Nine Hundred Million People at the Bottom of One Machine

The question was never who owns the model. It was who allocates the mind.

Palantir is the hard-power model of state-adjacent AI. It enters through defense, intelligence, targeting, border control, emergency coordination, and public-sector crisis. Its political language is martial: the West must recover seriousness, and private technology must reattach itself to the survival of the republic.

Microsoft and OpenAI are a different species of power.

They do not enter through the battlefield. They enter through the document, the inbox, the spreadsheet, the hospital note, the cloud contract, the procurement system, and the classified deployment environment. Palantir asks the republic to accept the stack as hard power. Microsoft and OpenAI teach institutions to experience the stack as ordinary work.

The distinction is one of emphasis, not a clean line. Palantir also runs through civilian and administrative workflows, and Microsoft and OpenAI are now deep inside defense and intelligence. But the center of gravity differs. Palantir’s terrain is action. Microsoft and OpenAI work one layer earlier, on the cognition that runs before action: what intelligence is available before coordination, targeting, and execution begin.

That earlier layer is the whole game. The Microsoft-OpenAI model is not simply another case of sovereignty migrating into infrastructure. It is the emergence of something quieter and broader, an access regime for machine intelligence.

The question is no longer only who owns the model, who controls the cloud, or who profits from enterprise AI. The more important question is who decides how intelligence itself is distributed: who receives it, at what tier, under what permissions, with what audit requirements, through which interface, and under whose terms.

The state keeps the authorization. The partnership increasingly runs the execution layer. And the mission language explains why this is not capture but progress.

The Access Regime Is Not the Security State

The access regime is the emerging order in which access to machine intelligence becomes tiered, permissioned, priced, audited, and differentially distributed across society.

It is not the welfare state, which allocates benefits. It is not the administrative state, which allocates rules. It is not the security state, which allocates suspicion and force. The access regime allocates cognition.

It decides which actors reach machine reasoning, summarization, prediction, and decision support, and which institutions receive consumer, enterprise, audited, classified, or sovereign deployment.

A free-tier user receives one version of intelligence. A paid subscriber receives another. An enterprise customer receives a different set of capabilities, privacy guarantees, and data-retention terms. A government agency receives another configuration again, and a classified user inside a defense or intelligence environment receives yet another, deployed through restricted infrastructure and surrounded by authorization, auditing, and compliance.

This is usually described as product segmentation. That framing is too small. When intelligence becomes a general-purpose input into institutional action, product segmentation becomes governance. It decides who can think with machines, at what level, and under what constraints.

The regime does not announce itself as a new order. It appears as pricing plans, compliance tiers, cloud regions, model-access rules, procurement certifications, and enterprise dashboards. Power no longer needs to appear as command. It can appear as access.

The Partnership Is an Access Architecture

The Microsoft-OpenAI partnership is usually read as a business story: the multibillion-dollar commitment, the cloud arrangement, the integration of OpenAI models into Microsoft products. That story is accurate and insufficient.

The more important fact is structural. OpenAI supplies frontier model capability. Microsoft supplies cloud infrastructure, enterprise distribution, government authorization pathways, security certifications, and the productivity interface through which much of the institutional world already works. Together they do not simply sell AI. They build the architecture through which access to AI is granted, restricted, monitored, and normalized.

When Microsoft announced its third investment phase in OpenAI in January 2023, the formal language was partnership. The structural reality was more specific: Azure became the exclusive cloud for OpenAI workloads, OpenAI’s compute ran on Microsoft’s infrastructure, and Microsoft’s products ran on OpenAI’s models. The dependency was mutual and deliberate. [1]

The relationship has since hardened into something closer to constitutional structure, then loosened again without losing its core. The October 2025 restructuring converted OpenAI’s commercial arm into a public benefit corporation and left Microsoft holding roughly 27 percent of OpenAI Group PBC, an investment the company valued at about 135 billion dollars. The April 2026 amendment then relaxed the exclusivity: OpenAI can now serve its products across any cloud, while Microsoft remains its primary cloud partner, OpenAI products ship first on Azure under defined conditions, and Microsoft keeps a non-exclusive license to OpenAI model and product IP through 2032. [2]

The structural point survives every amendment. Microsoft is a major shareholder, primary infrastructure partner, distribution channel, and enterprise interface for OpenAI-derived intelligence. The deeper return was position: Microsoft placed itself between institutional demand for intelligence and the infrastructure required to deliver it, and OpenAI gained one of the most important enterprise and government distribution channels in the world.

The result is a hybrid that does not fit existing regulatory categories. OpenAI keeps its mission governance and nonprofit parent; Microsoft holds commercial deployment rights, cloud priority, and the productivity layer. Neither fully controls the other, and neither is directly accountable to the institutions that increasingly depend on their combined system. This is not a software company and its cloud vendor. It is a distributed access architecture with no sovereign above it.

Azure Is the Control Plane

It helps to stop thinking of cloud as storage or compute. In the AI age, cloud is the control plane for permissioned intelligence. It determines where models run, what data they touch, which institutions can deploy them, what audit trails exist, and which users reach which capabilities.

This is most visible in government. In May 2024, Microsoft deployed GPT-4 to an isolated, air-gapped Azure Government Top Secret cloud for the Department of Defense. [3] By September 2024, Azure OpenAI had received FedRAMP High authorization and Defense Information Systems Agency approval for DoD Impact Levels 4 and 5. [4] In January 2025, GPT-4o was authorized for use at the intelligence community’s top-secret level under Intelligence Community Directive 503. [5] By April 2025, Azure OpenAI had been authorized across all U.S. government data classification levels, from unclassified through the Secret-tier Impact Level 6. [6]

Within roughly a year, OpenAI models moved from enterprise productivity tools to authorized infrastructure for classified defense and intelligence work at every classification level the government recognizes.

This did not happen only because Microsoft is powerful, though it is. It happened because the institutions adopting Azure OpenAI had a coordination problem existing systems could not solve. Agencies, military commands, and intelligence bodies produce more information than they can integrate, and they need systems that surface signals, summarize fragmented data, and deliver outputs into operational workflows at machine speed. Azure offered the path of least resistance: secure cloud, model access, compliance history, procurement familiarity, and a vendor already embedded across the enterprise state.

The certifications are not theater. FedRAMP High and the DISA authorizations are serious security processes, and Microsoft earned them. They do not prove dependency. They create the conditions under which dependency becomes institutional. Once an agency deploys Azure OpenAI at high classification levels, its workflows, training cycles, procurement habits, and institutional routines begin reorganizing around that infrastructure. The question stops being whether the system is safe enough and becomes whether the institution can function without it. The state certifies the infrastructure, and the infrastructure becomes part of the state’s capacity.

Copilot Is the Civilian Interface

Azure is the substrate; Copilot is the interface. This is where the model departs from harder forms of state-adjacent AI. Palantir is visible as power because it enters domains where power already appears nakedly. Microsoft enters through routine.

The meeting needs a summary. The doctor needs a note. The analyst needs a draft, the manager a report, the inbox a triage. Nothing feels constitutional. Everything feels useful. That is the genius of the model.

By late 2024, Microsoft reported that nearly 70 percent of the Fortune 500 was using Microsoft 365 Copilot. [7] A UK government trial of 20,000 civil servants reported average savings of 26 minutes a day, though that number was self-reported, and other government trials, one with a control group, found smaller gains or none at all. [8] The exact figures are contested. The direction is not: the interface is arriving inside institutional work at scale.

The productivity framing is true. It is also a retail explanation. The structural reality is that Copilot changes the surface through which institutions think. It does not only save time. It changes what gets surfaced before human judgment arrives. It drafts the document that gets revised rather than written, summarizes the meeting that will be remembered instead of the meeting that occurred, and runs the compliance check that decides whether an action proceeds. Each is a small relocation of cognition.

Before Copilot, a human decided what to look at, what to include, what to omit, and how to summarize. After Copilot, those choices are increasingly made first by a system whose parameters were set in Redmond and San Francisco, whose behavior was fixed by training and alignment processes no ordinary institution reviewed, and whose outputs arrive already formatted as helpfulness. The institution still decides. But the field of decision has already been shaped.

This is why Copilot matters beyond productivity. It is one of the most widely deployed civilian interfaces between private AI infrastructure and institutional cognition. The danger is not that Microsoft owns the state. It is that institutional cognition begins to pass through Microsoft’s interface before anyone experiences that as governance.

Tiered Access Is Governance

OpenAI’s deployment architecture is explicitly stratified, from the free consumer tier up through enterprise, government, and classified deployments, each with its own controls, data protections, and audit rules. [9]

The scale is easy to miss because the floor is so familiar. Roughly 900 million people now use ChatGPT every week. [10] They occupy the consumer level of an architecture whose upper tiers reach into air-gapped classified environments. One structure. Nine hundred million people at the bottom of it, a few thousand cleared analysts at the top.

That image is a gradient of permission and governance, not a single chain of command, and not a simple ladder of raw capability. The cleared analyst’s system is not necessarily smarter than the consumer’s; authorization is slow, and the classified tier may run older or more restricted models. What differs across the tiers is what each is allowed to do with machine intelligence, under what oversight, and against what data. The student and the cleared analyst do not answer to the same officer. They draw on related model families inside one connected commercial architecture, permissioned differently at each level.

That pricing tiers are themselves a political technology, that formal access can be broad while effective access stays narrow, is the argument of a companion essay. The point here is narrower and architectural: the same model family, governed differently at each tier, produces categorically different institutional objects. The Pentagon analyst running GPT-4o in a classified environment is not using the same thing as the student on a consumer chatbot. The model may be related. The governance context is not.

The architecture decides who gets what mind.

No single actor allocates all of this. The distribution is the product of many hands: customer demand, Microsoft’s pricing, OpenAI’s policy, government procurement, security rules, and rival vendors. But the aggregate behaves like an allocation. The sum of these choices, none of them a public decision, determines who reaches which capability. That is how access regimes form: not through a constitutional convention, but through product architecture. By the time the political system recognizes that access to machine intelligence has become a public question, the distribution may already be embedded.

Every such regime also produces a residual: actors who cannot afford the tier, cannot satisfy the compliance requirements, or receive only a degraded version of machine cognition. In the access regime, intelligence is not equally available. It is permissioned.

The Therapeutic Shield

Every transfer of power into private infrastructure needs a legitimacy argument. Palantir’s is martial: Western survival, democratic hard power, civilizational defense. Microsoft and OpenAI use a different one, therapeutic and humanitarian. AI will empower workers, save time, reduce burnout, return clinicians to patients, and benefit all of humanity.

This language is not false, which is why it works. Copilot does save time. Azure does let institutions reach capabilities they could not build. The benefits are real. But a shield does not have to be false to be structural. It takes a complex transfer of power into infrastructure and makes it narratable: dependency reframed as empowerment, automation as augmentation, unequal access as innovation rather than allocation.

OpenAI’s mission language does the same work at the civilizational level. The company was founded to ensure that artificial general intelligence benefits all of humanity, and every subsequent governance turn, the capped profit, the restructuring fights, the public benefit corporation, has been narrated as preservation of that mission. The point is not that the mission is fake. It is that mission language helps stabilize a structure in which accountability is hard to locate. The mission says the partnership is accountable to humanity. The structure says no specific public institution governs the access regime. That is the contradiction.

Dependency Arrives as Convenience

A hospital that deploys clinical documentation AI does not experience itself as surrendering capacity. It experiences less administrative burden. An agency that deploys Azure OpenAI does not experience outsourced sovereign cognition. It experiences faster analysis. Dependency does not announce itself. It accumulates as convenience.

An institution can change its CRM, migrate email with enough pain, replace a project tool. An institution that has reorganized its analytical workflows around a specific AI interface faces a different problem. The interface has not merely stored its information. It has shaped how the institution sees. That is not only vendor lock-in. The deeper risk is that the institution forgets how to work without the interface.

And the interface is not neutral. A meeting summary decides what mattered. A document draft sets the first structure of an argument. A compliance assistant encodes a model of risk. Each output can become the first version of reality the institution sees, and institutions are path-dependent: the initial summary becomes the shared memory, the first draft becomes the basis for revision. Human review remains present, but review is not authorship. A person editing a machine-generated summary is not in the position of a person who decided from scratch what the meeting meant. The human stays in the loop. The loop has changed.

The access regime does not need to replace human judgment. It only needs to supply the environment in which judgment operates.

Where Accountability Goes When Allocation Fails

In October 2023, New York City launched a chatbot called MyCity, built on Microsoft’s Azure AI cloud and trained on the city’s own regulations, to help small businesses navigate the rules. Within months, an investigation by The Markup found it telling business owners they could take a cut of workers’ tips, that landlords could turn away tenants with housing vouchers, and that a business could refuse to accept cash, each of them illegal under New York law. [11] Asked identical questions, it returned different answers to different users.

What happened next is the part that matters. The mayor acknowledged the bot was wrong in places and declined to take it down. [12] The city added a disclaimer telling users not to treat its answers as legal advice. Microsoft said it was working on accuracy but would not say what was causing the errors. The bot stayed live for nearly two more years, until a new administration shut it down as unusable. [13]

Notice where the accountability went. The judgment ran through a private system the city did not build and could not fully explain. The wrong answers could have entered at the model, the retrieval design, the system prompt, the source material, or the contractor’s implementation, and the city could not publicly say which. When it failed, the vendor pointed at ongoing fixes, the city pointed at a disclaimer, and the wrong answers stayed online. No one owned the output. The disclaimer, which told citizens to check the government’s own tool against the government’s own website, is the legitimacy wrapper in its purest form: the human is told to stay in the loop precisely so the institution does not have to answer for the machine.

The reflex has a cleaner statement in a Canadian tribunal case: an airline argued it could not be liable for its chatbot’s wrong fare advice because the bot was, in its telling, a separate legal entity responsible for its own actions. [14] The tribunal rejected it, but the argument is the tell. When an automated system produces a consequential error, the deploying institution’s first move is to place the authorship somewhere else.

This is what an earlier essay called the allocation state: the public institution held accountable for judgments produced in systems it does not govern. The access regime supplies the cognition; the allocation state inherits the liability.

The same shape recurs higher up, where it is harder to see. Consider the chain behind a government analyst using Azure OpenAI in a classified environment. OpenAI trained the model. Microsoft deployed and secured it. DISA and other bodies authorized the environment. A contracting officer approved procurement, an integrator built the workflow, an analyst used the tool, and a decision-maker acted on the output. If that system surfaces incomplete intelligence or shapes a decision badly, each link points to the next, and none is answerable for the whole.

A classified analytical system differs from MyCity in almost every way: its users, its data, its oversight, its stakes. It adds one more difference that matters here. No outsider can run the test. A journalist could type questions into a public chatbot and publish the wrong answers; no one can do that to a classified system. The conditions for the same diffusion of accountability exist there, in the tiers where the stakes are highest, precisely where it cannot be seen. Security review can confirm that a system meets technical controls. It cannot determine how machine-generated cognition reshapes institutional judgment over time. The institution that outsources part of its perception may also outsource its ability to know what it is missing. That is the failure no one is ready to govern.

The State Certifies What the Stack Has Already Made Necessary

Regulators are not absent. The EU has scrutinized Microsoft’s bundling. Antitrust authorities have examined the partnership. State attorneys general have watched OpenAI’s restructuring. Defense and intelligence authorizations required serious review. The state is present. But it arrives downstream. It reviews contracts after dependency has begun, certifies security after architecture is designed, and investigates market power after the product is embedded.

Deployment moves at the speed of enterprise procurement. Regulation moves at the speed of law. The stack moves first, and the state arrives later to certify, regulate, or legitimize a system that has already become useful. The access tiers were designed before the political theory caught up. The state does not lose sovereignty in name. It loses the operational independence that sovereignty needs in order to mean anything.

Does This Survive Fragmentation

The strongest objection is that this regime is already breaking apart. Open-weight models are catching up. Rival clouds compete for the same workloads. Sovereign-cloud and data-localization rules push deployment onto national infrastructure. And the April 2026 amendment freed OpenAI to serve its products on any cloud, ending Azure’s exclusivity.

These are real, and some cut the right way: competition and open weights can lower prices and weaken any single vendor’s grip. What they do not do is make the allocation equal, transparent, or publicly governed. The April 2026 amendment is the clearest case: OpenAI’s intelligence can now be served across Azure, AWS, and others, which does not flatten access so much as add a layer of tiering between clouds. A world with three frontier labs, four hyperscalers, and a dozen sovereign stacks is not a world in which machine intelligence is equally available. Fragmentation changes who holds the toll, and how many tolls there are. It does not remove them.

The Access Regime Enters the IPO Pipeline

In early June 2026, the access regime entered the IPO pipeline. Anthropic confirmed on June 1 that it had confidentially submitted a draft S-1 to the Securities and Exchange Commission. Within the week, OpenAI announced its own confidential filing, noting that timing remained undecided. [15]

The financial press asked the financial questions: whether revenue supports the valuations, whether the margins exist, when the burn ends. The constitutional question is different. If these companies proceed to public listings, the firms helping allocate machine cognition to states, enterprises, hospitals, schools, and ordinary users acquire a new governing constituency: public shareholders.

A listing changes who a company answers to. A public benefit corporation’s directors owe their duties to the corporation and are bound to weigh its mission alongside profit, so the mission language can survive on paper. But the daily gravity of a listed company is disclosure, growth, guidance, and the share price, and that gravity pulls in one direction. The access regime may go public in precisely the wrong sense, owned by the public as investors before it is governed by the public as citizens.

The infrastructure math explains the pull. In late 2025, Sam Altman described roughly 1.4 trillion dollars in infrastructure commitments over eight years. By February 2026, OpenAI was reportedly telling investors that it expected around 600 billion dollars in compute spending through 2030. The number fell; the dependency did not. A company operating at either scale has to keep raising, and to keep raising it has to satisfy the markets that supply the capital. [16] The allocator of intelligence becomes, structurally, an applicant to capital. The listing does not corrupt the access regime. It completes it: every tier now has a price, including the company’s own.

The Constitutional Question

Microsoft and OpenAI are the clearest present example of a structure that will not stay unique to them. Wherever frontier AI requires massive compute, safety evaluation, and government authorization, a small number of private actors will sit between institutional demand and machine intelligence. Their power will not come only from owning models. It will come from controlling access.

The future is not one universal intelligence system available equally to everyone. It is a layered regime shaped by capability, price, data protection, security clearance, and regulatory permission. That makes the distribution of machine cognition one of the central political-economic questions of the age. The politics of AI will not only be about whether models are aligned. It will be about who receives which aligned model, under what conditions, and through whose infrastructure. A society that distributes machine intelligence unequally is not simply adopting a technology. It is building a cognitive class structure.

The constitutional question of the AI age may not begin with weapons, borders, or surveillance. It may begin with something quieter: who allocates intelligence, and who governs the interface before the decision is made.

By now the pieces have names: the substrate, the interface, tiered access, the therapeutic shield, the dissolving of accountability, and the discipline of the capital markets. Together they form the soft infrastructure through which institutions think.

This is not capture in the old sense, a company seizing public authority from outside. It is subtler: a private-public intelligence regime becoming necessary to how institutions function. The state still speaks. The institution still decides. The human stays in the loop. But the intelligence available before the decision is increasingly allocated through private infrastructure, and the access regime does not need to own the decision. It only needs to govern what can be seen, summarized, ranked, and drafted before the decision is made.

The question was never who owns the model. It was who allocates the mind.

Notes

[1] Microsoft’s January 23, 2023 announcement of its third investment phase in OpenAI named Azure as the exclusive cloud for OpenAI workloads across research, products, and API services. Microsoft, “Microsoft and OpenAI extend partnership,” Official Microsoft Blog, January 23, 2023: https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/

[2] Microsoft’s investment in OpenAI has totaled roughly 13 billion dollars since 2019. Under the October 28, 2025 restructuring, OpenAI’s commercial operations became a public benefit corporation and Microsoft’s stake in OpenAI Group PBC was valued at about 135 billion dollars, roughly 27 percent on an as-converted basis. On April 27, 2026 the companies amended the partnership: OpenAI may serve products on any cloud, Microsoft remains primary cloud partner with OpenAI products shipping first on Azure under defined conditions, and Microsoft retains a non-exclusive license to OpenAI model and product IP through 2032. Microsoft, “The next chapter of the Microsoft-OpenAI partnership,” October 28, 2025: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/ and “The next phase of the Microsoft-OpenAI partnership,” April 27, 2026: https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/

[3] Microsoft deployed GPT-4 to an isolated, air-gapped Azure Government Top Secret cloud for Department of Defense use, announced May 7, 2024. DefenseScoop, “Microsoft deploys GPT-4 large language model for Pentagon use in top secret cloud,” May 7, 2024: https://defensescoop.com/2024/05/07/gpt-4-pentagon-azure-top-secret-cloud-microsoft/

[4] In September 2024, Azure OpenAI Service was approved within the FedRAMP High authorization for Azure Government and within the DISA DoD Impact Level 4 and 5 provisional authorization. Microsoft Azure Government blog, “Azure OpenAI, including GPT-4o, approved as a service within the FedRAMP High Authorization,” updated September 3, 2024: https://devblogs.microsoft.com/azuregov/azure-openai-fedramp-high-for-government/

[5] GPT-4o was authorized for use in Azure Government Top Secret under Intelligence Community Directive 503, announced January 16, 2025. DefenseScoop, “OpenAI’s GPT-4o gets green light for top secret use in Microsoft’s Azure cloud,” January 16, 2025: https://defensescoop.com/2025/01/16/openais-gpt-4o-gets-green-light-for-top-secret-use-in-microsofts-azure-cloud/

[6] Microsoft Azure Government blog, “Azure OpenAI Service now authorized for all U.S. Government data classification levels,” April 16, 2025: https://devblogs.microsoft.com/azuregov/azure-openai-authorization/

[7] Microsoft, “Ignite 2024: Why nearly 70% of the Fortune 500 now use Microsoft 365 Copilot,” Official Microsoft Blog, November 19, 2024: https://blogs.microsoft.com/blog/2024/11/19/ignite-2024-why-nearly-70-of-the-fortune-500-now-use-microsoft-365-copilot/

[8] UK Government Digital Service, “Microsoft 365 Copilot Experiment: Cross-Government Findings Report,” trial of about 20,000 civil servants across 12 organizations, September to December 2024, published June 2025; findings summarized in a UK parliamentary written statement, June 2, 2025: https://questions-statements.parliament.uk/written-statements/detail/2025-06-02/hlws667. The 26-minute figure is self-reported; a later Department for Work and Pensions study using a control group found a 19-minute average, and a Department for Business and Trade trial found no clear productivity gain.

[9] OpenAI’s standard API tier retains inputs and outputs for up to 30 days by default; eligible enterprise customers and endpoints can receive Zero Data Retention; classified government deployments on Azure operate under separate access controls and authorization frameworks.

[10] OpenAI and subsequent reporting placed ChatGPT at roughly 900 million weekly active users in 2026.

[11] Colin Lecher, “NYC’s AI Chatbot Tells Businesses to Break the Law,” The Markup, co-published with Documented and THE CITY, March 29, 2024: https://themarkup.org/artificial-intelligence/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law. The MyCity chatbot was built on Microsoft’s Azure AI cloud.

[12] Colin Lecher, Katie Honan, and Maria Puertas, “Malfunctioning NYC AI Chatbot Still Active Despite Widespread Evidence It’s Encouraging Illegal Behavior,” The Markup and THE CITY, April 2, 2024: https://themarkup.org/artificial-intelligence/2024/04/02/malfunctioning-nyc-ai-chatbot-still-active-despite-widespread-evidence-its-encouraging-illegal-behavior. Mayor Adams acknowledged the errors at an April 2, 2024 press conference and left the chatbot online.

[13] Colin Lecher and Katie Honan, “Mamdani to Kill the NYC AI Chatbot We Caught Telling Businesses to Break the Law,” The Markup and THE CITY, January 30, 2026: https://themarkup.org/artificial-intelligence/2026/01/30/mamdani-to-kill-the-nyc-ai-chatbot-we-caught-telling-businesses-to-break-the-law

[14] Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal). The tribunal held Air Canada liable for inaccurate fare information provided by its website chatbot and rejected the airline’s argument that the chatbot was a separate legal entity responsible for its own actions: https://canlii.ca/t/k2spq

[15] Anthropic, “Anthropic confidentially submits draft S-1 to the SEC,” June 1, 2026: https://www.anthropic.com/news/confidential-draft-s1-sec. OpenAI announced its own confidential S-1 submission in early June 2026, noting that timing remained undecided.

[16] OpenAI’s annualized revenue crossed 20 billion dollars in 2025. Reporting has placed OpenAI’s total infrastructure commitments as high as 1.4 trillion dollars.