Introduction: The Day the Pilot Began to Disappear
On September 10, 2026, the U.S. General Services Administration announced what, at first glance, looked like another routine government technology-procurement agreement. Under the next phase of GSA’s OneGov AI strategy, OpenAI would make its ChatGPT models available through a 27-month governmentwide arrangement using discounted, consumption-based pricing, scheduled to begin on October 1, 2026 and run through December 31, 2028.[1][4] Eligible government organizations would pay no platform-access fee — a license that normally costs fifteen dollars per user per month would fall to zero — would face no minimum-order or spending commitment, and would receive a fifty percent discount on token-based usage across eligible models, including models operating in FedRAMP-authorized environments.[1][2] More consequentially, the arrangement’s reach extended well beyond the executive agencies of Washington: federal legislative and judicial entities, state governments, local governments, and tribal governments could all participate through multiple procurement channels, from direct engagement to resellers to supported cloud marketplaces.[1][36]
OpenAI simultaneously framed the agreement in far broader terms than a chatbot license. State, local, tribal, and federal organizations would be eligible for the same basic economics, while every verified government entity would be approved for discounted access to Daybreak Blue, the company’s advanced cyber-defense system, together with scaled training intended to help public-sector cyber defenders find vulnerabilities before adversaries do.[2] The package on offer therefore combined language generation, reasoning, coding, analysis, cybersecurity, workflow support, and — through the inclusion of frontier systems such as GPT-6 Astra — increasingly agentic execution, all wrapped inside a government purchasing architecture explicitly designed for mass adoption rather than cautious experimentation.[2][35]
The numbers make the transition more significant than the press-release language alone would suggest. GSA reported that its OneGov strategy had already generated approximately $1.68 billion in savings for the government, with roughly $1.4 billion of that total attributable to AI agreements that had expanded advanced-model access to approximately 3.5 million federal employees during the promotional era.[1][3] OpenAI, for its part, estimated that the new terms would extend eligibility to a combined public-sector workforce of roughly 23 million people across every level of American government.[2] Bloomberg noted that the original one-dollar-per-year arrangement, struck in August 2025, had represented the steepest technology discount that the government’s central purchasing arm had ever negotiated — a concession whose entire purpose was to make experimentation costless.[3] Government AI, in other words, was no longer confined to a handful of laboratories, chief information officers, defense research programs, or experimental innovation teams. The distribution infrastructure had already been built. The next stage was making advanced intelligence cheap enough, compliant enough, and simple enough to procure that its absence — not its presence — could eventually require explanation.
This development sits inside a much larger policy transformation whose outlines were drawn in the spring of 2025 and elaborated across the eighteen months that followed. OMB Memorandum M-25-21, “Accelerating Federal Use of AI through Innovation, Governance, and Public Trust,” instructs federal agencies to pursue a forward-leaning approach to AI adoption while maintaining safeguards for privacy, civil rights, civil liberties, and public trust, and requires agencies to publish AI strategies, designate Chief AI Officers, and apply heightened controls to what the memo calls “high-impact AI” — systems whose outputs serve as a principal basis for decisions with legal, material, or significant effects on rights or safety.[6][7] Its companion, M-25-22, “Driving Efficient Acquisition of Artificial Intelligence in Government,” directs agencies toward timely and cost-effective acquisition while warning explicitly about vendor concentration, data and model portability, interoperability, and the danger of costly long-term dependence on individual suppliers, embedding those warnings across the entire acquisition lifecycle from market research to contract closeout.[8][9] America’s AI Action Plan goes further still, calling for aggressive government adoption, a GSA-managed AI procurement toolbox, greater interagency coordination, and the broad integration of AI into federal operations as a matter of national strategy.[5] AI adoption has consequently become an administrative objective of the American state rather than merely a technological experiment conducted at its edges.
This is the setting in which the concept of Administrative Default becomes useful. The profound change now underway is not that governments will suddenly allow machines to govern citizens autonomously; that scenario, whatever its long-run plausibility, is not the transformation actually occurring in 2026. The change is subtler and, precisely for that reason, more difficult to see and to govern: AI is becoming embedded one administrative function at a time — summarizing regulations, analyzing procurement documents, comparing grant applications, translating citizen communications, searching enormous case files, writing and testing software, inspecting cybersecurity logs, modeling policy alternatives, drafting correspondence, preparing briefings, managing institutional knowledge, assisting investigators, answering public inquiries, detecting anomalies, and coordinating the increasingly complicated bureaucratic workflows through which the modern state actually touches the lives of its citizens.
Once these systems become sufficiently inexpensive, secure, capable, authorized, and widely distributed, the institutional question changes its polarity. Government officials will no longer ask only whether they should use AI for a given task. They will increasingly be asked — by budget officers, by inspectors general, by legislators, by citizens accustomed to machine-speed service in every other domain of their lives — why they are still doing this particular task without it. That reversal, from a technology that must justify its use to a technology whose non-use must be justified, is the essence of Administrative Default, and it is the subject of this paper.
Why the Title “Administrative Default”
The title deserves a word of explanation, because the choice of terms is itself an argument. The important transformation examined here is not simply government adoption of artificial intelligence; governments have experimented with AI for years, and the scholarly record of that experimentation reaches back well before the generative era. As early as February 2020, the landmark report Government by Algorithm — prepared for the Administrative Conference of the United States by David Freeman Engstrom and Daniel E. Ho of Stanford Law School, Catherine M. Sharkey of NYU, and then-California Supreme Court Justice Mariano-Florentino Cuéllar — documented that nearly half of the 142 most significant federal departments, agencies, and sub-agencies had already experimented with AI and machine learning for tasks spanning regulatory enforcement, adjudication, risk monitoring, and public engagement.[19] What the present moment adds is not experimentation but normalization at scale.
The more consequential threshold arrives when AI becomes the normal starting condition of administrative work — the assumed analytical, computational, linguistic, cybersecurity, and eventually agentic layer surrounding the civil servant, in the same way that email, word processing, and networked databases became assumed layers in earlier generations of government modernization. A pilot requires justification to begin. A default eventually requires justification to avoid. The word “Administrative” points directly at the machinery through which the state actually operates: agencies, procurement systems, benefits programs, regulatory processes, public records, grants, licenses, investigations, cybersecurity operations, policy analysis, and citizen services. The word “Default” describes the moment frontier AI ceases to be an optional productivity application and starts becoming part of the ordinary operating environment of the state — unremarkable, budgeted, audited, and assumed.
Section 1: From AI Pilot to Administrative Default
1.1 The End of the Experimental Era
For much of the first generation of government AI adoption, agencies treated artificial intelligence the way careful institutions treat any unproven technology: as an experiment to be contained. Small innovation teams tested models in sandboxed environments; agencies created approved-use lists and acceptable-use policies; employees joined pilot programs whose very name signaled impermanence; and procurement officials negotiated individual contracts whose modest ceilings reflected modest expectations. The Stanford-NYU study for ACUS captured this era precisely, finding broad but shallow and largely uncoordinated adoption across the federal government, with most use cases still in early or planning stages and concentrated in a minority of technically sophisticated agencies.[19] The scholars who led that work understood, even in 2020, that something larger was coming.
“We are at the dawn of a revolution in how government uses AI” [20]
— David Freeman Engstrom, Stanford Law School
That experimental model, however well suited to an era of uncertain capability, does not scale to millions of government workers using frontier systems daily. The September 10, 2026 GSA agreement represents a fundamentally different procurement philosophy, one designed for saturation rather than sampling. No platform fee, no minimum commitment, consumption-based pricing, standardized terms, governmentwide availability, and multiple ordering channels together dramatically reduce the institutional friction required to experiment with — and subsequently to depend upon — frontier AI.[1][4] The significance of cheap AI therefore extends well beyond budget savings, real as those savings are. Cheap access accelerates institutional normalization, because every barrier removed from the act of trying a technology is simultaneously a barrier removed from the process of coming to rely on it. GSA’s own leadership described the promotional era in exactly these terms: the deeply discounted first-year offers existed so that agencies could evaluate powerful tools before committing to lengthy acquisitions, and the shift to consumption pricing marks the moment evaluation gives way to operation.[10][1]
“As agencies increasingly integrate AI into their regular operations, providing consumption-based access is the next logical step” [1]
— Laura Stanton, Acting Commissioner, GSA Federal Acquisition Service
It is worth pausing on how quickly the experimental era compressed. OpenAI joined OneGov in August 2025 at one dollar per agency per year; Anthropic matched those terms within days for its Claude models; Google followed with discounted Gemini for Government; Microsoft folded a government-exclusive Microsoft 365 and Copilot suite into a multibillion-dollar OneGov arrangement; xAI offered Grok access for cents; Meta contributed open-weight Llama models; and by mid-2026 additional agentic platforms such as CORAS had joined the roster with discounts of up to eighty percent.[12][34] All three of the flagship model agreements — OpenAI’s, Google’s, and Anthropic’s — were structured to expire on September 30, 2026, which means the September 10 announcement is best read as the first answer to the question every observer of the program had been asking: what happens when the free samples end and the technology has to survive on an ordinary bill?[13] The answer chosen — zero license fees, half-price consumption, and a 27-month horizon — is the answer of an institution that expects usage to grow, not shrink.
1.2 From Seat Licenses to Intelligence Consumption
Traditional enterprise software is commonly purchased through licenses: a government buys a defined number of seats for a defined application, and the cost of the software is essentially fixed regardless of how intensively any given employee uses it. Frontier AI increasingly behaves differently, and the difference is not cosmetic. Under the new OneGov structure, the government purchases what can only be described as units of intelligence: tokens, inference calls, reasoning cycles, coding sessions, searches, agentic actions, multimodal analyses, and cybersecurity investigations all become measurable, meterable forms of consumption, billed in proportion to the cognitive work actually performed.[1][4] This creates a new administrative economics with a simple grammar: employees consume labor hours; datacenters consume electricity; AI-enabled governments consume intelligence.
The private sector reached this conceptual destination first, and its earnings statements now describe the phenomenon with striking directness. When NVIDIA reported its results for the second quarter of fiscal 2027 on August 26, 2026 — revenue of $96.2 billion, up 106 percent from a year earlier, with data-center revenue of $89.0 billion driven by the Blackwell Ultra ramp — its founder framed the milestone not as chip sales but as the industrialization of machine cognition itself.[26][27]
“AI has reached its inflection point. It’s doing useful work.” [27]
— Jensen Huang, Founder and CEO, NVIDIA
Microsoft’s fiscal 2026 results, announced July 29, 2026, told the same story from the platform layer: annual revenue of $331.8 billion, Microsoft Cloud surpassing $214 billion, Azure exceeding $100 billion for the first time on 41 percent growth, capital expenditures of roughly $116 billion for the year, and more than 30 million paid seats for the Microsoft 365 Copilot assistant — with the company explicitly adding usage-based billing alongside per-seat licensing because seats alone no longer capture how intelligence is actually consumed.[29][30][31]
“We are advancing the frontier on the cost-to-outcome curve” [29]
— Satya Nadella, Chairman and CEO, Microsoft
The September 10 arrangement imports this consumption logic into public finance, and that importation is strategically important in a way that a simple discount would not be. Consumption pricing means government AI expenditures can gradually migrate from experimental innovation budgets — small, discretionary, easily cancelled — toward ordinary operating expenses, the recurring line items that define what an institution actually is. A government that budgets for intelligence the way it budgets for payroll and power has, in accounting terms, already declared AI to be infrastructure. The FinOps guidance, spend controls, and usage-estimation support that OpenAI committed to provide alongside the agreement are the mundane tooling of exactly this transition: nobody builds spend-management dashboards for an experiment.[2]
1.3 When AI Becomes Horizontal Infrastructure
Most government technologies historically serve identifiable, vertical functions. Tax software serves taxation. Defense systems serve defense. Case-management systems serve the particular agencies that commissioned them, encode those agencies’ particular workflows, and are largely useless anywhere else. This verticality has shaped everything about how government buys, governs, and audits technology: the system and the mission arrive together, and oversight of one is oversight of the other. Frontier models break this pattern, because the same underlying model can perform economically meaningful work across thousands of administrative functions simultaneously. An advanced model can assist an attorney at the Justice Department in the morning, a cybersecurity analyst at Homeland Security in the afternoon, a procurement specialist at GSA later that day, and a county administrator, a tribal government employee, or a state emergency-management official through essentially the same technological foundation, differentiated only by prompts, permissions, and connected data.
That horizontality makes frontier AI increasingly resemble an administrative utility rather than an application — closer in institutional character to electricity, telecommunications, or the payments system than to any single software product. The empirical record of 2025 and 2026 supports this reading. The Stanford 2026 AI Index found that generative AI reached roughly 53 percent global population-level adoption within three years of its debut — faster than either the personal computer or the internet diffused — while organizational adoption reached 88 percent and 70 percent of surveyed organizations reported using generative AI in at least one business function.[23][24] Global corporate AI investment more than doubled in 2025 to $581.7 billion.[23] A technology that diffuses at that speed, across that many functions, at that scale of capital formation, is not a tool that institutions adopt; it is an environment into which institutions move. Administrative Default names the moment the state completes that move.
1.4 The Five-Layer AI Economy Meets the State
Administrative Default also connects directly to what this author has elsewhere called the Five-Layer AI Economy, and the connection converts an industrial framework into a theory of state capacity. Layer 1, Energy, supplies the electricity required for AI infrastructure — an input now so binding that memory scarcity and power availability feature in the forward guidance of the world’s most valuable semiconductor company.[28] Layer 2, Chips, supplies the GPUs, accelerators, networking equipment, memory, and increasingly specialized inference processors whose demand NVIDIA’s accelerating quarters make visible. Layer 3, Datacenters, provides the physical compute environment — Microsoft alone added 31 datacenters across five continents in a single quarter, bringing its yearly total to 88, in response to demand that continues to exceed available capacity.[30] Layer 4, Models, supplies the frontier systems themselves: ChatGPT and GPT-6 Astra, Gemini, Claude, Grok, Llama, and their successors. Layer 5, Applications and Agents, is where the transformation becomes visible to citizens — the chat interfaces, copilots, coding agents, cyber defenders, and workflow systems through which intelligence actually performs administrative work.
Table 1. The Five-Layer AI Economy as State-Capacity Infrastructure
| Layer | What It Supplies | Government Exposure | Illustrative 2025–2026 Evidence |
| 1. Energy | Electricity for training and inference | Permitting, grid policy, siting; operational dependence of every AI workflow | Power and memory constraints cited in chip-sector guidance[28] |
| 2. Chips | GPUs, accelerators, networking, memory | Export controls, supply security, industrial policy | NVIDIA Q2 FY27: $96.2B revenue; $89.0B data center[26] |
| 3. Datacenters | Physical compute environments | FedRAMP/IL-rated capacity; state siting and tax policy | Microsoft: 88 new datacenters in FY26; ~$116B capex[30][31] |
| 4. Models | Frontier reasoning systems | OneGov agreements; classified deployments | GSA–OpenAI 27-month deal; DoD multi-vendor IL6/IL7 pacts[1][15] |
| 5. Apps & Agents | Copilots, agents, cyber defenders | Daily administrative work; citizen-facing services | GenAI.mil: 100,000+ agents built by DoD personnel[16] |
Government may interact primarily with Layer 5, but it becomes structurally dependent upon every layer beneath it, in exactly the way that a government dependent on electric light is dependent on generation, transmission, and fuel. This is why Administrative Default makes AI infrastructure a question of state capacity rather than merely of technology policy. A state whose permitting offices, benefits agencies, cyber defenders, and courts assume the continuous availability of inexpensive machine reasoning has acquired a new critical dependency — one that runs through private energy markets, a concentrated semiconductor supply chain, hyperscale datacenter operators, and a handful of frontier laboratories. The resilience of the administrative state and the resilience of the AI stack become, from that point forward, the same subject.
1.5 The Default Reversal
These threads converge in what can be called the Default Reversal, the pivot around which this entire paper turns. During the pilot era, AI use required justification: a business case, a risk assessment, an approval chain, a designated experiment. During the Administrative Default era, non-use increasingly requires justification instead, and the demand for that justification arrives from every direction at once. If an AI-assisted procurement review takes three hours while the previous procedure required three weeks, agency managers will eventually question why the slower process survives, and their question will carry the weight of the OneGov savings figures already circulating through the budget process.[1] If cybersecurity agents can continuously examine government networks — the explicit purpose of extending discounted Daybreak Blue access to every verified government entity — then operating without continuous machine triage becomes progressively harder to defend before an inspector general reviewing the aftermath of a breach.[2] If AI can translate public documents instantly into dozens of languages, limited-language service stops looking like a resource constraint and starts looking like a choice.
Table 2. The Default Reversal: Pilot Era versus Administrative Default Era
| Dimension | Pilot Era (c. 2018–2025) | Administrative Default Era (2026–) |
| Burden of proof | Use must be justified | Non-use must be justified |
| Budget location | Innovation / experimental funds | Ordinary operating expense; consumption metering |
| Procurement form | Individual pilots, seat licenses | Governmentwide vehicles; token-based consumption[1] |
| Workforce posture | Volunteer early adopters | Assumed layer for the whole civil service |
| Oversight question | “Is this safe to try?” | “Is this governed, portable, auditable at scale?”[9] |
| Failure mode feared | Wasted pilot spending | Operational lock-in; accountability gaps[8] |
Technology, in this reversal, moves from capability to expectation — and from expectation to administrative standard. The movement is familiar from every prior wave of government modernization: no statute ever commanded agencies to answer email, yet an agency that refused would today be judged derelict. What distinguishes the AI version of this progression is its speed and its depth. The progression from one-dollar pilots to governmentwide consumption pricing took thirteen months.[3][12] And the capability being normalized is not a communications channel but cognition itself — reading, comparing, reasoning, drafting, and deciding — which is to say, the substance of administrative work rather than its packaging. The remainder of this paper examines what follows when that normalization completes: for the civil service, for procurement and federalism, for constitutional accountability, for the design of the AI-native state, and for the international competition over state capacity that will define the decade’s end.
Section 2: Building the AI Operating Layer of the State
2.1 Government Work Is Primarily Information Work
To understand why the state is so exposed to advances in machine intelligence, one must begin with an unglamorous observation about what government actually does all day. Modern states process extraordinary quantities of information; that processing is not incidental to governance but constitutive of it. Governments collect taxes, issue benefits, administer contracts, review applications, create regulations, investigate fraud, manage records, inspect infrastructure, analyze intelligence, operate courts, administer healthcare programs, respond to disasters, and defend computer networks — and every one of these functions decomposes, on inspection, into a chain of informational operations: reading, searching, classifying, comparing, reasoning, drafting, checking, routing, summarizing, and deciding. The Stanford-NYU canvass for ACUS organized federal AI use into precisely such functional categories — enforcement prioritization at the SEC, CMS, and IRS; regulatory monitoring and analysis at NOAA, the FDA, and the FCC; adjudication support at the SSA and USPTO; citizen engagement and service delivery across dozens of agencies — demonstrating that the taxonomy of administrative work maps almost perfectly onto the taxonomy of tasks at which large models are improving fastest.[19][22]
This mapping is the deep reason the Administrative Default thesis is not merely a story about one procurement agreement. The 2026 AI Index documents that frontier models now meet or exceed human baselines on PhD-level science questions, multimodal reasoning, and competition mathematics, and that performance on SWE-bench Verified — a benchmark requiring models to resolve real software issues — climbed from 60 percent toward saturation within a single year.[23] Capabilities of that character do not sit adjacent to government work; they sit inside it. A state whose daily output consists of documents read, cases compared, rules interpreted, and letters drafted is, structurally, the largest single customer for machine cognition in any economy — whether or not it has yet organized itself to act like one. The unusual exposure runs in both directions: government has the most to gain from well-deployed machine intelligence, and the most to lose from deploying it badly, because its informational decisions carry the coercive and redistributive force of law.
2.2 The Civil Servant Becomes an AI Supervisor
The near-term transformation is unlikely to eliminate the civil servant, and the honest version of this paper must resist both the utopian and the dystopian temptation on this point. What changes first is not the existence of the government employee but the employee’s production function — the mix of activities through which a unit of administrative output actually gets produced. A future government analyst may supervise multiple AI agents conducting research, reviewing datasets, preparing policy alternatives, monitoring regulatory changes, drafting reports, and identifying inconsistencies, in the way a newsroom editor supervises reporters or an audit partner supervises staff. The human employee increasingly supplies judgment, authority, institutional knowledge, legal responsibility, ethical interpretation, and final accountability; the machine supplies speed, scale, memory, search, simulation, and continuous execution. The administrative workforce consequently moves from performing every informational step itself toward supervising increasingly autonomous informational systems — a shift in kind, not merely in degree.
The evidence that this shift is already underway inside the American state is strongest where adoption has been measured most publicly: the Department of Defense. GenAI.mil, the Pentagon’s enterprise generative-AI platform, launched in December 2025 and reached more than 1.2 million unique users by April 2026, with five of the six military services designating it their default enterprise AI platform; by the end of August 2026 the department reported more than 1.7 million unique users out of a workforce of roughly three million, with a January Navy memorandum making the platform an enterprise requirement for controlled unclassified information.[16][17] Most tellingly for the supervisory thesis, Pentagon personnel had by mid-2026 built more than 100,000 discrete AI agents — automated tools constructed by ordinary employees to handle specific recurring tasks — which is to say that hundreds of thousands of public servants have already crossed from using AI to directing it.[16] Officials credited the platform with cutting task times from months to days on research, drafting, and data-analysis work.[15] Whatever one concludes about the wisdom of the rollout’s speed — and the department has published no error rates or output-quality assessments, a silence that deserves more scrutiny than it has received[18] — the direction of the workforce transformation is no longer hypothetical.
Economists studying the broader labor market reach a compatible conclusion: the productivity effects of AI are real, uneven, and contingent on how organizations choose to deploy the technology. Stanford’s Erik Brynjolfsson, who directs the Digital Economy Lab and serves on the AI Index steering committee, has argued consistently against the assumption that AI’s value lies chiefly in removing labor costs, contending instead that the larger gains come from augmenting employees — an argument with obvious force for a public sector that cannot simply exit its statutory missions the way a firm exits a product line.[37][23] At the same time, the Index records early displacement signals that public-sector workforce planners cannot responsibly ignore, including a nearly 20 percent employment decline since 2024 for software developers aged 22 to 25 and employer surveys in which one-third of respondents anticipate workforce reductions within a year.[24] The civil service that emerges from Administrative Default will be smaller in some occupations, larger in others — evaluation, oversight, AI operations, security — and different in nearly all of them.
2.3 From Copilot to Administrative Agent
Within the operating layer now being assembled, the distinction between generative AI and agentic AI becomes critical, because the two raise categorically different governance questions. A chatbot waits for instructions, produces an output, and stops; every consequential act still passes through human hands. An agent can pursue objectives: it decomposes a goal into steps, invokes tools, reads and writes to systems, and iterates until a condition is met or an exception forces escalation. An administrative agent might eventually monitor procurement deadlines, identify missing documentation, query authorized internal systems, draft correspondence, reconcile databases, prepare recommendations, and escalate only the exceptional cases to human officials — performing, in continuous background execution, work that today occupies entire branches of clerical government.
The present state of this transition is instructive precisely because it is unfinished. The 2026 AI Index finds organizational adoption of generative AI at 88 percent, yet deployment of autonomous agents still in the single digits across nearly all business functions — a gap the Index community attributes to the “jagged frontier” of capability, in which systems that perform brilliantly on intended tasks fail unpredictably on adjacent ones, and to responsible-AI benchmarking that has not kept pace, with documented AI incidents rising to 362 in 2025 from 233 the year before.[24][25] The Pentagon’s hundred thousand agents show the demand; the Index’s single-digit deployment figures show the caution; and the space between them defines the central operational question of 2027 through 2030. Between those years, this distinction — tool versus worker, output versus action — may become one of the most consequential questions facing government, because the state will not merely be acquiring models. It will be acquiring machine workers, and everything in public law that attaches to the concept of a worker — authorization, supervision, discipline, liability, records — will demand its machine analogue.
2.4 Government APIs Become Administrative Nervous Systems
Agentic government cannot function through chat windows alone, and this technical fact carries more institutional weight than it first appears to. For agents to perform administrative work rather than merely describe it, models must connect with case-management systems, financial databases, procurement systems, email, document repositories, public records, geospatial systems, cybersecurity platforms, benefits databases, and authorized external data. Application programming interfaces consequently become part of the administrative nervous system — the pathways along which perception, decision, and action travel through the body of the state. The OneGov architecture anticipates this: the agreements cover not only chat products but API and platform deployments in FedRAMP-authorized and government-cloud environments, precisely so that agency systems, not just agency employees, can consume intelligence.[35][2]
It follows that the most important government AI architecture may be neither the chatbot nor the model. It may be the authorization layer: the infrastructure that determines which artificial agent can access which system, for which purpose, under whose authority, for how long, and with what audit trail. This is where the classical concerns of administrative law — delegation, scope of authority, record-keeping, review — reappear in engineering form. Engstrom and Ho argued in their study of algorithmic accountability that the future of governing governmental AI lies less in constitutional grand theory than in the workaday machinery of administrative law, which modulates how tools are adopted, documented, and reviewed[22]; the authorization layer is that machinery’s technical embodiment. A permission scheme is a delegation doctrine written in code. An audit log is an administrative record generated at machine speed. Governments that treat identity, authorization, and logging as afterthoughts will discover that they have delegated real authority without the instruments to supervise it; governments that build these layers deliberately will possess something no paper-era bureaucracy ever had — a complete, queryable account of who and what acted in the state’s name.
2.5 The Rise of Machine-Speed Bureaucracy
Bureaucracy is most often criticized for slowness, and AI attacks that failure directly — which is precisely why the deepest risk of the transition is easy to miss. A government capable of processing one hundred times more information can also generate decisions, notices, audits, investigations, requests, and enforcement actions far faster than citizens, small businesses, and even large regulated entities can respond. Machine-speed administration therefore creates a paradox: every gain in the state’s processing capacity is simultaneously a potential gain in the state’s capacity to overwhelm the humans on the receiving end of its output. An agency that once mailed a thousand deficiency notices a month can, in principle, generate a million; the constitutional question is not whether it can but what happens to the practical meaning of notice, response, and appeal when it does.
Administrative efficiency, for this reason, cannot be measured solely by government processing speed, and a legitimate AI-enabled state must deliberately preserve human-scale rights inside machine-scale administration. Concretely, this means engineering asymmetries in the citizen’s favor: response clocks that do not accelerate merely because issuance did; consolidated rather than fragmented demands; AI assistance offered to the public for navigating AI-generated process, so that the technology that speeds the state also equips those who must answer it; and deliberate throttles on enforcement volume tied to the availability of meaningful review. The IMF’s warning about the macro-level transition applies with full force to the administrative version of it: the institution’s leadership has repeatedly emphasized that roughly 40 percent of global employment — and about 60 percent in advanced economies — is exposed to AI, and that the outcome depends on preparation rather than on the technology alone.[32]
“A tsunami is hitting the labour market” [33]
— Kristalina Georgieva, Managing Director, International Monetary Fund
A tsunami, in Georgieva’s figure, is survivable with seawalls and early warning and ruinous without them. Section 4 of this paper takes up the seawalls — the accountability, record-keeping, and escalation structures that machine-speed bureaucracy requires. But the design principle can be stated now, and it anchors everything that follows: the purpose of Administrative Default is a faster state that remains answerable at human speed, and any deployment that purchases the first at the price of the second has misunderstood the assignment.
Section 3: Federalism, Procurement, and the New Vendor–State Relationship
3.1 OneGov Becomes ManyGov
The most important element of the September 10 agreement may ultimately prove to be neither its price nor its duration but its geography. OpenAI’s offer explicitly reaches federal, state, local, and tribal governments, and the GSA framework opens participation to legislative and judicial entities alongside the executive branch — the first time a frontier-model arrangement of this scale has been architected for the whole vertical structure of American government rather than for Washington alone.[2][36] That design radically expands the potential surface area of government AI. The United States contains tens of thousands of governmental entities — states, counties, municipalities, school districts, special districts, tribal nations, courts, and public authorities — and the overwhelming majority lack the resources to develop proprietary frontier models, negotiate bespoke enterprise contracts, or staff large AI teams. Standardized purchasing changes their position entirely: a county clerk’s office or a tribal environmental department can now access, on identical terms and at half price, technological capabilities that eighteen months earlier were available only to the federal government and the world’s largest corporations. Administrative Default, in other words, will not trickle down through federalism; it is being piped down, deliberately, through procurement.
The inclusion of tribal governments deserves particular emphasis rather than a passing clause. Tribal entities often face distinctive funding constraints, procurement barriers, and connectivity challenges, and their sovereign status places them awkwardly within many federal technology programs; an arrangement that grants them the same zero-license, half-price consumption terms as a cabinet department is, quietly, one of the more significant acts of technological inclusion in recent federal practice.[2][9] Whether inclusion on paper becomes capability in practice will depend on the training, onboarding, and enablement layers — the expanded Government Academy and adoption support that accompany the agreement — and on whether smaller governments can supply the data governance and security practices that safe deployment presupposes.[2] The risk worth naming is a new administrative divide: not between governments with and without AI access, which the agreement largely dissolves, but between governments with and without the institutional capacity to use that access well.
3.2 The Governor as AI Executive
If the vertical extension of OneGov succeeds, governors become some of the most important AI executives in America, and it is worth spelling out why the gubernatorial role is structurally distinct from the federal one. States operate or decisively influence transportation, education, energy regulation, Medicaid, unemployment systems, professional licensing, environmental permitting, emergency response, economic-development incentives, public safety, and administrative workforces that collectively dwarf the federal civilian workforce. These are the programs where citizens actually meet the state — where a delayed permit stalls a housing project, a backlogged unemployment system deepens a recession’s damage, and a slow Medicaid determination becomes a family crisis. A governor who compresses those queues with well-governed AI changes more lives, more directly, than most federal deployments ever will; a governor who automates them badly will produce the era’s defining scandals.
States such as California, Texas, Pennsylvania, Michigan, Indiana, Virginia, and Arizona occupy a second, compounding position: they sit at the intersection of the Five-Layer AI Economy itself, simultaneously regulating or hosting datacenters, power generation, semiconductor facilities, transmission infrastructure, advanced manufacturing, and large technology investments. Virginia’s datacenter corridor, Texas’s grid politics, Arizona’s fabrication plants, and California’s frontier labs mean that a governor’s AI agenda is at once industrial policy (Layers 1 through 3), procurement strategy (Layer 4), and administrative modernization (Layer 5). The governor’s AI strategy consequently evolves from technology policy — a chief information officer’s portfolio — into administrative strategy: a theory of how the state itself will produce its work. The states’ famous role as laboratories of democracy acquires a literal new meaning here, and Section 6 returns to why this fragmentation may prove an underappreciated American advantage.
3.3 Procurement Becomes Architecture
Government procurement rules were historically designed to answer a bounded question: who may sell what to the government, at what price, through what competition? AI procurement must answer much more, because what is being purchased is not an artifact but an ongoing relationship with a system that learns, changes, and accumulates institutional context. Which model receives government data, and under what training and retention commitments? Can workloads migrate to another provider without reconstruction? Who owns prompts, outputs, embeddings, fine-tuning artifacts, and agent histories? Can government retrieve its institutional knowledge if a vendor fails, is acquired, or is excluded? Can one provider’s agents interact with another provider’s models? Can an agency switch platforms without rebuilding its entire AI workflow from the authorization layer up? Every one of these questions is simultaneously a contract clause and a systems-architecture decision, which is the sense in which procurement has become architecture.
Federal policy has, to its credit, recognized the problem in advance of the crisis. OMB M-25-22 requires agencies to consider vendor lock-in across the entire acquisition lifecycle, encouraging contract terms that mandate knowledge transfer, data and model portability, and transparency, while preserving government rights to data and outputs; M-25-21 directs agencies to prefer interoperable products precisely to sustain a competitive federal AI marketplace.[7][8][9] The September 10 agreement itself reflects this posture in its data commitments — ChatGPT Enterprise does not use government business data, including inputs and outputs, to train or improve models, and participating organizations retain the protections of the services they select.[4][2] These are necessary provisions. Whether they are sufficient is the subject of the next subsection, because the deepest dependency risks are not the kind that contract language alone can reach.
3.4 The Vendor–State Dependency Problem
The greatest commercial opportunity created by Administrative Default is also one of its greatest governance risks, and both should be stated plainly. Frontier laboratories are acquiring extremely deep institutional relationships with governments: OpenAI, Google, Microsoft, Amazon, NVIDIA, xAI, Oracle, Meta, Anthropic, and future providers are becoming embedded within different portions of government activity, from unclassified productivity platforms to classified analytical networks.[12][15] The problem this creates is not simply monopoly in the antitrust sense; on paper, the government’s AI marketplace is more plural than most federal software categories have ever been. The deeper issue is administrative dependence. If thousands of workflows, agents, databases, security processes, and bodies of accumulated prompt-craft and institutional memory become optimized around one model ecosystem, replacing the provider can become prohibitively expensive even when the underlying procurement contract remains formally competitive. The state can become locked in operationally before it is locked in contractually — and operational lock-in never appears on a contract-award notice.
The events of early 2026 supplied a live demonstration of how abruptly a vendor relationship can rupture, and of what rupture costs when a model is widely embedded. The Pentagon’s dispute with Anthropic over usage guardrails — the company insisted on restrictions concerning surveillance and weapons applications as negotiators pressed for unrestricted-purpose language — ended with the department labeling a widely used supplier a supply-chain risk and barring its tools across the defense enterprise, even as officials acknowledged the company’s cyber-capable Mythos model as a separate national-security concern.[14][15] Whatever position one takes on the merits of that confrontation — and reasonable observers have taken sharply different ones, since it pits a vendor’s safety commitments against the government’s demand to define scope of use — its structural lesson is independent of the merits: agencies that had woven a particular model into daily work faced immediate transition costs measured in retraining, workflow reconstruction, and lost context. Dependency risk is not hypothetical; it has already been exercised.
3.5 The Government Model Portfolio
The solution emerging in practice resembles nothing so much as financial portfolio management, and the analogy is worth taking seriously rather than decoratively. Rather than relying exclusively on one frontier laboratory, agencies can maintain diversified portfolios of models: one for general reasoning, another for cybersecurity, another for scientific analysis, another for coding; open-weight models for highly controlled internal environments where data cannot leave government custody; and specialized small models for repetitive, well-characterized workloads where frontier capability would be wasted expense. The objective is not simply lower cost through competition, though consumption pricing makes cost comparison newly tractable. The objective is administrative resilience: the guarantee that no single vendor’s outage, acquisition, price change, capability regression, or policy collision can halt the business of government.
Table 3. A Government Model Portfolio: Illustrative Allocation Logic
| Portfolio Slot | Workload Character | Sourcing Logic | Resilience Function |
| General reasoning | Drafting, analysis, briefings | Frontier model via OneGov consumption terms[1] | Primary capacity; benchmarked quarterly |
| Cyber defense | Vulnerability discovery, alert triage | Specialized system (e.g., Daybreak-class)[2] | Continuous defense independent of office tools |
| Coding & software | Agency systems, automation | Coding-optimized models/agents[35] | Keeps modernization independent of chat vendor |
| Sensitive/internal | Data that cannot leave custody | Open-weight models in gov environments[12] | Sovereignty hedge; exit insurance |
| High-volume routine | Classification, extraction, routing | Small specialized models | Cost control; graceful degradation |
| Classified/analytic | IL6/IL7 workloads | Multi-vendor classified agreements[15] | No single point of national-security failure |
The Pentagon’s 2026 strategy provides the early illustration at the largest possible scale. In May 2026 the department announced agreements with eight frontier companies — SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle — to deploy their models on classified IL6 and IL7 networks, explicitly spreading data synthesis, situational understanding, and decision support across a diverse AI stack rather than relying on any single vendor, and building atop a GenAI.mil platform that had already reached more than 1.3 million users in its first five months.[14][15] Reuters and subsequent reporting made clear that diversification was not incidental but the stated design principle of the classified rollout.[14] Government AI architecture, in short, is converging on the same redundancy principles long required in energy, telecommunications, banking, and defense procurement — the domains where society decided, after hard experience, that critical infrastructure must never have a single point of failure. Administrative Default makes intelligence critical infrastructure; the portfolio is what treating it that way looks like.
Section 4: Cybersecurity, Legitimacy, and the Constitutional Limits of Machine Administration
4.1 AI Becomes Both Administrative Tool and Attack Surface
The September 10 announcement deliberately connects routine government AI adoption with cybersecurity, and that connection deserves sustained attention rather than a nod, because it captures the dual character of the entire transition. As governments become more dependent upon models, agents, APIs, and automated workflows, those very systems become attractive attack surfaces — and they are attack surfaces of a genuinely new kind, because what is being attacked is not merely data at rest but reasoning in motion. Prompt injection can turn a helpful agent into an unwitting insider; poisoned training or retrieval data can bias the analytical layer on which thousands of decisions rest; compromised agent credentials inherit every permission the authorization layer granted; manipulated retrieval systems can feed confident falsehoods into official work products; model extraction can exfiltrate capability itself; malicious tools can weaponize the agent’s own ability to act; and autonomous cyber operations compress the timeline of attack below the timeline of human response. Each of these vectors targets precisely the capacities — reading, trusting, acting — that make administrative AI valuable in the first place. The AI that improves government efficiency thereby simultaneously enlarges the government’s digital attack surface, and any accounting of Administrative Default that records the first effect without the second is propaganda rather than analysis.
The concentration effects examined in Section 3 compound the exposure. When 1.7 million defense personnel work through a single platform, and when tens of thousands of civilian entities converge on a handful of model providers, the compromise of any shared layer — a model, an orchestration service, an identity system, a widely reused agent template — propagates with a speed and breadth that the fragmented, inefficient legacy environment, whatever its other failings, structurally resisted.[17][12] Homogeneity is efficient, and efficiency is contagious in both directions. This is the security argument for the model portfolio of Section 3.5, restated: diversity in the AI stack is not only procurement hygiene but epidemiological defense.
4.2 Daybreak and the Cybersecurity State
OpenAI’s decision to bundle discounted Daybreak Blue access for every verified government organization — at fifty percent off commercial pricing, with scaled training so that public-sector defenders can locate vulnerabilities before adversaries exploit them, and with the more offensive-capable Daybreak Red available to government partners at standard pricing — demonstrates how quickly frontier-model providers are moving beyond productivity applications toward national cyber-defense infrastructure.[2][1] The significance of this move is easy to understate. It suggests that the frontier AI laboratory is becoming, simultaneously, a government software vendor, a cybersecurity contractor, an intelligence-infrastructure provider, an enterprise platform, and a strategic national-security partner — five roles that the American state has historically procured from different industries, governed under different legal regimes, and overseen through different committees. Their convergence in single firms deserves far greater scrutiny than ordinary software-as-a-service procurement receives, because the leverage that accompanies each role compounds the leverage of the others: the vendor that defends the network also observes it; the platform that hosts the workflow also shapes it; the partner whose model is strategically indispensable negotiates from a position no office-software supplier ever occupied.
None of this is an argument against the arrangement; a government facing AI-accelerated adversaries can hardly decline AI-accelerated defense, and the case for putting frontier defensive capability into the hands of chronically under-resourced state, local, and tribal security teams is overwhelming on its own terms.[2] It is an argument for governing the arrangement with instruments proportionate to its novelty: security-evaluation regimes for the defensive models themselves, contractual separation between defensive telemetry and commercial product development, congressional visibility into the cumulative footprint of each provider across the five roles, and the multi-vendor redundancy that the Pentagon’s classified strategy has already made doctrine.[15] The cybersecurity state being assembled around frontier models will either be governed deliberately at its founding or retrofitted painfully after its first crisis; there is no third schedule.
4.3 Human Accountability Cannot Be Outsourced
No matter how capable AI becomes, constitutional government cannot allow responsibility to dissolve into an algorithm, and this principle — easy to affirm, demanding to implement — is the moral center of the Administrative Default era. A citizen denied a benefit should be able to learn why, in terms a person can contest. A company denied a permit should have a path of appeal that terminates in an accountable official rather than in a model version. An individual investigated by government retains legally established protections that do not thin merely because the investigation was machine-prioritized. A contractor rejected from a procurement should know which authority made the determination and on what record. Administrative Default therefore requires a simple organizing principle: automation may assist authority; it cannot erase accountability. The federal framework already gestures in this direction — M-25-21 preserves requirements surrounding privacy, civil rights, civil liberties, public trust, and risk management, and its high-impact AI category attaches heightened documentation and oversight duties exactly where outputs principally determine consequential decisions.[6][7] The scholarly foundation is older still: the ACUS study warned from the outset that agencies adopting these tools would confront hard questions about the respective scope of human and machine decision-making, and about accountability, transparency, and non-discrimination.[19][21]
“this report unearths the current broad and uncoordinated use of AI by our government” [21]
— Daniel E. Ho, Stanford Law School
What was “broad and uncoordinated” at the moment Ho and his colleagues surveyed it has since become broader and — through OneGov, the OMB memoranda, and agency AI strategies — considerably more coordinated. The unresolved question is whether coordination of adoption will be matched by coordination of accountability. The honest answer in September 2026 is: not yet, and not automatically. Accountability structures are built by institutions under pressure to move fast, and the entire design of the current moment is pressure to move fast. That is exactly why the principle must be stated as a constraint rather than an aspiration, and why the next two subsections translate it into the two concrete institutions on which it depends: the record and the escalation right.
4.4 The Administrative Record Must Become an AI Record
Administrative law runs on records. Judicial review of agency action is review of the record; the reasoned-decision-making requirement is, in practice, a requirement that the record disclose the reasoning; and the entire apparatus of administrative accountability presupposes that what the agency considered, and why it concluded as it did, can be reconstructed after the fact. AI complicates the concept of a record at its foundation, because when analysis flows through a model, the traditional record — final documents, signed determinations, docketed submissions — no longer captures how the conclusion was actually produced. Engstrom and Ho’s work on algorithmic accountability demonstrated that conventional ex ante and ex post review doctrines strain badly against algorithmic governance tools, and argued that administrative law, not constitutional law, would have to carry the adaptive burden.[22] The adaptation begins with the record itself.
Future agencies may need to preserve, for consequential decisions, not merely final documents but a computational layer of the record: which model was used and which version; which system instructions and configuration governed it; which government data sources were accessed and which external tools were called; whether an agent acted autonomously at any step and under whose authorization; what confidence, uncertainty, or dissent the system expressed; which human reviewed and approved the result; and whether the model changed between initial analysis and final decision — a question with no analogue in the paper era, when the reasoning instrument did not receive weekly updates. The administrative record of the AI era, in short, becomes partly computational, and the audit trails discussed in Section 2.4 are not merely security instrumentation but the raw material of legality. There is a quiet optimism available here that critics of automation rarely concede: a well-instrumented AI process can be more reviewable than the human process it replaced, because human deliberation never logged itself. Whether that potential is realized depends entirely on whether record-generation is engineered in from the start — which is to say, on procurement, which is to say, on the architecture decisions being made right now under M-25-22’s lifecycle requirements.[9]
4.5 The Right to Human Escalation
One safeguard deserves elevation above the general catalogue: a right to human escalation for consequential government decisions. The logic of Administrative Default is that AI appropriately absorbs enormous quantities of routine administrative work — the classification, extraction, drafting, and reconciliation that consume the civil service’s hours without exercising its judgment. But when decisions materially affect liberty, benefits, employment, taxation, immigration, healthcare, licensing, property, or legal status, citizens require meaningful access to accountable human review — not a human rubber stamp appended to a machine conclusion, and not a nominally available appeal buried beneath machine-speed process, but a genuine reconsideration by an official empowered to disagree with the system and answerable for the result. The high-impact AI framework of M-25-21 supplies the natural doctrinal home for such a right, since it already isolates the category of systems whose outputs principally determine consequential outcomes[7]; what remains is to convert documentation duties into an enforceable individual entitlement.
The design challenge is locating the boundary between efficiency and legitimacy honestly, because an escalation right defined too broadly simply reconstitutes the pre-AI queue — with its months of delay that were themselves a due-process failure — while a right defined too narrowly becomes ornamental. The resolution lies in asymmetry: let machines handle volume, let humans handle stakes, and let the definition of stakes be set publicly, in advance, through the ordinary instruments of rulemaking rather than discovered privately in the exception-handling logic of a vendor’s workflow product. Administrative Default works only if government becomes faster without becoming less accountable; the escalation right is where that sentence stops being rhetoric and becomes an architecture requirement.
Section 5: 2027–2030 — The AI-Native State
5.1 From Digital Government to AI-Native Government
Each preceding wave of government technology changed where administrative work happened while leaving its essential production process intact. Digitization put government forms online, but a digital form was still a form, completed by a citizen and processed by an official. Cloud computing moved infrastructure out of agency basements, but the applications running on rented servers embodied the same workflows as before. AI changes something categorically deeper: it changes how administrative work itself is produced, entering the workflow not as a new container for human labor but as a new source of the cognitive labor itself. An AI-native government is therefore not one that simply possesses ChatGPT accounts — by that trivial standard, the American federal government became AI-native during the promotional year, when 3.5 million employees gained access.[3] An AI-native government is one whose workflows are designed around the assumption that inexpensive machine reasoning is continuously available: whose forms are designed to be read by extraction agents, whose case files are structured for machine synthesis, whose review chains presume a first-pass machine analysis, and whose exception paths are engineered for the minority of matters that genuinely require human deliberation. That is a fundamentally new institutional architecture, and between 2027 and 2030 the governments that merely bought AI will diverge, visibly and measurably, from the governments that rebuilt around it.
5.2 The Administrative Intelligence Budget
Public finance will register the transition before political rhetoric does. Today agencies budget for employees, contractors, software, cloud infrastructure, facilities, and communications; the intelligence consumed by the organization has never been a line item because it arrived bundled inside salaried human beings. Consumption-priced AI unbundles it. Under the OneGov structure, an agency’s intelligence consumption is metered in tokens and billed monthly, which means that for the first time in administrative history, machine cognition appears in the ledger as a distinct, elastic, manageable input.[1][4] Future agencies may consequently budget explicitly for machine intelligence: token budgets and agent budgets, inference allocations by mission, premium reasoning reserves for complex analytical work, cybersecurity-agent capacity, and model-routing infrastructure that steers each task to the cheapest system adequate to it. The FinOps disciplines that enterprises developed for cloud spending — and that OpenAI’s government support materials now explicitly import into the public sector[2] — become instruments of public financial management. Government will need to know not merely how many workers it employs but how much intelligence each mission consumes; and legislatures, in turn, will discover that appropriating intelligence is a new and consequential form of policy choice, since the missions granted generous token budgets will simply be able to think more than the missions denied them.
5.3 Administrative Productivity Becomes Measurable
One of the most politically important consequences of the metered state is that administrative productivity becomes measurable with an accounting precision it has never had. Governments will increasingly be able to compare processing time before and after AI assistance, human hours saved, backlogs reduced, fraud detected, citizen response times, software-development cycles, cybersecurity vulnerabilities discovered, and — the summary statistic of the era — cost per administrative transaction, decomposed into its human and machine cognitive inputs. The early public numbers already sketch the genre: GSA’s $1.68 billion in OneGov savings, the Pentagon’s months-to-days task compression, GenAI.mil’s adoption curve published like a growth chart.[1][15][17]
Measurement of this kind creates opportunity and danger in equal measure, and the paper would be incomplete if it recorded only the first. Well-designed measurement could expose genuine improvements, direct investment toward what works, and give the public an unprecedented view into the machinery it funds. Poorly designed metrics could reward agencies for automating easy tasks while degrading the difficult services that require human judgment; could convert the citizen-facing exception — the complicated case, the non-standard applicant, the person who needs a conversation — into a cost anomaly to be minimized; and could recapitulate, at the scale of the state, the oldest pathology of performance management: the metric becomes the mission. The Pentagon’s own rollout illustrates the asymmetry, publishing user counts and speed gains while disclosing no error rates or output-quality assessments — adoption measured meticulously, accuracy not at all.[18] The corrective is not less measurement but complete measurement: every efficiency dashboard paired with quality, error, equity, and appeal-outcome indicators, so that the state optimizes for administration that is better, not merely administration that is faster and cheaper on the tasks easiest to count.
5.4 The State Becomes a Model Buyer — and Model Shaper
Government procurement at this scale does not merely select from the market; it shapes what the market builds, and this reflexive power is among the least appreciated instruments of AI governance available to democracies. When governments specify requirements for security, auditability, interoperability, data portability, model behavior, logging, cybersecurity, identity, and human oversight — as M-25-22’s lifecycle requirements and the OneGov terms already begin to do — vendors serving a 23-million-person addressable public workforce have powerful commercial incentives to build those capabilities into their products, whereupon the capabilities become available to every private customer as well.[9][2] The state therefore does not merely consume frontier AI; through procurement, it indirectly shapes frontier AI. This is why government purchasing standards could become as consequential as conventional AI regulation, and in some respects more so: regulation constrains from outside and invites resistance, while procurement pays from inside and invites compliance. A requirement that appears in a governmentwide vehicle — for exportable audit logs, for model-version pinning, for portability of fine-tuning artifacts — becomes industry infrastructure within a product cycle. The governments of 2027–2030 that understand this will write their values into the stack; those that do not will inherit the values the market wrote for them.
5.5 Administrative Default and the International Competition for State Capacity
The final dimension is geopolitical, and it reframes the entire technology race of the decade. The strategic AI competition between the United States and China will not ultimately be measured only by which country possesses the fastest GPU, the largest model, or the biggest datacenter — the Layer 1-through-4 scoreboard that dominates headlines and, as the 2026 AI Index records, has effectively converged, with the leading American model ahead of the best Chinese systems by a margin of just 2.7 percent as of March 2026 after repeated exchanges of the lead.[23] When model capability converges, differentiation migrates to deployment: which state learns to use artificial intelligence most effectively across its institutions. The crucial competition becomes concrete and administrative. Which government issues permits faster? Which detects cyberattacks sooner? Which processes infrastructure projects faster? Which military moves information more efficiently — the explicit ambition of the Pentagon’s declared transformation toward an AI-first force?[15] Which tax system detects fraud more accurately? Which scientific agencies convert research into policy sooner? Which country coordinates energy, chips, datacenters, models, and applications — all five layers — most effectively in support of its public purposes?
The Five-Layer AI Economy therefore intersects directly with state capacity, and the intersection yields this paper’s longest-range claim: by 2030, the strongest AI nation may not merely be the country with the strongest AI industry. It may be the country whose government learned how to operate with intelligence most effectively — while retaining, and this qualification is not decoration, the legitimacy that makes its operations durable. An autocracy can adopt machine-speed administration without the constraints of Section 4; the democratic wager of Administrative Default is that accountable machine-speed administration, though harder to build, proves stronger over time, because it compounds public trust rather than spending it. The American federal experiment now underway — 3.5 million federal users, a multi-vendor classified stack, fifty-plus state and thousands of local laboratories about to come online under the extended OneGov terms — is, among other things, the test of that wager.[3][15][2]
Section 6: What Have We Learned?
Six sections of argument reduce to seven pillars — the durable lessons that survive the specifics of any single agreement, administration, or vendor. They are offered not as predictions, which the pace of this field humbles quickly, but as structural findings about how frontier AI and the machinery of the state are now joined.
Pillar 1 — AI Is Becoming Administrative Infrastructure
The central lesson is that frontier AI is crossing the boundary between application and infrastructure, and that the crossing is observable in the mundane evidence of budgets, contracts, and adoption curves rather than in speculation. When millions of government employees can access advanced models through standardized governmentwide agreements — 3.5 million federal users during the promotional year, an addressable public workforce of roughly 23 million under the extended terms — AI no longer belongs to innovation laboratories; it enters the everyday machinery of administration.[3][2] Infrastructure has a signature: it is priced by consumption, assumed by workflows, invisible when it works, and catastrophic when it fails. The September 10 agreement gave government AI the first property; the 2027–2030 buildout will confer the rest. Administrative Default starts when intelligence becomes an assumed resource, and by every institutional indicator assembled in this paper, that assumption is now being poured like concrete.
Pillar 2 — Procurement Policy Is Becoming AI Industrial Policy
Government purchasing is no longer a downstream activity that merely acquires what the market has already decided to build. Procurement decisions now determine which AI companies achieve scale in the world’s largest institutional market, which security standards become normal, which interoperability requirements spread through the industry, and which architectures agencies — and, by imitation, enterprises — adopt. OMB’s insistence on competition, portability, interoperability, and the avoidance of costly single-vendor dependency is therefore not administrative housekeeping; it is emerging AI industrial policy, conducted through contract clauses rather than statutes.[8][9] The corollary cuts in both directions: a wisely designed governmentwide requirement can raise the safety and accountability floor of the entire industry, while a carelessly designed one can entrench incumbents and freeze architectures for a decade. The contracting officer has quietly become one of the more consequential AI policymakers in the American system, and neither the training pipeline nor the oversight structure of the acquisition workforce has yet caught up to that fact.
Pillar 3 — Federalism Will Multiply AI Adoption
The next wave of government AI will not remain confined to Washington, because the September 10 architecture was expressly built to escape it: state, local, tribal, legislative, and judicial entities can all draw on the same zero-license, half-price consumption terms through multiple channels.[2][36] Tens of thousands of governments — states, counties, cities, tribal nations, courts, public universities, emergency agencies — can now access frontier capabilities without building frontier infrastructure, and each will embed those capabilities in different statutes, cultures, and constraints. This creates an enormous distributed laboratory for administrative AI: different governments experimenting with different models, rules, workloads, and safeguards, generating the comparative evidence about what works that no single national deployment could ever produce. The United States’ famously fragmented governmental structure — long lamented as a modernization handicap — could paradoxically become an innovation advantage, provided the experiments are observed, evaluated, and shared rather than merely conducted. Building that evaluation layer — the institutional memory of the fifty-state experiment — is among the highest-return investments available to federal AI policy.
Pillar 4 — Administrative Intelligence Requires Constitutional Guardrails
Faster administration is not automatically better government, and nothing in this paper’s enthusiasm for state capacity should be read as forgetting it. AI systems that touch benefits, liberty, licenses, taxes, and legal status must preserve accountability, appeal, privacy, civil liberties, transparency, cybersecurity, and identifiable human responsibility — the requirements that the federal framework already names and that Section 4 translated into the two indispensable institutions of the era: the computational administrative record and the right to human escalation.[6][7][22] The critical policy challenge is therefore not choosing between adoption and regulation, a framing that flatters both camps while illuminating nothing. It is designing institutions capable of achieving high administrative intelligence and high democratic legitimacy simultaneously — machine-scale processing that remains answerable at human scale. The governments that solve this design problem will not merely avoid scandal; they will discover that legitimacy is itself a performance advantage, because administration that citizens trust generates compliance, cooperation, and data quality that no enforcement budget can buy.
Pillar 5 — The Vendor–State Relationship Is a New Object of Governance
Between the market and the state, this period has created a third thing: the embedded frontier vendor — simultaneously software supplier, cybersecurity contractor, platform operator, and strategic partner — whose products carry the daily cognition of the government it serves.[2][15] The Anthropic–Pentagon rupture of early 2026 demonstrated that this relationship can fracture over values as readily as over price, and that when it fractures, the costs land on agencies whose workflows had quietly grown around the departed model.[14] The lesson is that the vendor–state relationship must be governed as deliberately as the technology itself: through portfolio diversification, through exit-tested portability, through transparency about each provider’s cumulative governmental footprint, and through honest advance deliberation — before the crisis, not during it — about whose terms prevail when a vendor’s stated principles and a government’s stated requirements collide. A state that has not answered that question in peacetime will answer it badly under pressure.
Pillar 6 — Measurement Will Decide Whether the Default Deserves Its Name
Consumption pricing makes the state’s cognition legible for the first time — metered, budgeted, and comparable across agencies and years — and what becomes legible becomes governable, for better and for worse.[1][4] The productivity story of Administrative Default will be told in dashboards: backlogs, cycle times, cost per transaction, vulnerabilities found, hours redeployed. The legitimacy story must be told in the same dashboards or it will not be told at all: error rates, appeal outcomes, equity of treatment across populations, the fate of the complicated case. The Pentagon’s adoption metrics without accuracy metrics are the cautionary template.[18] The pillar, stated as a rule: no efficiency indicator without its paired quality indicator, published on the same page, audited by the same offices. A default that cannot demonstrate it made government better — and not merely faster — has not earned the permanence the word implies.
Pillar 7 — State Capacity Becomes an AI Competition
The deepest implication extends beyond productivity to power. Artificial intelligence changes the effective capacity of the state itself, and capacity — the ability to perceive, decide, and act — is the substrate of every other national advantage. Countries able to convert frontier models into faster infrastructure approvals, stronger cybersecurity, better public services, superior military logistics, more productive scientific institutions, and more responsive administration may gain a strategic edge independent of raw model benchmarks — an edge that matters more, not less, as benchmark leadership converges to margins of a few percent.[23] The AI race therefore expands from its familiar question — who builds the best model? — to its consequential one: who builds the most capable AI-enabled society and state? The IMF’s framing of AI as a new industrial revolution applies with special force here, because industrial revolutions have always ultimately been won by the societies that reorganized their institutions around the new technology, not merely by those that manufactured it.[32] Administrative Default is the name of that reorganization as it reaches the state.
Conclusion: Administrative Default
September 10, 2026 may eventually look less important because of the price of a particular ChatGPT government agreement than because of the institutional direction it revealed. A 27-month term, a zero platform fee, discounted consumption pricing, no minimum commitment, FedRAMP-compatible access, state and local eligibility, tribal-government participation, structured training, and advanced cybersecurity availability collectively remove nearly every barrier that historically separated government AI experiments from everyday operations.[1][2][4] The three great promotional pilots — OpenAI’s, Google’s, and Anthropic’s dollar-and-cents offers of 2025 — expire in the same month this paper is written, and the first of their successors has chosen the economics of permanence.[13] The technological story has therefore become an administrative story, and the administrative story is the one this paper has tried to tell.
Frontier artificial intelligence is moving downward through the Five-Layer AI Economy into the foundations of governance. Energy enables chips. Chips populate datacenters — at a pace that produced $89 billion of data-center revenue for a single supplier in a single quarter.[26] Datacenters run models — across clouds whose government-dedicated capacity now spans unclassified enterprise platforms and IL7 classified networks alike.[15][35] Models power applications and agents — by the hundred thousand, built by ordinary public servants.[16] And those agents are now entering the institutions through which governments tax, spend, regulate, protect, investigate, communicate, plan, procure, and serve.
The most consequential transformation will probably not arrive with a presidential announcement declaring that America has become an AI government. It will happen quietly, in the accumulating texture of ordinary work. A federal analyst will stop manually reading thousands of pages because an AI agent already summarized them, and will spend the recovered day on the judgment the summary cannot supply. A county employee will stop translating forms individually because translation became automatic, and a resident who never received service in her language will simply receive it. A cybersecurity team will stop reviewing every alert manually because agents continuously triage threats, and will hunt the adversaries the triage surfaces. A procurement office will compare hundreds of bids automatically; a governor’s staff will simulate policy alternatives before the morning meeting; an agency lawyer will ask an internal model to trace every relevant rule, case, contract, and precedent; a government software team will generate and test code alongside machine collaborators. Eventually these behaviors become so routine that employees will barely describe them as “using AI.” They will simply describe them as working.
That is why Administrative Default fits this paper, and why its two words carry its whole argument. “Administrative” locates the transformation inside the machinery of government — the agencies, records, budgets, contracts, and daily decisions through which the state actually exists — rather than in abstract debates about artificial intelligence. “Default” identifies the critical threshold at which adoption no longer depends upon extraordinary permission, pilot funding, specialized expertise, or experimental justification. A default is what remains when a technology stops feeling exceptional. The defining public-sector AI question of the late 2020s is therefore no longer whether government should experiment with artificial intelligence. It is whether a modern government can operate effectively — and legitimately, and securely, and accountably — without it. That transition, from optional experiment to ordinary institutional assumption, is Administrative Default. The experiments are ending. The defaults are being written now, in procurement clauses and authorization layers and budget categories — and the governments, scholars, and citizens who engage with those unglamorous instruments in the next four years will decide what kind of state the default produces.
Endnotes:
[1] Nextgov/FCW (E. Graham). “GSA unveils new, token-based OneGov discount with OpenAI,” September 2026 (reporting GSA press release; Laura Stanton quotation; $1.68B OneGov savings; $1.4B AI-related; terms and eligibility). https://www.nextgov.com/acquisition/2026/09/gsa-unveils-new-token-based-onegov-discount-openai/415908/
[2] OpenAI. “Expanding AI access and cyber defense for federal, state, local, and tribal governments,” September 10, 2026 ($0 license fee; 50% usage discount; 27-month term; Daybreak Blue; GPT-6 Astra; ~23 million eligible public servants; Government Academy; data-use commitments). https://openai.com/index/expanding-ai-access-us-government/
[3] Bloomberg News. “OpenAI Gives US Agencies 50% Off Models, Ending $1 Per Year Deal,” September 10, 2026 (3.5 million federal users; steepest GSA technology discount; $1.4B savings). https://www.bloomberg.com/news/articles/2026-09-10/openai-gives-us-agencies-50-off-models-ending-1-per-year-deal
[4] Securities.io (M. Tan). “GSA and OpenAI Set 27-Month OneGov ChatGPT Deal With $0 License Fee,” September 2026 (October 1, 2026 – December 31, 2028 term; FedRAMP-authorized environments; no minimums; enterprise data-training commitments). https://www.securities.io/gsa-and-openai-set-27-month-onegov-chatgpt-deal-with-0-license-fee/
[5] U.S. General Services Administration. “GSA Announces New Partnership with OpenAI, Delivering Deep Discount to ChatGPT Gov-Wide Through MAS,” August 6, 2025 (OneGov strategy; America’s AI Action Plan alignment; M-25-21/M-25-22 support; Altman statement). https://www.gsa.gov/about-gsa/newsroom/news-releases/gsa-announces-new-partnership-with-openai-delivering-deep-discount-to-chatgpt-08062025
[6] Hunton Andrews Kurth LLP. “OMB Issues Revised Policies on AI Use and Procurement by Federal Agencies,” April 2025 (analysis of M-25-21 and M-25-22; EO 14179 implementation; lock-in protections). https://www.hunton.com/privacy-and-cybersecurity-law-blog/omb-issues-revised-policies-on-ai-use-and-procurement-by-federal-agencies
[7] Wiley Rein LLP. “Trump Administration Revamps Guidance on Federal Use and Procurement of AI,” April 2025 (high-impact AI category; agency AI strategies; anti-lock-in solicitation provisions). https://www.wiley.law/alert-Trump-Administration-Revamps-Guidance-on-Federal-Use-and-Procurement-of-AI
[8] Covington & Burling LLP, Inside Government Contracts. “OMB Issues First Trump 2.0-Era Requirements for AI Use and Procurement by Federal Agencies,” April 2025 (vendor lock-in across the AI acquisition lifecycle). https://www.insidegovernmentcontracts.com/2025/04/omb-issues-first-trump-2-0-era-requirements-for-ai-use-and-procurement-by-federal-agencies/
[9] Regulations.AI. “OMB Memorandum M-25-22: Driving Efficient Acquisition of Artificial Intelligence in Government” (lifecycle acquisition approach; portability; government rights to data and outputs; GSA coordination). https://regulations.ai/regulations/RAI-US-NA-OMMDEXX-2025
[10] Fox Business (A. Nikolic). “GSA and OpenAI partner to give US federal workers 50% AI discount,” September 2026 (GSA Administrator Edward C. Forst and Laura Stanton statements on OneGov strategy and try-before-buying rationale). https://www.foxbusiness.com/technology/trump-admin-partners-openai-equip-federal-employees-artificial-intelligence-tools
[11] Nextgov/FCW. “Nearly 3.4M users across government can use AI through OneGov, GSA official says,” May 2026 (Deputy Administrator Mike Lynch; $1.15B interim savings figure). https://www.nextgov.com/artificial-intelligence/2026/05/nearly-34m-users-across-government-can-leverage-ai-through-onegov-gsa-official-says/413588/
[12] Potomac Officers Club. “OneGov Has Surpassed $1.4B in AI Savings. What’s Next for the GSA Initiative?,” September 2026 (timeline of OpenAI, Anthropic, Google, Microsoft, Meta, xAI, Perplexity, and CORAS OneGov agreements). https://www.potomacofficersclub.com/articles/onegov-ai-gsa-openai-xai-google-anthropic/
[13] Startup Fortune. “OpenAI Ends Its $1 ChatGPT Deal With the US Government,” September 2026 (FedScoop reporting on simultaneous September 30, 2026 expiration of OpenAI, Google, and Anthropic OneGov deals). https://startupfortune.com/openai-ends-its-1-chatgpt-deal-with-the-us-government/
[14] NBC News (with Reuters reporting). “Pentagon reaches agreements with leading AI companies,” May 2026 (multi-vendor classified agreements; GenAI.mil surpassing 1.3 million personnel; Anthropic supply-chain-risk designation and guardrail dispute). https://www.nbcnews.com/tech/tech-news/pentagon-reaches-agreements-leading-ai-companies-rcna343071
[15] WinBuzzer (M. Kroker). “Pentagon Clears 8 AI Firms for Classified IL6/IL7 Networks,” May 3, 2026 (SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, AWS, Oracle; AI-first fighting force statement; months-to-days task compression). https://winbuzzer.com/2026/05/03/pentagon-classified-ai-agreements-nvidia-microsoft-aws-google-openai-spacex-oracle-reflection-xcxwbn/
[16] Military.com. “1.2M US Troops, Pentagon Civilians Are Using GenAI.mil — What That Is, What It Means,” 2026 (adoption trajectory; five of six services designating default platform; 100,000+ AI agents built by personnel). https://www.military.com/1-million-us-troops-pentagon-civilians-are-using-genaimil-what-that-is-what-it-means
[17] Military Times (T. Vergun). “The military’s ChatGPT is now live via the Pentagon’s GenAI platform,” August 31, 2026 (1.7 million unique users; Navy CUI/IL5 enterprise mandate). https://www.militarytimes.com/industry/techwatch/2026/08/31/the-militarys-chatgpt-is-now-live-via-the-pentagons-genai-platform/
[18] The Next Web (A. M. Constantin). “The Pentagon’s AI platform went from 80,000 users to 1.5 million in six months,” June 2026 (absence of disclosed error rates and quality assessments; adoption dynamics). https://thenextweb.com/news/pentagon-genai-mil-1-5-million-users-google-gemini-military-ai
[19] David Freeman Engstrom, Daniel E. Ho, Catherine M. Sharkey & Mariano-Florentino Cuéllar. Government by Algorithm: Artificial Intelligence in Federal Administrative Agencies, Report to the Administrative Conference of the United States, February 2020 (Stanford RegLab). https://reglab.stanford.edu/publications/government-by-algorithm/
[20] Administrative Conference of the United States. “ACUS, Stanford Law School, and NYU School of Law Announce Report on Artificial Intelligence in Federal Agencies,” February 2020 (Engstrom and Cuéllar statements). https://www.acus.gov/newsroom/news/acus-stanford-law-school-and-nyu-school-law-announce-report-artificial-intelligence
[21] Stanford Law School. Press announcement of the ACUS report, February 2020 (Daniel E. Ho statement on broad and uncoordinated federal AI use). https://law.stanford.edu/press/acus-stanford-law-school-and-nyu-school-of-law-announce-report-on-artificial-intelligence-in-federal-agencies
[22] David Freeman Engstrom & Daniel E. Ho. “Algorithmic Accountability in the Administrative State,” Yale Journal on Regulation 37 (2020): 800 (SSA and SEC case studies; limits of ex ante and ex post review; administrative-law adaptation). https://dho.stanford.edu/wp-content/uploads/EngstromHo.pdf
[23] Stanford Institute for Human-Centered AI (AI Index Steering Committee: R. Perrault & Y. Gil, co-chairs; incl. E. Brynjolfsson, J. Clark, J. Manyika, R. Altman). The 2026 AI Index Report, Stanford University, April 2026 (53% population adoption; 88% organizational adoption; $581.7B corporate investment; 2.7% U.S.–China frontier gap; benchmark gains). https://hai.stanford.edu/ai-index/2026-ai-index-report
[24] Stanford HAI. The 2026 AI Index Report — Economy chapter ($172B U.S. consumer surplus; 70% of organizations using generative AI; single-digit agent deployment; U.S. 24th at 28.3% population adoption; entry-level developer employment decline). https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
[25] Forbes (S. Wolfe Pereira). “Stanford’s AI Report Card: Agents Are Ready. Companies Are Not,” April 14, 2026 (jagged frontier; 362 documented AI incidents in 2025; Index governance and steering committee). https://www.forbes.com/sites/stevenwolfepereira/2026/04/14/stanfords-ai-report-card-agents-are-ready-companies-are-not/
[26] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027,” August 26, 2026 ($96.2B revenue, +106% YoY; $89.0B Data Center revenue; Q3 FY27 guidance of ~$108B). https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
[27] NVIDIA Corporation, U.S. SEC Form 8-K Exhibit. Q2 FY2027 press release as filed, August 26, 2026 (Jensen Huang statement on AI’s inflection point and compute as revenue). https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm
[28] CNBC (K. Leswing). “Nvidia earnings takeaways: Huang forecasts 70% fiscal 2028 revenue growth,” August 26, 2026 (memory-scarcity commentary; AWS 2-million-GPU expansion; margin guidance). https://www.cnbc.com/2026/08/26/nvidia-nvda-earnings-report-q2-2027-live-updates.html
[29] Microsoft Corporation. “Microsoft Cloud and AI strength fuels fourth quarter results,” July 29, 2026 (FY2026 revenue $331.8B; Nadella statement on the cost-to-outcome curve; Anthropic investment gain). https://news.microsoft.com/source/2026/07/29/microsoft-cloud-and-ai-strength-fuels-fourth-quarter-results-4/
[30] Microsoft Corporation, Investor Relations. Fiscal Year 2026 Fourth Quarter Earnings Conference Call transcript, July 29, 2026 (Microsoft Cloud $214B+; Azure $100B+, +41%; 31 new datacenters in the quarter, 88 for the year). https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4
[31] CNBC (J. Novet). “Microsoft (MSFT) Q4 earnings report 2026,” July 29, 2026 (30 million+ paid Microsoft 365 Copilot seats; FY26 capital expenditures near $116B; Azure growth detail). https://www.cnbc.com/2026/07/29/microsoft-msft-q4-earnings-report-2026.html
[32] International Monetary Fund (M. Cazzaniga, F. Jaumotte, L. Li, G. Melina, A. J. Panton, C. Pizzinelli, E. J. Rockall & M. M. Tavares). “Gen-AI: Artificial Intelligence and the Future of Work,” IMF Staff Discussion Note SDN/2024/001 (≈40% of global employment exposed to AI; ≈60% in advanced economies; new-industrial-revolution framing). https://www.imf.org/-/media/files/publications/sdn/2024/english/sdnea2024001.pdf
[33] Gulf News (N. Dayanand). “AI tsunami could disrupt 40% of jobs worldwide, IMF chief warns,” World Economic Forum, Davos, January 2026 (Kristalina Georgieva remarks on labour-market transformation and readiness). https://gulfnews.com/business/markets/ai-tsunami-could-disrupt-40-of-jobs-worldwide-imf-chief-warns-1.500417389
[34] U.S. General Services Administration. “GSA Announces CORAS Partnership Through OneGov, Expanding Federal AI Access and Delivering Cost Savings of up to 80%,” July 28, 2026. https://www.gsa.gov/about-gsa/newsroom/news-releases/gsa-announces-coras-partnership-through-onegov-07282026
[35] OpenAI. “Solutions for Government” (ChatGPT Enterprise, Codex, and API availability; FedRAMP 20x Moderate accreditation; Azure OpenAI and AWS GovCloud deployment paths). https://openai.com/solutions/industries/government/
[36] Quartz (C. Tolomia). “OpenAI ends $1-a-year federal deal, offers 50% discount on AI,” September 11, 2026 (legislative and judicial agency eligibility; state, local, and tribal extension; OneGov context). https://qz.com/openai-federal-government-50-percent-discount-usage-pricing-091126
[37] SmarterX / Marketing AI Institute. “Stanford’s 2026 AI Index Just Made the Overhype Argument Hard to Defend,” April 2026 (Erik Brynjolfsson on AI productivity effects appearing in national economic data; Index synthesis). https://smarterx.ai/smarterxblog/stanford-2026-ai-index-report
