising control over their data, operations or digital future,” by “bringing Mistral’s frontier European models into our sovereign cloud portfolio” (Microsoft Source, 21 Jul 2026). What it pays: the sovereign-cloud portfolio is a Microsoft product line, and the Mistral agreement is a “multibillion dollar commitment” that expands Microsoft’s capacity in Europe. Best case, Smith’s: “we are honoring the European Digital Commitments we made.” Convergence: Mistral AI, whose chief executive says “With Microsoft as our partner, our models reach enterprises and public institutions at global scale.”

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

• Social accounts and feeds. Feed: Microsoft On the Issues. X: none verified in this record.

Position in the Apparatus

Microsoft sits in the apparatus as landlord, stakeholder, counterparty and, since 2025, developer. As landlord it hosts the models of other frontier labs (OpenAI, Anthropic, Mistral AI, xAI) and sells compute to OpenAI and Anthropic. As stakeholder it holds roughly 27 percent of OpenAI and an Anthropic investment on which it booked a $3.2 billion gain in one quarter. As counterparty to governments it has a memorandum with NIST’s centre, a collaboration agreement with the UK AI Security Institute, an MoU with Australia’s institute and with Singapore’s IMDA, and the GSA agreement (Microsoft, 2026 Responsible AI Transparency Report). As developer it publishes a framework, a transparency report and a model paper, and names no outside evaluator in the last two. Its regulators and rule-writers include the EU’s AI Office (Microsoft is among the signatories the Commission lists for the Code of Practice), the US and UK institutes, and the courts (European Commission). Its peers are OpenAI, Anthropic, Google DeepMind, Meta, xAI and Mistral AI, and it is a founding member of the Frontier Model Forum. The White House accord of 29 September 2026 “was signed by Trump, Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Nvidia CEO Jensen Huang, and Elon Musk, founder of xAI,” and names no Microsoft signatory (NPR / AP).

Actions & Leadership Choices

The model with the dangerous-capability tests out of scope. MAI-Thinking-1 is a “35B active, ~1T total parameters, MoE” model, “Microsoft AI’s first reasoning model,” with a technical paper dated June 2026 and public preview in Foundry as of 12 August 2026 (Microsoft AI, MAI-Thinking-1). The paper’s safety section reports internal and independent red teaming and says dangerous-capability and uplift evaluations are “out of scope for this release” (Microsoft AI, technical paper). The Frontier Governance Framework runs “a leading indicator assessment” on “any model that is in scope for frontier model requirements under applicable laws” and on substantially fine-tuned third-party models; the paper does not say whether MAI-Thinking-1 was in scope.

The president’s October interview. To the BBC, Brad Smith called for “a notification mechanism for victims whose systems are breached by a rogue AI agent, as well as an ’emergency brake’ available to developers,” in Fox Business’s summary, and said “We know how to do this for other products. We should build on that knowledge and think about how to apply it to AI” (Fox Business). In July he told Fortune, of the government’s action on Anthropic’s Fable, “when the government gets that information, I think it’s right to act,” and “what the government found was that it only had one regulatory tool it could use: an export control tool” (Fortune).

The Mistral expansion (21 July 2026). Microsoft Source: “Microsoft will leverage Mistral’s expanded Europe-based GPU infrastructure to increase capacity for AI development and to support the delivery of MSFT’s cloud and AI services,” and “Mistral Medium 3.5 and OCR 4 are now available in Microsoft Foundry, and Mistral Medium 3.5 is now in Microsoft Copilot Studio” (Microsoft Source). Mistral’s chief executive: “Our mission has always been to put frontier AI in the hands of every organization while keeping them in control of their technology.”

The lawsuit. Microsoft Corporation is a named defendant in The New York Times Company v. Microsoft Corporation, No. 1:23-cv-11195 (S.D.N.Y., filed 27 December 2023), a copyright action now consolidated in In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143 (CourtListener docket).

Leadership choices. The president speaks for rules and the CEO for the platform; the lab is run by Suleyman’s MAI Superintelligence Team inside Microsoft AI, and the framework puts the deployment decision with “the Executive Officers (or their delegates).” The chief responsible AI officer reviews framework updates: “Any updates to our practices will be reviewed by Microsoft’s Chief Responsible AI Officer prior to their adoption” (Microsoft, Frontier Governance Framework). The documented counterpart is the list of government counterparties in the transparency report, the Alignment Project and Frontier Model Forum funding, and the university labs in the External Red Team Alliance.

CONDUCT: ASKS FOR BINDING RULES, HOSTS ITS RIVALS’ MODELS. Microsoft’s president asks for licensing, pre-release government evaluation and an emergency brake, and the company holds the testing agreements that would be the vehicle. The same company sells compute to OpenAI and Anthropic, buys capacity from Mistral, offers xAI’s model, publishes no outside evaluator for its own, and is not on the accord’s signature list. Each is a lawful choice on the record.

Alliance Map

These run at the same time. One company holds 27 percent of one frontier lab, an investment gain in a second, a multibillion-dollar capacity agreement with a third, and a platform that sells a fourth; it funds the labs’ grant fund, trade forum and a UK government research fund; it holds testing agreements with the US and UK institutes; it sells Copilot to the federal government free for a year; it is a defendant in the copyright suit; and its president asks for licensing, pre-release evaluation and an emergency brake. The overlap is the finding, not a coordination claim.

Reach Assessment

Institutional: The reach is platform, seat and test channel read together: “Azure revenue surpassed $100 billion for the first time,” Microsoft 365 Copilot “over 30 million paid seats,” a no-cost first year in the federal government’s procurement vehicle, a memorandum with NIST’s centre and an agreement with the UK institute (Microsoft, FY2026 Q4 results; GSA). Memetic: The framework and the transparency report are Microsoft’s published account of its practices, and its president’s October interview put the comparison to washing machines and milk in public. Civilizational: The company says in its own paper that dangerous-capability evaluations are “out of scope” for its first model, and in its framework that its executive officers decide on release.

Score basis. Atk 8 is reach: Nadella says “Azure revenue surpassed $100 billion for the first time, and Microsoft 365 Copilot reached over 30 million paid seats”; GSA’s agreement “makes Microsoft 365 Copilot available at no cost for up to 12 months for Microsoft G5 customers”; and the transparency report records a memorandum with CAISI and a collaboration agreement with the UK institute (Microsoft, FY2026 Q4 results; GSA; Microsoft RAI report). It is an 8 and not a 9 because no statute obliges anyone to use Microsoft’s own models and its largest frontier-model reach runs through OpenAI’s and Anthropic’s. Def 8 is funding insulation: “Revenue was $90.0 billion” and “Operating income was $40.6 billion” for the quarter, with the Anthropic gain and the OpenAI stake the company’s own to hold (Microsoft, FY2026 Q4 results; Microsoft, 28 Oct 2025); it is not a 9 because OpenAI’s “incremental $250B of Azure services” and Anthropic’s “$30 billion” tie the backlog to two counterparties. Hp 9 is entrenchment: IP rights “through 2032,” a commercial backlog that “increased 84% to $678 billion,” and a federal vehicle that runs “through September 2026” for opt-in with pricing “for up to 36 months” (Microsoft, 28 Oct 2025; Microsoft, FY2026 Q4 results; GSA). Peers: xAI holds the same 8/8/9 (hazard 83) on a smaller user base and a single dominant customer; OpenAI holds 9/8/9 (88) with a consumer product of its own; Google DeepMind holds 9/8/9 (87) with its own frontier model in the lead role where Microsoft’s is a medium-sized reasoning model.

Sources: Microsoft, Frontier Governance Framework, Feb 2026 (PDF); Microsoft, 2026 Responsible AI Transparency Report (PDF); Microsoft On the Issues, 1 Sep 2026; Microsoft AI, technical paper, MAI-Thinking-1 (PDF); Microsoft AI, Introducing MAI-Thinking-1; Microsoft AI, MAI-Thinking-1; Microsoft AI, Towards Humanist Superintelligence, 6 Nov 2025; Microsoft, 28 Oct 2025; Microsoft, 18 Nov 2025; Microsoft Source, 21 Jul 2026; Microsoft, FY2026 Q4 results; Microsoft, Satya Nadella; Microsoft, Brad Smith; Microsoft, Kevin Scott; Microsoft On the Issues, 25 May 2023; Tech Policy Press, Senate Judiciary hearing transcript, 12 Sep 2023; Fortune, 9 Jul 2026; Fox Business, Oct 2026; GSA, 2 Sep 2025; ExecutiveGov, 6 May 2026; NPR / AP, 30 Sep 2026; European Commission, GPAI Code of Practice; UK government, Frontier AI Safety Commitments; Mistral AI, Mistral Large; NVIDIA, Open Weights and American AI Leadership (PDF); Frontier Model Forum, AI Safety Fund; Alignment Project; AVERI, About; SpaceXSI on X; CourtListener, NYT v. Microsoft; LDA, Q1 2026; LDA, Q2 2026; LDA, Q3 2026.