SOVEREIGN COMPUTE¶

Infrastructure Sovereignty as the Foundational Prerequisite for Digital Autonomy in the Age of AI¶

Derek Chapman

The Eno Project

Version 1.0, June 2026

theenoproject.com

Foreword¶

The system is extractive. I've known that for a long time. Not in an abstract, ideological way, but in the practical, everyday sense. Economic value flows consistently upward. A smaller number of companies own more of the things we depend on each year. The architecture of modern life is arranged so that more of it runs through fewer chokepoints. Most people feel this even if they'd struggle to describe it. They feel it when they check their bank account, when they scroll through what used to be an open internet, when they try to build something and find the on-ramps controlled by someone else.

At first I just vented and complained, mostly to people who didn't care much or who's eyes glaze over when you start talking global macro-economics and solving the world's structural problems. Complaining is easy and cheap. It doesn't actually change anything and more complaining only makes you more frustrated. I tried to think about mechanisms, not grievances. Starting a blog felt pointless; a YouTube channel, more noise in an already deafening room. I wanted something that could actually change the structure, not just describe why it's broken.

Almost ten years ago I got into crypto for the reasons most people in this space did. Decentralization mattered. Sovereignty mattered. The idea that you could opt out of captured systems, even partially, was worth something real. But there was a contradiction I couldn't get comfortable with. The technology calling itself decentralized mostly ran on centralized servers owned by Amazon, Google, and Microsoft. And actually running a node yourself, participating in the network as a first-class citizen, required a level of technical knowledge that almost no one had. The principle and the practice didn't match.

What changed is AI. When I understood that AI could run the node for you, that the person operating the infrastructure no longer needed to be an engineer, something clicked into place. The barrier that kept ordinary people out wasn't ideological. It was operational. And now that barrier is close enough to gone that we can build around what remains.

That realization leads somewhere specific. In a world where intelligence can act independently, where AI agents execute tasks and make decisions without waiting for human input, the question of who owns the underlying compute stops being theoretical. It's the whole question. The compute is where control actually lives. Owning it, or not owning it, determines whether decentralization is real or just a story we tell ourselves.

This paper is about that: what sovereign compute means, how it works, and why it matters now.

If you understood why Bitcoin mattered, this will feel familiar. The decentralized movement built something real, but it's always been missing a layer. Actual physical infrastructure, owned and operated by the people it's supposed to serve, connected to real communities rather than floating above them. That's what both sides have needed to succeed. This is it.

I'm not asking you to take a leap of faith. I'm asking you to help build something consistent with what you already believe. Read what follows. Then let's get to work.

Derek Chapman, The Eno Project

Abstract¶

Every major economic epoch is defined by the entity that controls its foundational means of production. In the agrarian economy, that layer was arable land. In the industrial economy, it was energy and heavy capital: rail, steel, oil. In the digital economy, the foundational layer is compute, the capacity to process, store, and serve information at scale.

This paper argues that compute concentration isn't a temporary market condition amenable to regulatory correction, but a self-reinforcing phenomenon driven by capital intensity, network effects, and technical lock-in. Three entities control approximately 63% of global cloud infrastructure spending, which exceeded $400 billion in 2025. A single semiconductor company commands 85 to 87% of the AI accelerator market by revenue. The capital expenditure required to compete at the infrastructure layer now exceeds $100 billion annually per firm, and the four largest hyperscalers are projected to spend more than $700 billion collectively in 2026, almost entirely on AI infrastructure.

The emergence of artificial intelligence transforms this concentration from an economic inconvenience into an epistemic dependency. Unlike previous software paradigms, AI systems don't merely process instructions given by users. They mediate understanding, filter information, generate analysis, and increasingly make decisions on behalf of individuals and institutions. Whoever controls the compute layer that trains and serves these models controls the epistemic infrastructure of the digital economy: not what people are allowed to say, but what they are able to think.

Underneath the argument is a simpler claim. Intelligence is becoming a commodity. Given the same model, a mechanic can reason like a physicist and a solo operator can produce like a team. When intelligence is everywhere and roughly equal, it stops being the thing that separates people or firms. What stays scarce is the compute it runs on, and the question of who owns it. Whoever controls that substrate controls what the intelligence is allowed to do. That is why this paper treats compute, not intelligence, as the ground worth fighting for.

The consequences extend beyond epistemology. Enterprise AI adoption has crossed a threshold: 72% of enterprises now report at least one AI deployment in production, and the competitive dynamics of adoption have become coercive. When one competitor deploys AI to reduce its cost structure by 30 to 40%, every other competitor faces a choice built into the situation: match, or lose on margin. The productivity surplus AI generates follows a predictable path: from enterprise operations to shareholder earnings, and from every enterprise's AI deployment to the infrastructure providers who extract rent from all of it. This isn't a prediction about what AI might do. It's a description of what it's doing now.

The direction of AI's broader economic restructuring depends on who controls the compute. In one path, AI is deployed top-down at the enterprise level, automating roles, consolidating headcount, and routing economic activity through platform intermediaries that capture the value of AI-driven efficiency while individuals provide labor at diminishing returns. In the other path, AI amplifies individual capability at the edge, enabling solo operators and small practitioners to deliver enterprise-scale output without institutional overhead or platform extraction. The first outcome concentrates economic power. The second distributes it. The variable that determines which outcome prevails is infrastructure control.

This paper demonstrates that application-layer interventions (privacy tools, encryption, data portability standards, and regulatory frameworks) are insufficient because they operate within infrastructure controlled by entities that can be pressured, regulated, compelled, or that may unilaterally alter their terms of service. Sovereignty at the application layer without sovereignty at the infrastructure layer is a lease, not ownership.

We propose two complementary responses. First, for individuals and organizations: the immediate construction of personal AI stacks, sovereign compute infrastructure capable of running inference, fine-tuning models, and serving applications without dependency on centralized providers. We argue this is the direct analog to historical acts of economic self-determination: homesteading, electrification, and the construction of independent communications infrastructure. Second, for the broader network: the Eno Project, an open-source foundation that serves as the orchestration layer connecting the components of a sovereign digital economy (open-source software, decentralized network infrastructure, tokenomic incentive structures, and AI agents), all of which already exist in some form. The missing piece is the foundational hardware platform that breaks the disposable-hardware cycle and eliminates the cloud dependency that these other components currently can't escape. The Eno Foundation's specific mandate is to fund and design that platform, and to coordinate the existing components into a coherent, self-sustaining system, not through ideology, but through superior economic alignment between infrastructure providers and the communities they serve.

The argument is empirical, not political. The data is public. The architecture is open by design. The reader is invited to verify every claim independently.

1. The Pattern: Control of Foundational Infrastructure Across Economic Epochs¶

1.1 The Thesis¶

Economic power doesn't accrue primarily to those who create value at the application layer. It accrues to those who control the infrastructure upon which all application-layer activity depends. This isn't a novel observation. It's a pattern that has repeated across every major economic transition in recorded history. What changes between epochs is the identity of the foundational layer, not the dynamics of its control.

The pattern operates through a consistent process: a new technology or resource emerges that enables a qualitatively new category of economic activity. Early in its adoption, access is distributed. As the technology matures, returns to scale emerge (network effects, capital barriers, regulatory capture, or technical lock-in) that concentrate control among a decreasing number of entities. Once concentration is achieved, the controlling entities extract rent from all economic activity that depends on their infrastructure, and they acquire the ability to set terms for participation in the economy their infrastructure enables.

Understanding this pattern is essential for correctly diagnosing the present situation. Compute concentration isn't a policy failure, a market inefficiency, or a temporary condition that competition will self-correct. It's the natural and predictable consequence of infrastructure economics operating as they always have.

1.2 The Agrarian Epoch: Land as Foundational Infrastructure¶

In pre-industrial economies, land was the foundational layer of production. All economic activity (agriculture, resource extraction, habitation, commerce) required access to land. The entity that controlled land controlled the economy.

The pattern is visible across civilizations. In feudal Europe, land ownership was synonymous with political power. The manorial system concentrated arable land under aristocratic control, with the majority of the population providing labor as tenants under terms set by landowners. The English Enclosure Acts of the 18th and 19th centuries formalized this concentration, converting roughly 6.8 million acres of common land to private ownership. The result was not merely economic inequality but fundamental dependency: those without land had no means of independent production and were compelled to sell their labor on terms set by those who controlled the foundational resource.[^1]

In colonial contexts, the same pattern replicated at continental scale. The displacement of indigenous populations across the Americas, Africa, and the Pacific was fundamentally an infrastructure seizure: the transfer of foundational productive capacity from distributed local control to concentrated colonial control. The economic consequences persisted for centuries after formal colonialism ended, precisely because the infrastructure transfer was baked in rather than merely political.[^2]

The critical insight is that those who controlled land didn't merely become wealthy. They acquired the power to define the terms on which everyone else participated in the economy. This is the distinction between wealth and systemic power, and it recurs in every epoch.

1.3 The Industrial Epoch: Energy and Transport as Foundational Infrastructure¶

The Industrial Revolution shifted the foundational layer from land to energy and the physical infrastructure required to harness and distribute it: coal mines, oil wells, rail networks, electrical grids, and the heavy capital goods that comprised the means of industrial production.

The United States provides the canonical example. Between 1870 and 1910, railroad companies controlled not only transportation but the economic viability of entire regions. The railroad barons (Vanderbilt, Harriman, Gould, Stanford) didn't merely profit from transport services. They controlled the infrastructure upon which agriculture, mining, manufacturing, and commerce depended. A town not served by a railroad was economically nonviable. A farmer whose grain couldn't reach market had no economy to participate in, regardless of how productive his land.

Standard Oil demonstrated the same dynamic in energy. At its peak in 1904, Standard Oil controlled approximately 91% of U.S. oil refining capacity and 85% of final sales. This wasn't achieved through superior product quality. It was achieved through infrastructure control: refining capacity, pipeline networks, railroad preferential rates, and the vertical integration of supply chains. The 1911 Supreme Court dissolution of Standard Oil into 34 companies recognized, in legal terms, what this paper argues in economic terms: that control of foundational infrastructure constitutes a qualitatively different category of market power than control of application-layer products.[^3]

The electrical grid repeated the pattern once more. Samuel Insull's holding company empire, which at its peak controlled electrical utilities serving roughly 4 million customers across 32 states, demonstrated that whoever controlled generation and distribution infrastructure could set terms for every business and household that depended on electricity.[^4]

In each case, the underlying dynamics were identical: high capital barriers created natural monopolies or oligopolies; control of the infrastructure layer conferred pricing power over all application-layer activity; and political power followed economic control, not the reverse.

1.4 The Transition to Digital: The Pattern Repeats¶

The digital economy followed the same course. Its foundational layers emerged in sequence: telecommunications infrastructure (1950s to 1980s), internet backbone and connectivity (1990s to 2000s), and cloud computing and data center infrastructure (2010s to present). At each stage, early distribution gave way to concentration as returns to scale asserted themselves.

The Internet, initially designed as a distributed and resilient network, concentrated at the physical layer into a small number of backbone providers and at the application layer into a small number of platform companies. Cloud computing, which emerged in the mid-2000s with the explicit promise of democratized access to enterprise-grade infrastructure, has consolidated into a three-firm oligopoly that controls the majority of the market and gains share year over year.

The pattern is the expected outcome of infrastructure economics. What has changed is the nature of the infrastructure, and, as the next sections argue, the consequences of its concentration.

2. The Current Epoch: Compute as the Means of Production¶

2.1 Defining Compute Concentration¶

Compute concentration is the consolidation of computational infrastructure (processing capacity, storage, networking, and the associated software platforms) into a small number of entities that control the conditions of access for the rest of the economy. It's measurable along several dimensions, each of which reveals a market structure that is concentrated, increasing in concentration, and inherently resistant to deconcentration.

2.2 Cloud Infrastructure: The Three-Firm Oligopoly¶

As of Q3 2025, three companies (Amazon Web Services, Microsoft Azure, and Google Cloud) account for 63% of global enterprise spending on cloud infrastructure services. The worldwide market value reached $107 billion in that quarter alone, a 57% increase from $68 billion eight quarters prior. For the IaaS segment specifically, Gartner reports that the three firms captured nearly 71% of global spend in 2024, with AWS alone holding approximately 38%.[^5]

The concentration is increasing, not decreasing. The Big Three's combined share rose from 61% to 63% over an eight-quarter period in which the total market nearly doubled in absolute terms. Smaller providers collectively lost share despite the rising tide; no single competitor outside the Big Three holds more than low-single-digit market share.[^6]

Full-year 2025 cloud infrastructure revenues exceeded $400 billion for the first time. Public cloud end-user spending across all service models reached an estimated $723 billion in 2025.[^7]

2.3 AI Accelerator Hardware: A Near-Monopoly¶

The concentration at the hardware layer is more extreme still. NVIDIA's revenue share of the AI accelerator market peaked at approximately 85 to 87% in 2024 to 2025 across data center GPU segments, in a market that grew from $15 billion in 2022 to over $100 billion annually by 2024. Data from Bloomberg and company-reported segments show NVIDIA's share of AI data center revenue rising from 25% in 2021 to 86% in 2025.[^8]

NVIDIA's data center revenue reached $51.2 billion in a single quarter (Q3 FY2026), with gross margins of approximately 73.4% (GAAP). The company closed its fiscal year 2025 with $130.5 billion in total revenue.[^9]

This isn't merely a market share number. It represents a technical lock-in that compounds over time. NVIDIA's CUDA software suite, built over nearly two decades with more than four million developers and a deep catalog of optimized libraries and applications, creates switching costs that function analogously to the railroad gauge standards of the 19th century. Migrating workloads from CUDA to alternative platforms (AMD's ROCm, Google's TPU/JAX network, or custom ASIC solutions) requires rewriting code, retraining staff, and accepting performance uncertainty. Most organizations won't voluntarily incur these costs while NVIDIA's platform remains dominant.

The competitive landscape confirms the concentration. AMD's AI chip revenue, while growing, represents 5 to 8% of the market; hyperscaler custom ASICs are projected to reach 10 to 15% by 2026 but serve primarily internal workloads. Custom silicon supplements rather than replaces the dominant platform.[^10]

2.4 Capital Expenditure: The Barrier That Enforces Concentration¶

The capital required to compete at the infrastructure layer has reached a scale that functionally eliminates new entry. The four largest hyperscalers (Amazon, Alphabet/Google, Meta, and Microsoft) are projected to spend more than $700 billion collectively on capital expenditure in 2026, with the overwhelming majority directed at AI infrastructure.[^11]

Individual firm commitments are staggering: Amazon committed to approximately $200 billion in capex for 2026, Alphabet to $180 to 190 billion, Meta to $125 to 145 billion, and Microsoft tracking toward roughly $190 billion. These figures have more than doubled in two years. Even at this scale, multiple hyperscalers report that their markets are supply-constrained rather than demand-constrained. Microsoft, for one, has described its cloud capacity as constrained by power availability rather than demand, while carrying a commercial backlog (remaining performance obligations) of roughly $625 billion as of early 2026.[^12]

The capital barrier is self-reinforcing. These expenditures are increasingly funded not from operating cash flow but from debt markets. Hyperscalers issued approximately $121 billion in new debt in 2025, with over $90 billion raised in the final three months of the year alone. Morgan Stanley and JP Morgan project the technology sector may need to issue as much as $1.5 trillion in new debt over the coming years to finance AI infrastructure.[^13]

2.5 The Training Cost Ratchet¶

The economics of frontier AI model development further entrench concentration. Research from Epoch AI reveals that the amortized hardware and energy cost of training frontier models has grown at a rate of approximately 2.4x per year since 2016. GPT-4's amortized training cost was estimated at approximately $40 million; the hardware acquisition cost to build the cluster on which it trained was approximately $800 million. Anthropic CEO Dario Amodei stated in 2024 that training runs approaching one billion dollars were imminent. Epoch AI's projections indicate the most expensive publicly announced training runs will exceed one billion dollars by 2027.[^14]

OpenAI's 2024 compute expenditure illustrates the scale: of an estimated $5 billion compute budget, the large majority was allocated to experimental training runs and research rather than the final training of released models (Epoch AI estimates that under $1 billion went to the models that actually shipped). Frontier AI development requires not merely the capital to execute a single training run, but the sustained capital to fund hundreds of experimental runs for every model that reaches production.[^15]

The Stargate Project, a joint venture led by OpenAI, SoftBank, Oracle, and MGX, has announced plans to invest $500 billion in AI infrastructure over the coming years, with $100 billion deployed in 2025 alone. Whether these specific figures materialize, they signal the scale at which serious participants in frontier AI believe they must operate.[^16]

3. Why AI Changes the Stakes: From Inconvenience to Existential Dependence¶

Begin with the change that makes everything else follow. Intelligence is becoming ubiquitous, and it is becoming agentic: it no longer only answers questions, it takes actions. As it spreads, it stops being scarce. Given the same model, a mechanic reasons at the level of a physicist; a solo operator produces at the level of a team. The capability that used to separate people is turning into a commodity anyone can run.

Two consequences follow, and they define the rest of this paper. First, when intelligence is everywhere, the only thing that remains scarce is the compute it runs on and the question of who owns it. Whoever controls that substrate controls what the intelligence is permitted to see, say, and do. Compute, not intelligence, becomes the locus of power. Second, if intelligence is everywhere, the only thing that still distinguishes one outcome from another is what people choose to do with it. Human intent becomes the deciding variable, and owning the compute beneath the intelligence is what keeps that choice in human hands.

This is the whole argument in one line: in a world of ubiquitous intelligence, the compute it runs on is the last thing that matters, and the first thing we have to own. The sections that follow trace what happens when that compute is concentrated, and what becomes possible when it is not.

3.1 The Qualitative Shift¶

Previous forms of compute concentration were economically significant but epistemically neutral. A business running its ERP system on AWS was dependent on Amazon for uptime and pricing, but the software itself executed deterministic logic defined by its developers. The infrastructure provider hosted the computation; it didn't shape the computation's conclusions.

Artificial intelligence inverts this relationship. AI systems (particularly large language models, recommendation engines, and autonomous agents) don't execute deterministic logic. They generate outputs shaped by their training data, their alignment procedures, their inference-time constraints, and the priorities of the entities that control each of these parameters. The infrastructure provider is no longer a passive host. It's an active mediator of the output's content, framing, and boundaries.

This qualitative shift moves compute concentration from the category of "economic inconvenience" to "epistemic dependency." It's about what the computers tell people to think, not what people are allowed to do with their computers.

3.2 Epistemic Infrastructure: A Precise Definition¶

By "epistemic infrastructure" we mean the systems that mediate how individuals and institutions acquire, evaluate, and act on information. Historically, epistemic infrastructure included libraries, universities, newspapers, broadcast media, and search engines. Each shaped not just the availability of information but the frameworks through which information was interpreted.

AI systems are now assuming this role at a scale and depth that surpasses all prior epistemic infrastructure. When an individual asks an AI assistant to summarize a legal dispute, analyze a medical condition, evaluate a business opportunity, or explain a political situation, the AI's response isn't a neutral relay of facts. It's a synthesized interpretation shaped by: the model's training data (which data was included and excluded); alignment tuning (what the model was trained to emphasize or avoid); inference constraints (how much computation was allocated to the response); and the provider's content policies (what topics the model will engage with and how).

Each of these parameters is controlled by the entity that operates the compute infrastructure on which the model runs. The user has no visibility into, and no control over, any of them.

3.3 The Dependency Cascade¶

The epistemic dependency compounds across domains. Consider the following cascade, each step of which is either already occurring or is the stated roadmap of major AI providers:

Information access. AI systems increasingly serve as the primary interface through which individuals interact with information. Search engines are being augmented or replaced by AI-generated summaries. The user doesn't see the sources; they see the model's synthesis.

Analysis and decision support. Businesses deploy AI systems for financial analysis, legal research, medical diagnosis support, and strategic planning. The AI's framing of options directly shapes the decisions that follow.

Content generation. AI systems produce text, code, images, and video that constitute an increasing share of the content individuals encounter. The distinction between "AI-assisted" and "AI-generated" content is already blurring.

Autonomous action. AI agents that execute tasks on behalf of users (scheduling, purchasing, communicating, negotiating) introduce a layer where the AI's decisions replace the user's decisions entirely for routine operations.

At each level, the user becomes more dependent on the AI system's integrity, and the entity controlling the compute infrastructure gains more influence over the user's economic and cognitive activity.

3.4 The Alignment Problem Is a Sovereignty Problem¶

The technical AI safety community has extensively documented the "alignment problem": the challenge of ensuring AI systems behave in accordance with human values and intentions. This paper argues that the alignment problem has a dimension that is underexamined: even a perfectly aligned AI system (one that faithfully executes its instructions) is aligned to the values and priorities of whoever controls it, not necessarily to the values and priorities of the user.

An AI model operated by a publicly traded company is subject to the company's commercial incentives, regulatory obligations, terms of service, content policies, and the political pressures to which the company is exposed. These constraints aren't hypothetical. They're documented, visible in model behavior, and change over time without the user's consent or necessarily even the user's awareness.

This isn't a complaint about any particular company's policies. It's an observation: a system that mediates your understanding of reality, and whose behavior is controlled by an entity whose interests aren't identical to yours, is a dependency that no amount of application-layer tooling can resolve. The only remedy is to control the compute layer on which the model runs.

3.5 Who Trains the Models Mediates Reality¶

The training process of large language models is itself an act of epistemic construction. The selection of training data (what is included, what is excluded, what is weighted) determines the model's representation of knowledge. Reinforcement learning from human feedback (RLHF) and constitutional AI methods further shape the model's outputs according to criteria set by its developers. The result isn't a neutral knowledge base but a curated epistemic framework that reflects the priorities, assumptions, and constraints of its creators.

This isn't inherently malicious. All epistemic infrastructure involves curation. The question is not whether curation occurs but who performs it and whether the end user has the ability to choose among competing curations, modify the curation to suit their own needs, or operate without curation entirely. In the current market structure, the answer to all three questions is effectively "no" for any user who doesn't control their own compute infrastructure.

A user running a locally hosted, open-weight model on their own hardware can inspect the model's weights, modify its system prompt, fine-tune it on their own data, and operate it without any content restrictions imposed by a third party. A user accessing the same model through a cloud API can do none of these things with certainty, because the provider can modify the model, restrict access, log queries, change terms, or shut down the service at any time.

The difference between these two users isn't a matter of technical preference. It's the difference between epistemic sovereignty and epistemic dependency: the difference between owning your means of understanding and renting them.

3.6 The Enterprise Arms Race Is Already in Motion¶

From epistemic dependency, the consequences extend directly into economics. The economic restructuring AI is producing is current, measurable, and accelerating.

Enterprise AI adoption has crossed a threshold of competitive compulsion. As of 2026, 72% of enterprises report at least one AI deployment in production (a figure that was near 50% only two years prior). Gartner projects that 40% of enterprise applications will be integrated with task-specific AI agents by end of 2026, up from less than 5% in 2025.[^17]

The distinction that matters isn't the adoption rate itself but the nature of what's driving it. A technology that 72% of enterprises have deployed isn't an emerging trend; it's becoming a competitive baseline. When one firm in a market segment deploys AI to reduce its cost by 30 to 40%, every competing firm faces a direct choice: match, or lose on margin. No executive has discretion to opt out on principle. The competitive pressure makes adoption compulsory, independent of any individual firm's preferences about workforce composition.

The path of adoption follows a recognizable progression:

Phase 1 (current for most enterprises): AI augments existing workers, improving per-worker output. Headcount holds or grows slightly as firms explore capabilities.

Phase 2 (current for early adopters; 12 to 24 months for mainstream): AI replaces discrete functions (back-office operations first: data entry, report generation, customer service triage), then mid-skill knowledge work (legal research, financial modeling, code generation, content production).

Phase 3 (emerging in 2025 to 26): Full reorganization. Entire departments reconstitute around AI-first workflows with dramatically reduced headcount. The organizational chart changes shape, not just size.

The speed of this progression isn't governed by AI capability; the capability for most Phase 2 and many Phase 3 use cases already exists. It's governed by enterprise adoption cycles: procurement, integration, change management, and institutional inertia. But the direction is unambiguous, and the competitive pressure is relentless.

Goldman Sachs Research estimates that 300 million jobs globally are exposed to automation by AI (a figure the firm has reaffirmed as recently as March 2026). At the U.S. level, Goldman estimates 6 to 7% of the workforce faces direct displacement risk if AI capability spreads across the economy at the pace experts project.[^18]

3.7 The Productivity Surplus Capture Process¶

The central economic question isn't whether AI creates value (it does). The question is who captures it.

AI generates a productivity surplus: the delta between pre-AI cost and post-AI cost for a given unit of output. In the current market topology, the capture pattern operates as follows:

Enterprise deploys AI, reducing labor input per unit of output, improving cost profile.
Cost improvement flows to operating margin; operating margin flows to earnings; earnings flow to shareholders.
Displaced labor enters a market where the same forces play out across competing firms, reducing demand for the displaced skill set and creating wage pressure on remaining positions.
AI infrastructure providers (the hyperscalers, NVIDIA, frontier model companies) capture revenue from every enterprise's AI deployment, compounding their margins and market capitalization.

The result is a two-tier extraction: the enterprise captures surplus from its own workforce, and the infrastructure provider captures surplus from every enterprise. The labor force is on the losing end of both tiers.

This pattern is the standard path by which returns to capital outpace returns to labor when a new means of production is introduced. Research by Acemoglu and Restrepo (2024) finds that automation systematically targets high-rent tasks, dissipating worker rents and amplifying wage losses. Their analysis estimates that inefficient rent dissipation has offset 60 to 90% of the productivity gains from automation in the U.S. since 1980 (meaning that the macroeconomic productivity dividend from automation has not, historically, translated proportionally into wage growth for the workers whose tasks were automated).[^19]

What distinguishes AI from previous automation waves isn't the underlying process; it's the breadth of application. Previous automation technologies affected specific domains of physical or routine cognitive tasks. AI automation addresses cognitive tasks across virtually every knowledge-work domain simultaneously. The displacement isn't sector-specific; it's economy-wide.

3.8 Why This Displacement Differs Historically¶

Every previous major automation wave triggered a counterargument ("automation creates more jobs than it destroys"), and that counterargument has historically been correct. Agricultural automation displaced farm labor, which was absorbed by manufacturing. Industrial automation displaced assembly-line work, which was absorbed by the service economy. Early digital automation displaced clerical and administrative work, which was absorbed by knowledge work and the technology sector.

In each case, the new technology automated the previous era's foundational task while creating demand for a new category of human work that the technology couldn't perform. The question for AI is whether this pattern holds.

The honest answer is that it isn't yet clear, and specific reasons justify concern.

AI automates the cognitive layer itself. Previous automation technologies affected physical tasks or routine information processing. AI automates analysis, synthesis, judgment, communication, and creative production (the functions that defined the knowledge-work employment layer that absorbed previous waves of displacement). The technology is automating the destination, not just the origin.

The capability frontier moves faster than retraining cycles. Previous transitions unfolded over decades, allowing workforce retraining and generational adjustment. AI capability advances on a timeline measured in months to years. The gap between displacement speed and adaptation speed is wider than in any previous transition.

New AI-era roles carry infrastructure dependency. New roles created by AI (prompt engineering, AI operations, model fine-tuning, AI-augmented analysis) require access to AI infrastructure. If that infrastructure is concentrated, the new employment layer depends on the same entities whose products caused the displacement. This is circular dependency, not absorption.

This doesn't establish that mass permanent unemployment is inevitable. It does establish that the historical pattern of sector-to-sector absorption can't be assumed for this transition, and that the response (ensuring individuals have their own productive AI infrastructure) is the path that creates the replacement employment layer: sovereign operators producing enterprise-scale output on their own terms.[^20]

3.9 The Corporate Incentive Architecture: Design, Not Malice¶

The consolidation path described above isn't a worst-case scenario that might be avoided through corporate goodwill or better industrial policy. It's the default outcome of existing institutional incentives encountering a technology that dramatically reduces the cost of cognitive labor.

The publicly traded corporation is a legal entity with a specific design: it exists to maximize returns to shareholders. This isn't an accusation; it's the entity's legal architecture, established across virtually every major jurisdiction through fiduciary duty doctrine. Boards of directors who knowingly forgo profit maximization face legal exposure to shareholders. This is how the system works, not a discretionary choice.[^21]

When AI enables a corporation to produce the same output with fewer employees, the corporation will deploy AI and reduce headcount. Not because its executives are indifferent to displaced workers (many aren't), but because the entity's incentive architecture, legal obligations, and competitive environment all point in the same direction. A CEO who refuses to automate while competitors automate will deliver inferior margins, face board pressure, and eventually be replaced by someone who will.

This framing matters precisely because it closes off the most common objection to the paper's economic argument: that better corporate behavior or more socially minded leadership could alter the outcome. The entity is designed to optimize for shareholder return. AI gives it a powerful new lever for doing so. It will pull that lever. The only question is how fast.

The implication is direct: avoiding the consolidation path requires not corporate goodwill but intervention at the infrastructure level. Specifically, ensuring that the population has productive infrastructure that isn't dependent on the entities doing the displacing.

3.10 The Full Compression Stack: How AI Consolidates Every Layer Simultaneously¶

The productivity surplus capture pattern described in Section 3.7 operates at the labor layer. But AI-driven consolidation doesn't stop there. It compresses the economy at every layer simultaneously (infrastructure, software markets, labor, and household economics) in a cascade that's already underway and that the existing governance structure of the digital economy has no way to arrest.

The first compression is at the infrastructure layer, documented in Section 2: compute concentrates among three cloud providers and one chip company, with capital barriers that eliminate new entry. The second compression operates at the application layer, where frontier AI is eliminating the market for the software products built on top of that infrastructure.

The SaaS market that emerged over the past two decades was built on a specific premise: specialized software vendors could build durable businesses by owning particular workflow categories (customer support, legal research, financial modeling, content production, knowledge management). That premise is dissolving. Frontier AI models now perform, at inference cost, the core functions that hundreds of mid-market SaaS vendors spent years and billions of dollars building. The vendors who can't retrain their entire product stack around frontier models (which is most of them, because the capital required to do so rivals that of the hyperscalers themselves) will be absorbed or displaced. The application layer is consolidating for the same reason the infrastructure layer consolidated: returns to scale, this time in model capability rather than data center economics.

The scale of this compression is visible in real time. Entire product categories that commanded premium recurring revenue two years ago (AI writing tools, legal research platforms, customer service software, mid-tier code generation tools) now face direct competition from the base models embedded in the hyperscalers' own platforms. The winners in this consolidation are the same entities that win at the infrastructure layer: the companies with the capital to deploy frontier models at scale. The losers are the thousands of software vendors whose differentiated products now compete with a feature included for free in an enterprise AI subscription.[^22]

The third compression operates at the labor layer, already analyzed in Sections 3.6 to 3.8. Its scope deserves calibration against current projections.

The AI 2027 scenario analysis, developed by former OpenAI researchers using trend extrapolation and expert forecasting, projects AI systems with superhuman coding ability by 2027 and the internal deployment of hundreds of thousands of AI research assistants within that same window. These aren't fringe predictions; they're the stated planning assumptions of the organizations building the systems. The capability for large-scale cognitive displacement isn't a decade away. It's the current development path of systems in production now.[^23]

The fourth compression doesn't stop at cognitive labor. AI-guided robotics represent the extension of the same economic logic to physical work. The barrier to physical automation has historically been hardware cost, reliability, and the flexibility required to operate in unstructured environments. Each of these barriers is collapsing under AI-driven progress in perception, dexterity, and real-time decision-making. Boston Dynamics, Figure, Apptronik, and a dozen well-capitalized entrants are deploying general-purpose humanoid robots in warehouse and manufacturing environments now. Tesla's Optimus program, backed by the same capital structure and vertical integration that defines Tesla's competitive posture in every market it enters, targets manufacturing deployment at scale.

The physical automation wave isn't a separate phenomenon from the cognitive automation wave. It's the same phenomenon, extended by AI's ability to generalize across task types that previously required human dexterity and judgment. The corporate incentive structure analyzed in Section 3.9 doesn't change when the task moves from a keyboard to a warehouse floor. When the capex to deploy a robotic workforce becomes favorable relative to the ongoing cost of human labor (and the margin improvement from that substitution accrues permanently to shareholders), the corporation deploys the robots. This is a financing decision, not a philosophical one. The entities with the capital to make that financing decision at scale are the same entities accumulating capital from cognitive automation now.

Nothing in the governance structure of the current digital economy suggests a way to prevent this outcome. The regulatory frameworks that exist address data privacy, algorithmic transparency, and content moderation. None address the built-in incentive that drives a corporation with unlimited access to debt capital, facing competitors doing the same, to replace its workforce with machines whenever the economics permit. The entities best positioned to fund physical automation at scale are precisely those accumulating the greatest returns from cognitive automation today.

The fifth compression is at the household economics layer, and it's the one that closes the trap. Displacement doesn't happen in a vacuum; it lands on individuals whose cost of participation has been converted into permanent monthly subscriptions, including the AI tools they now need to stay competitive. The result is a contradiction examined in detail in Section 4.2: AI is most disruptive to the income streams of the people most exposed to subscription cost inflation, and centralized AI pricing extracts rent from the displaced at the very moment of displacement.

Sovereign compute infrastructure breaks this trap at the fundamental level. A person who owns their AI stack (their hardware, their models, their data) has converted a recurring subscription cost into a capital asset. The marginal cost of their AI capability approaches zero. The trajectory these forecasts describe, if accurate, is an argument for urgency in building sovereign AI infrastructure. The individual who owns their productive AI infrastructure before that displacement reaches them is in a fundamentally different position from the individual who rents it from the same entities doing the displacing.

3.11 From Epistemic Dependency to Economic Dependency: The Diverging Futures¶

The dependency cascade described in Section 3.3 doesn't terminate at epistemology. The same infrastructure control that creates epistemic dependency creates economic dependency. AI concentration doesn't merely control what individuals are able to think; it controls the terms on which they're able to work.

Two paths are diverging, and the difference between them is a direct consequence of infrastructure topology:

Path One: AI as a Tool of Consolidation. AI is deployed top-down at the enterprise level. Organizations use AI to automate roles, consolidate headcount, and reduce labor costs. Economic activity is routed through platform intermediaries that control the matching, pricing, and terms of engagement between service providers and consumers. The individual becomes a dependent of the platform (visible only through the platform's algorithm, competing on the platform's terms, subject to the platform's extraction). This is the gig-economy model generalized: the platform captures the value of AI-driven efficiency while the individual provides labor at diminishing returns. In this world, AI is a tool of consolidation. Jobs are automated at the institutional level. The surplus accrues to platform operators and their shareholders.

Path Two: AI as a Tool of Distribution. AI amplifies individual capability at the edge. An individual operating their own AI stack (their own models, their own agents, their own data, their own workflows) acquires capabilities that previously required organizational scale. A solo consultant with a well-built AI stack can perform analysis, generate deliverables, manage client relationships, and coordinate projects at a level that previously required a team. A bookkeeper with sovereign AI serves clients directly rather than through an intermediary platform that dictates pricing and takes a percentage. A developer, a designer, a strategist, a tradesperson with intelligent scheduling and client management (each operates as a full-service entity without the overhead of institutional structure and without the extraction of a platform intermediary).

This isn't a return to a nostalgic small-business past. It's something fundamentally new: individual-scale operators with enterprise-scale capability, transacting in a peer-to-peer economy where they choose who they do business with rather than being matched by an algorithm on someone else's infrastructure.

The variable that determines which path prevails isn't policy, regulation, or corporate goodwill. It's infrastructure control. Centralized compute produces centralized economic structures. Distributed compute enables distributed economic structures. The topology of the infrastructure determines the topology of the economy.

4. Why It Matters: The Human Ledger¶

The preceding sections establish an argument grounded in infrastructure, incentives, and outcomes: compute is concentrated, AI has made that concentration epistemically and economically consequential, and the enterprise arms race is already in motion. That argument is, by design, cold. Analysis at the level of systems operates in ratios, capital flows, and outcomes. What it doesn't convey, without additional work, is what this looks like from the inside (for the specific population of people that this is heading toward).

This section provides that account. What follows is analytic, not sentimental, but it's analytic about people, not systems.

4.1 The Offshoring Parallel: This Has a Precedent¶

The United States lived through one major episode of systematic labor-cost arbitrage before this one. From roughly 1980 through the 2010s, manufacturing employment fell by approximately 7 million jobs. The dynamic was wage arbitrage: American manufacturing workers earning roughly $25 to $40 per hour could be replaced by workers in China, Mexico, and Southeast Asia earning a fraction of that cost. The infrastructure required to execute this arbitrage (container shipping, telecommunications, global supply chains, free-trade agreements) was assembled over decades and framed as an unambiguous policy success by the mainstream economic consensus.

The productivity and consumer surplus from manufacturing offshoring were real. Goods became cheaper. Corporate margins improved. GDP grew. The consensus held that the gains were net positive and broadly shared, and that displaced workers would retrain for the knowledge economy that was absorbing the productivity dividend.

What the consensus obscured was the distribution of the adjustment cost. The productivity gains accrued to capital and to consumers of cheap goods. The adjustment costs fell with geographic precision on industrial communities (factory towns in Ohio, Michigan, Pennsylvania, and the upper Midwest) whose entire economic structure depended on the industries that departed. The retraining narrative was empirically hollow. Coal miners didn't become programmers. Autoworkers didn't retrain as software developers. Research by Autor, Dorn, and Hanson established that regions most exposed to import competition from China experienced persistent, multi-decade declines in employment, wages, and population (effects that conventional trade theory predicted would be temporary but that proved durable). Anne Case and Angus Deaton documented the downstream consequence: elevated mortality rates in deindustrialized communities, driven by overdoses, alcohol-related illness, and suicide (what they termed "deaths of despair").[^24]

The lesson here isn't that comparative advantage is wrong. It's that labor-cost arbitrage concentrates its costs precisely on the people least able to absorb them, and that the institutional supports required to manage that distribution (robust social insurance, genuine transition support, wage floors) were inadequate, politically inconvenient to the interests that benefited from the arbitrage, and never corrected. The communities left behind by offshoring have not recovered. Not in one decade, not in two. They're still losing.

AI displacement is wage arbitrage applied to cognitive work. The economic logic is identical: a knowledge worker earning $80,000 to $150,000 per year can be replaced by AI inference costing a fraction of that in marginal compute. The infrastructure required to execute this arbitrage isn't being assembled; it's already built, already deployed, and already processing the work.

The parallel extends beyond the underlying dynamic to distribution and political invisibility. The knowledge worker whose role is being automated is, in the popular imagination, the winner of the previous transition. They did what they were told: went to college, developed marketable analytical skills, entered the knowledge economy that was supposed to be safe. They weren't warned that the knowledge economy would become the next target of labor-cost arbitrage (this time executed not by container ships and free-trade agreements, but by a technology with no borders, no import tariff, and no social or political cost to the corporation deploying it).

There's one critical dimension in which the analogy understates the current situation. Manufacturing offshoring unfolded over thirty years, allowing partial generational adjustment. AI is automating cognitive work on a timeline measured in months. The one relief valve (generational adjustment) that provided even partial relief in the prior transition is unavailable for this one.[^25]

4.2 The Subscription Economy Trap: Fixed Costs in a Gig World¶

The manufacturing displacement of the 1980s to 2000s occurred against an economic backdrop where the primary costs of living were dominated by large, irregular expenditures (housing, food, vehicles) with comprehensible annual rhythms. A displaced factory worker could, in theory, restructure their cost of living. The obligations were large but reducible at the margin.

The displacement of cognitive labor by AI is occurring against a fundamentally different backdrop: one in which the cost of participating in modern professional life has been systematically converted from periodic large purchases into permanent monthly obligations.

This conversion wasn't accidental. The shift from software-as-purchase to software-as-subscription was a deliberate capital structure optimization executed category by category over the past fifteen years. Adobe's Creative Cloud conversion in 2013 is the canonical example: a professional design suite that cost $600 to $2,600 as a one-time purchase became $55 to $85 per month, indefinitely, with no option to own. Microsoft followed with Office 365. Enterprise software vendors replicated the model universally. The streaming era converted media consumption in the same direction. Cloud storage, project management, fitness services, news access, food delivery, and increasingly AI tools (each converted to recurring monthly charges, each individually framed as trivially affordable).

Research by C+R Research finds that the average American actually spends $219 per month on subscription services (more than 2.5 times their own estimate of $86). West Monroe Partners places the figure higher still at $273 per month. Across all subscription categories, the average American household carries between $200 and $275 per month in recurring digital obligations they can't meaningfully reduce without losing access to the tools of professional and social participation. Seventy-four percent of consumers admit these charges are easy to forget, structured to be invisible: auto-renewed, fragmented across multiple billing sources, and individually small enough that the aggregate doesn't register.[^26]

This subscription cost structure is priced for income stability. It was designed for employees with predictable paychecks. It wasn't designed for the labor market that has actually emerged over the past two decades: gig workers, freelancers, independent contractors, and platform-mediated service providers whose income is variable, seasonal, and client-dependent.

The gig economy and the subscription economy are fundamentally incompatible. The gig economy produces variable income. The subscription economy produces fixed monthly obligations. When the income stops, the obligations don't. When the freelance contract ends, the Adobe subscription doesn't. When the platform algorithm deprioritizes a gig worker's profile, their cloud storage bill doesn't.

The knowledge worker displaced by AI automation doesn't exit this system. They enter a more precarious version of it. And they face a trap that didn't exist in prior displacement waves: competing for whatever work remains after automation requires more AI capability, not less, at the exact moment income has been disrupted. ChatGPT Plus, Claude Pro, GitHub Copilot, Perplexity, and the expanding suite of AI productivity tools each carry monthly costs that now constitute the baseline of professional competitiveness in knowledge work. That monthly tax is levied by the same infrastructure providers whose models created the displacement in the first place.

The trap closes as follows: AI automates knowledge work, displaced workers enter a gig economy with variable income, they retain fixed monthly subscription obligations, competing for remaining work requires more AI access (not less), AI access is itself subscription-based, the displaced pay the entities doing the displacing for the privilege of competing against their products.

This isn't a prediction. It's a description of the economic position a knowledge worker occupies the day after their role is automated, given the market structures that already exist. It's also a dynamic that mainstream media, professional associations, policymakers, and those who make decisions in government have chosen not to name. The subscription economy's incompatibility with the gig economy isn't a niche concern. It's the central financial reality of the population most immediately exposed to AI displacement.

Sovereign compute infrastructure breaks this trap at the foundational level. The individual who owns their AI stack (their hardware, their models, their inference capacity) has converted a $200 to $275 per month recurring cost into a capital asset with near-zero marginal cost at the point of use. That difference, compounded over years of economic transition, isn't marginal. It's the difference between renting the tools of economic survival from the entities causing the displacement, and owning those tools outright.

4.3 Deployment Without Guardrails: The Governance Vacuum¶

The manufacturing displacement of prior decades, whatever its institutional failures, occurred within a pre-existing framework of labor protections built over the preceding century: unemployment insurance, OSHA workplace standards, collective bargaining rights, trade adjustment assistance programs. These supports were inadequate to the scale of displacement they faced. But their existence represented an institutional acknowledgment that economic transitions impose costs on specific people, and that those costs are a legitimate concern of governance.

No equivalent framework exists for AI displacement. The deployment of AI at enterprise scale (replacing knowledge-work functions, eliminating headcount, restructuring the terms of labor market participation) is proceeding without a regulatory framework that treats workers as stakeholders in how that transition unfolds. There's no AI equivalent of the NLRA, OSHA, or the EPA. There's no unemployment insurance designed for the pace and breadth of cognitive automation. There's no trade adjustment assistance for the displaced knowledge worker. The same political cycle that produced four decades of inadequate response to manufacturing displacement is now facing a transition moving an order of magnitude faster.

The organizations building frontier AI have not, in significant instances, maintained internal safety governance adequate to the capability of the systems being deployed. The people with the most direct visibility into AI's risks have been choosing, in public and in numbers, to leave. In February 2026, Mrinank Sharma (head of Anthropic's Safeguards Research team) resigned with a public letter stating that "the world is in peril," citing persistent internal pressure "to set aside what matters most," and writing that "throughout my time here, I've repeatedly seen how hard it is to truly let our values govern our actions." The same week, OpenAI researcher Zoë Hitzig resigned and published an op-ed in The New York Times warning that the technology carries "a potential for manipulating users in ways we don't have the tools to understand, let alone prevent." In the same period, OpenAI dissolved its mission alignment team entirely.[^27]

These aren't the first such departures. In May 2024, Jan Leike, head of OpenAI's Superalignment team, resigned alongside Ilya Sutskever and subsequently stated that at OpenAI "safety culture and processes have taken a backseat to shiny products," and that his team had been denied promised compute resources. OpenAI then disbanded the Superalignment team. The people who built these systems, assessing the institutional priorities governing their deployment, have voted with their careers.[^28]

The military application of AI provides the clearest demonstration of what governance by corporate preference, rather than law, looks like in practice. In early 2026, the U.S. Department of Defense demanded that Anthropic remove two restrictions from its Pentagon AI contract: a prohibition on using its models for the mass domestic surveillance of Americans, and a prohibition on fully autonomous weapons systems without human oversight. These had been explicit conditions since the original $200 million contract was signed in 2025. Anthropic refused. On February 27, Secretary of Defense Pete Hegseth designated Anthropic a "supply chain risk" (a designation historically reserved for companies doing business with foreign adversaries, never previously applied to a domestic firm), and President Trump directed all federal agencies to cease using Anthropic's technology.

Within hours, OpenAI's Sam Altman announced that OpenAI had concluded a Pentagon contract to fill the gap. Altman stated that OpenAI and the Pentagon share the same "red lines": human responsibility for the use of force, including autonomous weapons, and opposition to domestic mass surveillance. Former Pentagon officials and national security experts publicly assessed that assurance as unverifiable. The Electronic Frontier Foundation observed the implication directly: "the state of your privacy is being decided by contract negotiations between giant tech companies and the U.S. government."[^29]

The implication here isn't about Anthropic or OpenAI specifically. It's about governance architecture. The only effective constraint on the deployment of AI for domestic mass surveillance and autonomous kill decisions (applications with direct implications for individual liberty and human life) was the commercial preference of one corporation. That constraint evaporated within a day when a competitor calculated differently. No law prevented it. No regulatory body reviewed it. No democratic deliberation authorized it.

The governance vacuum that permitted this at the level of military AI is the same vacuum governing AI's economic consequences. AI-driven displacement of the cognitive labor force is proceeding under the same non-framework: corporate decisions, competitive pressure, timelines shorter than any regulatory process, and no mandatory accounting for the people absorbing the cost.

4.4 The Blindspot: Who Is Not in the Room¶

The dominant mainstream narrative on AI and labor has a consistent structure: AI will create jobs as well as eliminate them; historical automation was net positive; retraining will absorb the displaced; the gains will be broadly shared. This narrative appears in government documents, corporate communications, and mainstream media with a consistency that reflects aligned institutional incentives rather than independent verification. The organizations producing it (media companies dependent on technology sector advertising, policy institutions funded by technology philanthropy, elected officials receiving technology sector campaign contributions) share incentives to produce optimism and none bear the cost of being wrong.

The most informative data on AI's actual labor market consequences is produced not by these institutions, but by the architects of the transition in settings where they speak to their peers. At the BlackRock Infrastructure Summit in March 2026, Sam Altman told an audience of infrastructure investors that AI is fundamentally shifting the balance between labor and capital, and admitted directly that nobody knows what the policy response should be. The architects of the most consequential labor displacement in a generation are deploying it without a known policy response (and making that admission to the investors financing the infrastructure, at a venue not designed to be part of a national conversation about work).[^30]

The policymakers nominally responsible for the regulatory response are, by design, the last to understand the problem. Legislative processes operate in years. AI capability advances in months. Congressional technology committees were designed for an era of stable technical change and are staffed by generalists drawing on briefings prepared by industry lobbyists whose interests lie in minimizing perceived urgency. The federal government can't pass comprehensive AI labor legislation in the time it takes a frontier model to render an entire software product category obsolete.

Labor unions, which historically represented the interests of workers facing technological displacement, have not, with limited exceptions, developed the technical fluency required to engage substantively in AI policy debates. The professional associations representing knowledge workers are oriented toward licensing and credentialing, not labor market advocacy. The workers themselves (dispersed, often self-employed or contract workers, lacking institutional affiliation) have no organized voice in the process.

The communities that will absorb the first wave of AI displacement are not represented in the rooms where decisions about AI are being made. The executives whose corporations are deploying AI at scale are in those rooms. The investors financing that deployment are in those rooms. The workers whose roles are being automated are not.

This is precisely the condition that produced forty years of inadequate response to manufacturing displacement. The people who bore the cost of that transition weren't in the room when the trade agreements enabling it were negotiated. The communities still paying that cost (in elevated mortality rates, hollowed-out tax bases, and persistent unemployment) were told at the time that adjustment would be temporary, that gains would be shared, and that the market would sort it out.

The AI transition is moving faster. The subscription economy has made the affected population more financially brittle. The governance vacuum is total. And the same institutions that failed the last time have not been reformed to perform better this time. The historical precedent offers a precise calibration of what to expect when this mismatch isn't corrected: the communities that lost the last round of wage arbitrage are still losing, forty years later. AI doesn't allow forty years for a delayed response to arrive.

5. The Insufficiency of Application-Layer Solutions¶

5.1 The Layer Problem¶

The technology industry's response to concerns about privacy, censorship, and digital autonomy has focused almost entirely on application-layer solutions: end-to-end encryption, privacy-preserving browsers, VPN services, federated social protocols, data portability standards, and regulatory frameworks like the GDPR, CCPA, and the EU AI Act. These tools and regulations address real problems and provide meaningful protections within their scope. This paper doesn't argue that they are valueless. It argues that they are insufficient by design, because they operate within an infrastructure layer they do not control.

5.2 Encryption on Rented Infrastructure¶

End-to-end encryption protects the contents of a message in transit. It does not protect against the infrastructure provider's ability to observe metadata (who communicated with whom, when, from where, and how often), throttle or block traffic, deny service, comply with court orders to provide access, or modify the application itself to weaken encryption in future updates.

Signal is an exemplary encrypted messaging application. But Signal runs on AWS. If Amazon Web Services decides, whether for commercial, legal, or political reasons, to terminate Signal's hosting, Signal must migrate to another provider or cease operating. The encryption is intact; the availability is not.

This isn't a hypothetical concern. In January 2021, Amazon Web Services terminated hosting for the social media platform Parler, effectively shutting it down within hours. One can debate the merits of that particular decision, but the precedent is unambiguous: the infrastructure provider has the power to unilaterally terminate any application running on its infrastructure, regardless of that application's encryption, privacy features, or user base.[^31]

5.3 Data Portability Without Infrastructure Portability¶

Data portability regulations (the right to export your data from one service to another) address an important component of user autonomy. But data without the compute infrastructure to process it is inert. A user who exports their data from Google but has no alternative infrastructure on which to run email, storage, calendar, and AI services has accomplished a legal exercise with limited practical value.

The analogy to the agrarian epoch is instructive: a serf granted the right to take their harvested grain when they left the manor still needed land on which to grow next year's crop. Data portability is the digital equivalent of grain portability: necessary but insufficient without access to the means of production.

5.4 Regulatory Frameworks and Their Limitations¶

Regulatory approaches (antitrust enforcement, content moderation requirements, algorithmic transparency mandates) operate within the framework of nation-state governance. They face the same fundamental constraint: the regulated entities control infrastructure that the regulators do not. Enforcement depends on the regulated entities' compliance, which depends on the jurisdiction's enforcement capacity, which varies dramatically across countries and over time.

The EU AI Act, the most comprehensive AI regulatory framework to date, imposes requirements on AI system providers regarding transparency, safety testing, and prohibited use cases. These requirements apply to companies operating within the EU or serving EU citizens. They do not apply to an individual running an open-weight model on their own hardware. This distinction isn't a loophole; it's a feature that illustrates the point. Sovereignty at the infrastructure layer renders many regulatory concerns moot, because the individual controls the system rather than being subject to a third party's compliance with external rules.

Regulation is a necessary component of the overall response to compute concentration, but it isn't a substitute for infrastructure sovereignty. Laws change. Enforcement priorities shift. Regulatory capture is a well-documented phenomenon. A sovereignty strategy that depends on the continued benevolence and competence of regulatory bodies is not a sovereignty strategy. It's a hope.

5.5 VPNs, Tor, and the Privacy Theater Problem¶

Privacy tools like VPNs and the Tor network protect against certain categories of surveillance. They do not address compute dependency. A user who routes their AI queries through a VPN is still sending those queries to a centralized model provider. The provider still controls the model, still logs the interaction (or claims not to, with no verifiable guarantee), still shapes the response, and still retains the ability to modify or terminate the service.

The core issue is not whether the user's IP address is visible. The core issue is whether the user controls the compute that generates the response. Privacy without sovereignty is privacy at the pleasure of the infrastructure provider.

5.6 The Lesson: Sovereignty Is a Layer Problem¶

The consistent failure of application-layer solutions to deliver sovereignty is not a failure of engineering. It's a category error. Sovereignty is a property of the infrastructure layer, not the application layer. An application can provide privacy, security, and functionality, but only to the extent that the infrastructure on which it runs permits.

Building sovereignty at the application layer while renting infrastructure is analogous to building a house on leased land: the house is yours, but the ground it stands on is not, and the landowner can change the terms at any time.

This is the argument for sovereign compute. Not because cloud infrastructure is unreliable. Not because cloud providers are malicious. But because dependency on any entity whose interests are not identical to yours introduces a vulnerability that no application-layer tool can eliminate. The only remedy that addresses the problem at the correct layer is to own the infrastructure.

6. The Individual Imperative: Sovereign AI Stacks as Personal Infrastructure¶

6.1 The Practical Analog to Historical Self-Determination¶

Every epoch's concentration of foundational infrastructure produced a counter-movement of individuals and communities who sought to control their own means of production. In the agrarian epoch, this took the form of homesteading: the deliberate acquisition of land by individuals who rejected tenancy. In the industrial epoch, it took the form of rural electrification cooperatives, community banks, and mutual aid societies that built infrastructure where concentrated capital would not. In the early internet era, it took the form of self-hosted servers, open-source software, and the cypherpunk movement's insistence on encryption as a civil right.

The present moment demands an analogous response. The individual who builds and operates their own AI stack (a personal compute infrastructure capable of running inference, fine-tuning models, storing data, and serving applications without dependency on a centralized provider) is performing the equivalent of homesteading in the compute epoch.

6.2 What a Sovereign AI Stack Consists Of¶

A sovereign AI stack, at minimum, consists of the following components:

Compute hardware capable of running AI inference workloads: modern consumer GPUs, Apple Silicon systems, or purpose-built edge AI hardware
A locally hosted open-weight large language model from families including Llama 3.x, Mistral, Qwen 2.5, or DeepSeek V3
A local data store for documents, communications, and personal knowledge
An orchestration layer that routes queries, manages context, and coordinates multiple AI capabilities
A networking layer that enables the stack to communicate securely without traversing centralized infrastructure

This isn't science fiction. As of early 2026, a capable sovereign AI stack can be built for the cost of a mid-range laptop. An Apple Mac Mini with an M4 Pro chip and 64GB of unified memory runs models from 7 billion up to roughly 70 billion parameters (the largest at aggressive quantization) at usable inference speeds; a base M4 comfortably handles models up to around 32 billion parameters. A Hetzner VPS at approximately €25 per month provides remote compute for tasks that exceed local hardware capability. Open-weight models like DeepSeek V3, Llama 3.x, and Qwen 2.5 are freely available and perform competitively with proprietary models across a wide range of benchmarks.[^32]

The components exist. The capability is real. What's lacking is not technology but awareness and incentive structure.

6.3 The Practitioner's Role: A Historically Significant Service¶

If the argument presented in this paper is correct (that compute sovereignty is a prerequisite for digital autonomy, and that AI has made this sovereignty existentially important rather than merely convenient), then the person or organization that helps individuals and businesses build sovereign AI stacks is providing a historically significant service.

Consider the historical analogs. The surveyor who helped homesteaders identify and claim productive land provided a service whose value compounded for generations. The electrician who wired rural homes enabled participation in the industrial economy. The network engineer who helped organizations build their first internet presence in the 1990s enabled participation in the digital economy. Each of these roles was, at the time, understood as merely technical work, but each was, in practical terms, enabling a transition in the means of production.

The practitioner who helps a business build its own AI stack (selecting appropriate hardware, deploying open-weight models, configuring orchestration, integrating with existing workflows, and establishing data sovereignty over proprietary knowledge) is performing the same service. They are helping their clients transition from tenancy to ownership at the foundational layer of the digital economy.

The business case is straightforward: AI inference costs on centralized APIs scale linearly with usage and are subject to pricing changes by the provider. A self-hosted stack has a high initial cost but dramatically lower marginal cost, and the business retains control over its data, its model configurations, and its operational continuity. For any organization whose AI usage is substantial and ongoing, which increasingly means any organization, the arithmetic favors ownership before long. But cost is only the floor of the case. The recurring subscriptions a typical operator already carries, the productivity tools, the per-seat AI, the cloud, are rent paid in perpetuity to the same firms the work depends on, and they build no equity. Ownership converts that rent into an asset with near-zero marginal cost at the point of use. What it buys beyond the savings is the part that cannot be rented: autonomy over how the system runs, sovereignty over the data inside it, privacy by default, and the plain fact of owning the tool rather than borrowing it. The real question is whether you rent the means of your livelihood forever, or own it once.

6.4 The Window of Opportunity¶

The window for building sovereign compute infrastructure is open now but will not remain open indefinitely. Several converging factors make this moment critical.

Open-weight models are good enough now, and the gap is closing. The objection that a sovereign stack will always trail the frontier misreads what the frontier is for. There is a level of intelligence sufficient for almost all real work: drafting a contract, reconciling books, writing code, running a team of agents, and open-weight models have essentially reached it. The race past that threshold is a race to serve intelligence to billions of people at once, which is the hyperscalers' business and the reason their capital expenditure runs to hundreds of billions. It is not a requirement for running enough intelligence for one person, one household, or one firm. Serving the planet needs a data center; serving yourself and your agents needs a capable machine on a desk, and a million such machines, coordinated, begin to rival the data center on the workloads their owners care about. This is, in part, a bet that open models stay close enough to remain sufficient. The trend favors it: the distance between the best open and best closed models is now measured in months, not generations. The contingency runs the other way too. If regulation restricts open-weight releases, or if frontier hardware requirements outpace what individuals can own, the window narrows.

Consumer hardware is currently capable of running useful AI workloads. The rapid improvement in Apple Silicon, consumer GPUs, and edge AI chips has created a brief period in which the hardware required for personal AI infrastructure is affordable. If model architectures evolve to require hardware only data center operators can provide, the window closes.

The regulatory environment is still forming. Individuals and businesses who establish sovereign compute infrastructure now operate in a permissive environment. Future regulations may impose licensing requirements, reporting obligations, or operational constraints on private AI systems. Existing infrastructure provides a stronger foundation for defending these capabilities than attempting to build them after restrictions are in place.

The message to individuals and practitioners is clear: the time to build is now. Not because the political situation demands it (though it may), but because the economics of compute concentration make future action progressively more difficult and expensive.

7. The Economic Topology: From Sovereign Individuals to a Peer-to-Peer Economy¶

7.1 The Positive Vision¶

The defensive case is made: compute concentration creates dependency, application-layer fixes are insufficient, and sovereignty at the infrastructure layer is the remedy. That case tells you what to avoid. The positive case tells you what becomes possible: what a world of sovereign compute operators produces economically.

7.2 Individual Sovereignty Enables Economic Independence¶

An individual operating a sovereign AI stack is not merely protected from the risks of centralized dependency. They are equipped with capabilities that fundamentally alter their economic position. The solo consultant who runs their own models, manages their own data pipelines, and deploys their own agents is not competing with other solo consultants. They're competing with teams. The bookkeeper who serves clients through their own AI-augmented practice rather than through a platform intermediary retains the full value of their service rather than surrendering a percentage to an entity that controls the matching algorithm. The developer who builds on their own infrastructure ships on their own timeline, under their own terms, with their own data.

This isn't marginal productivity improvement. It's a shift in the relationship between the individual and the market. For the first time in the history of the digital economy, an individual with the right infrastructure can deliver enterprise-scale output (analysis, deliverables, coordination, client management) without enterprise-scale overhead and without platform-scale extraction.

The economic consequence is that the value created by AI-augmented work accrues to the operator rather than to the platform or the institution. This is the inverse of the consolidation path described in Section 3.10. Instead of AI automating the individual out of the value chain, AI amplifies the individual within it.

7.3 The Scope of the Argument¶

The capability-amplification thesis has the most direct force in domains where the primary input is cognitive labor and the primary output is information, analysis, or communication: consulting, professional services, legal research, financial analysis, software development, content production, education, and design. In these fields, a solo operator with a well-configured AI stack can plausibly deliver output competitive with small to mid-size teams. This is where the peer-to-peer economy becomes real, and where the individual's economic relationship to the market changes most quickly.

The scale of this domain makes the stakes clear. Service-producing industries account for roughly 70% of U.S. employment, and knowledge-intensive sectors have driven the majority of GDP growth over the past two decades.

Some domains combine cognitive work with physical delivery, capital requirements, or regulatory constraints. Healthcare combines cognitive diagnosis with regulated physical treatment. Real estate combines analytical work with legally complex transactions. Manufacturing design combines cognitive design with capital-intensive production. In these domains, sovereign compute amplifies the cognitive component without eliminating the non-cognitive constraints. The argument applies, but partially.

Manufacturing, construction, physical logistics, and infrastructure deployment have physical constraints that compute augments but cannot replace. A sovereign AI stack optimizes decisions (supply chain routing, production scheduling, demand forecasting), but it does not replace a factory floor or a construction crew. In these domains, the value of sovereign compute is decision-making autonomy, not production sovereignty.

These gradations matter for intellectual honesty, not for the thesis's macro significance. The domains where sovereign compute has the most direct force (knowledge work and professional services) represent the majority of employment and economic value creation in developed economies. The peer-to-peer economy this paper describes does not require every sector to transform. It requires the sectors where most people work and most value is created.

Stating these boundaries explicitly preempts the most common dismissal of the thesis: "You can't run a factory from a Mac Mini." Correct. The paper is not claiming you can. It claims that the majority of knowledge workers can own productive AI infrastructure that gives them enterprise-scale cognitive capability, and that this is sufficient to enable a peer-to-peer economy of meaningful scale.

7.4 The Missing Layer: Discovery and Exchange Without Extraction¶

Individual sovereign stacks solve the sovereignty problem for their operators. They do not, by themselves, solve the coordination problem. A thousand sovereign operators who cannot find each other, transact with each other, or verify each other's capabilities are a thousand islands: economically sovereign but economically isolated.

In the centralized model, this coordination function is performed by platforms. The platform matches buyers with sellers, handles payment, manages reputation, and sets the terms of engagement. In exchange, the platform extracts a percentage of every transaction and acquires the power to modify terms unilaterally.

The question is whether the coordination function can be performed without the extractive intermediary. Can sovereign individuals find each other, transact with each other, and build trust with each other through infrastructure that no single entity controls?

This is the question that connects individual sovereignty to the collective architecture. The answer is yes, but it requires building the connective tissue of a peer-to-peer economy: discovery patterns, reputation systems, exchange protocols, and the incentive structures that make participation self-sustaining. This isn't merely a technical problem. It's a design problem. And it's the problem that the collective layer is designed to solve.

7.5 The Alternative to the Platform Model¶

The peer-to-peer economy enabled by sovereign compute isn't merely a platform with lower fees. It's a different economic topology. In the platform model, all transactions route through a central node that captures data, sets prices, and controls access. In the peer-to-peer model, transactions occur directly between participants, with the network providing the infrastructure for discovery and trust without controlling the terms of any individual transaction.

The difference is analogous to the difference between a toll road and a public highway. Both connect the same destinations. But the toll road operator captures rent from every journey and can close the road at will. The public highway is maintained collectively and used freely. The economic outcomes for the individuals who depend on the road differ depending on which model prevails.

This is what makes the collective layer worth building: not merely shared compute, but an economic topology that differs from the extractive platform model at the foundational level.

7.6 A Note on Framing¶

This argument is structural, not populistic. It is not anti-corporate or anti-platform. Platforms have created enormous value and solved real coordination problems at scale. The observation is narrower: the economic consequences of infrastructure topology are predictable and measurable. Centralized compute produces centralized economic structures. Distributed compute enables distributed economic structures.

The paper presents both trajectories as consequences of architectural choices, not as moral narratives about good and evil actors. The reader should arrive at their own judgment about which path they prefer. The paper's job is to demonstrate that the choice exists and that it's a function of infrastructure control.

8. The Collective Architecture: Eno (Incentive Design for Decentralized Compute)¶

8.1 The Problem Individual Action Cannot Solve¶

Individual sovereign stacks don't address the systemic problem of compute concentration. What's missing is the incentive structure that converts individual sovereignty into collective infrastructure: a process that rewards individuals for contributing compute capacity to a shared network, enables coordination without centralized control, and creates the economic feedback loops that make the decentralized alternative self-sustaining.

The collective layer isn't merely shared compute; it's the peer-to-peer economy that replaces the platform model. What follows is the architecture designed to deliver it.

What distinguishes the Eno Project from prior decentralization efforts is a more precise diagnosis of what's actually missing. The components of a sovereign digital economy already exist. Open-source software has matured to the point where barriers to participation in the digital economy approach zero. Decentralized physical infrastructure networks demonstrate that distributed contributors can collectively provide capacity that rivals centralized alternatives. Tokenomic models exist that align contributor incentives with network health. Open-weight AI models run on consumer hardware today. The deficit isn't any of these components in isolation. The deficit is the hardware platform that makes all of them independent of the cloud layer, and the orchestration layer that connects them into a coherent, self-sustaining system.

The capital allocation logic that explains this gap is inherent rather than incidental. The entity with sufficient capital to fund foundational sovereign hardware infrastructure at scale is precisely the entity whose revenue model depends on the cloud dependency that hardware would disrupt. No suppression is required to explain the absence. Rational capital allocation produces it automatically. This isn't a market failure in the conventional sense. It's a predictable consequence of incentive structure. It's the same dynamic analyzed in Section 3.9 applied to investment rather than employment: the fiduciary obligation to maximize shareholder returns makes funding the infrastructure that would erode one's own moat the least rational available choice. The gap therefore cannot be closed by market forces alone. It requires a foundation-model (an entity whose mandate is explicitly not to maximize returns for a concentrated set of shareholders, but to deliver a public infrastructure good that the market is prevented from providing).

There is a precedent worth naming directly, because much of the audience for this argument lived it. The first decentralization movement, the one built around blockchains, was right about the principle and wrong about the substrate. It ran decentralized software on rented servers owned by the same handful of cloud providers this paper describes, a contradiction at the root of the project. And even where the will to self-host existed, the practice did not: running a real node or server demanded a level of technical skill that excluded almost everyone. The ethos was sound; the on-ramp was missing. Intelligence changes that. An AI assistant running on the device handles the configuration, maintenance, and troubleshooting that used to require a specialist. For the first time, the substrate can be genuinely owned by the people who use it, and operated by someone who is not an engineer. That is the unlock, and it is why this is possible now and was not before.

8.2 The Four Pillars of Digital Sovereignty¶

Digital sovereignty isn't a single technical problem with a single technical solution. It's a system, and systems have a property that collections of parts do not: remove any single load-bearing element, and the structure fails. The Eno Foundation organizes its work around four pillars that together constitute a complete architecture for digital sovereignty. They are pillars in the precise sense: each is necessary, none is sufficient, and together they form a structure that holds.

The order of the pillars reflects both the logic of dependency and the current state of the world. Three of the four already exist in meaningful form. One does not, and without it, the others cannot deliver sovereignty, only the appearance of it.

Software. Open-source software (including AI-generated code that further reduces the skill barrier to contribution) drives the cost of participating in the digital economy toward zero. It's a prerequisite for sovereignty because proprietary software, regardless of what hardware it runs on, reintroduces the dependency it was meant to escape. It's also an outcome: a world of sovereign hardware operators has a natural economic incentive to build and maintain open software, because open software compounds in value as the network of operators grows.

Intelligence. AI makes managing a personal compute node accessible to a mainstream user. The operational complexity that previously confined server management to specialists (configuration, maintenance, troubleshooting, optimization) is now addressable by an AI assistant running on the device itself. Modern chips, efficient model architectures, and advances in power and cooling have shrunk the hardware footprint from a server closet to a device that fits on a coffee table. Intelligence is both a prerequisite (it makes sovereignty operationally viable for non-specialists) and an outcome (sovereign operators retain the full value of AI-augmented productivity rather than surrendering it to a platform intermediary).

Network. A network of sovereign compute nodes, properly incentivized, provides the collective with access to the same scale of compute that centralized data centers provide, without concentrating ownership or control. Decentralized physical infrastructure networks demonstrate this is achievable. The critical design requirement is that the network must be genuinely sovereign at the node level: each participant owns and operates their hardware, and the network derives its resilience from that distributed ownership rather than from a central operator's reliability guarantees. The value created by the network flows to the participants who build and sustain it, not to the platform that connects them.

Infrastructure. This is the missing piece. The foundational hardware platform (designed for home deployment, built to last, upgradeable rather than disposable, and open in specification) is the only element of this system that does not yet exist in a form that enables true sovereignty. Without it, software runs on rented hardware, intelligence operates at the pleasure of a cloud provider, and the network is decentralized in topology but concentrated in physical dependency. The EnoStack is the Eno Foundation's specific contribution to this architecture: the hardware layer the market isn't positioned to fund, delivered by an entity whose mandate is precisely to fund it.

The four pillars are prerequisites and outcomes simultaneously. Build the hardware platform, and the software suite has a substrate worth building on. Add intelligence, and that substrate becomes accessible to anyone. Connect the nodes into a network, and the collective gains the compute density that has been the exclusive advantage of centralized data centers. The value that the digital economy generates flows to the people who participate in building it, rather than to the entities positioned as the mandatory intermediary between participants and the infrastructure they depend on.

8.3 Design Principles¶

The Eno Project's architecture is derived from four principles, each of which responds to a documented failure mode of prior decentralization efforts.

First: sovereignty must be physical, not merely logical. A decentralized network built on rented cloud infrastructure is decentralized in topology but concentrated in dependency. If the cloud provider terminates service, the network fails. The hardware layer (the EnoStack) is designed to be owned and operated by the individual participant, ensuring that no third party can unilaterally disrupt the network by withdrawing infrastructure.

Second: economic incentives must align contribution with reward. Prior decentralized compute projects have struggled with the free-rider problem: participants consume resources without contributing proportionally. The token economy is designed so that contribution of compute resources to the network is the primary way to earn participation rights. Nodes that provide verifiable compute earn tokens; tokens are required to access network services. The incentive is built in, not moral.

Third: the system must be open and auditable. A decentralized sovereignty initiative that requires trust in its operators has merely replaced one form of dependency with another. The Foundation's software stack will be open source, its token mechanics transparent, and its governance run through verifiable on-chain processes, so a participant can audit the system's behavior at every layer.

Fourth: the architecture must degrade gracefully. A decentralized network that requires full participation to function is fragile. The system is designed so that each node is independently useful: a personal sovereign compute stack that operates with or without network connectivity. The network layer adds capability (shared compute, distributed inference, collaborative training) but isn't required for basic operation.

8.4 The Hardware Layer: EnoStack¶

A clarification the rest of this section depends on: what follows describes the architecture the Foundation intends to build, not a system that already exists. The Eno Project is at its outset. It has the principles, the mission, and the design laid out here; execution is the work now beginning. And that execution is structured deliberately. This is not a capital-intensive product built in private and shipped finished. It is bootstrapped by design and built in the open, by a distributed community of contributors who earn tokens for the work they do, coordinated by the Foundation rather than owned by it. The hardware and software described below are specifications and reference designs: the blueprint the community is being convened to build.

EnoStack is the Foundation's reference design for a modular hardware platform built for home deployment. The design priorities are: sufficient compute for local AI inference (targeting 7B to 70B parameter models at interactive speeds); low power consumption suitable for residential electrical service; silent or near-silent operation; modular expansion capability (additional compute, storage, and networking modules); and open hardware specifications that enable third-party manufacturing and modification.

The EnoStack is not a data center appliance. It's a home appliance, designed with the same considerations of noise, power, and form factor that apply to consumer electronics. The design goal is that operating an EnoStack is no more complex or intrusive than operating a home router or NAS device.

Detailed specifications will be developed and released as the EnoStack Technical Reference, an open hardware specification under permissive licensing that any manufacturer can build from.

8.5 The Software Layer: EnoOS¶

EnoOS is the Foundation's open-source software stack, designed for three operational modes: standalone operation as a personal sovereign compute stack; local network operation for household or small-business multi-device coordination; and Eno network participation, in which the node contributes compute to and draws compute from the broader decentralized network.

As specified, the stack includes: a model serving layer supporting open-weight models across multiple frameworks; a data sovereignty layer providing encrypted local storage with user-controlled access policies; an orchestration layer for AI agent workflows, tool use, and multi-model routing; a network layer implementing peer-to-peer communication, distributed inference protocols, and reputation scoring; and a token interface that connects local compute contribution to the Eno token economy.

Source code will be released under the MIT License as it is written, in the open. The full architecture will be documented in the Eno Software Architecture Reference.

8.6 The Token Economy: Incentive Alignment¶

The Eno token ($ENO) is the incentive process that aligns the interests of everyone required to bring a decentralized compute network into existence and sustain it: developers who build the software, contributors who advance the platform, manufacturers who produce the hardware, and operators who provide compute to the network. It isn't a speculative financial instrument, a governance token with unbounded authority, or a general-purpose currency. It's how contribution is recognized and rewarded across every phase of the project's lifecycle.

The token economy operates across two phases, reflecting the reality that a decentralized network must be built before it can incentivize itself.

During the build phase, the Eno Foundation manages token distribution to fund development, incentivize early contributors, bootstrap community participation, and support the infrastructure required to bring the decentralized compute layer into existence. The foundation serves as the coordinating body during this period because there's no decentralized network yet to coordinate itself. The allocation model, vesting schedules, and distribution mechanics for this phase are set out in the Eno Tokenomics Reference.

As the decentralized infrastructure reaches operational maturity, the token economy transitions to programmatic, network-driven emission. Nodes that provide verifiable compute (serving inference, participating in distributed training, providing storage) earn $ENO in proportion to their verified contribution. Verification is performed through a proof-of-compute protocol. Users and applications that consume compute spend $ENO, with pricing determined by supply and demand within the network. Emission is gated by network-level utilization metrics to ensure token supply tracks actual economic activity rather than speculative demand.

The parameters governing network-phase emission (reward curves, rates, supply calibration) are determined at the point of network activation rather than predetermined. The variables that govern these decisions (hardware economics, compute pricing, network scale, model architecture requirements) change faster than any static model can anticipate. The architecture accommodates this calibration rather than foreclosing it. The network-phase emission model will be published as the infrastructure approaches operational readiness, calibrated to observed conditions and informed by the contributor community.

8.7 Why This Architecture Avoids Prior Failure Modes¶

Prior decentralized compute projects (including early file-sharing networks, blockchain-based compute marketplaces, and federated learning initiatives) have struggled with common failure modes: insufficient incentives for contributors, poor quality of service compared to centralized alternatives, lack of useful standalone functionality, and governance structures that centralized authority under the guise of decentralization.

The architecture is designed to address each of these. The token economy provides direct, proportional incentives for contribution. Quality of service comes from the reputation system and the ability of consumers to select among competing providers. Standalone functionality means participants receive immediate value from their hardware regardless of network state. Governance runs through on-chain voting weighted by verified compute contribution, not token holdings, which would recreate plutocratic concentration, and not identity, which would require centralized identity verification.

This is a claim that the architecture addresses the specific problems that have prevented prior decentralization efforts from achieving self-sustaining scale, and that it does so through design rather than ideology.

9. Conclusion: The Structural Choice¶

The argument presented in this paper is structural, not political. It doesn't depend on any particular view of government, any theory of corporate malice, or any prediction about which political party holds power in which jurisdiction. It depends on five empirical observations, one extension, and one logical inference.

Observation one: Compute infrastructure is concentrated. Three firms control 63% of cloud spending. One firm controls 85 to 87% of AI accelerator revenue. Capital barriers to entry exceed $100 billion annually per firm and are increasing. The four largest hyperscalers alone will spend more than $700 billion on AI infrastructure in 2026.

Observation two: AI systems mediate understanding and decision-making. They are not passive tools that execute user instructions. They are active mediators that shape outputs according to the priorities of their operators. This is an architectural property of how these systems work.

Observation three: The enterprise AI arms race is already underway and coercive by design. 72% of enterprises report AI in production. Competitive dynamics make adoption compulsory. Goldman Sachs estimates 300 million jobs globally are exposed to AI automation. The productivity surplus from AI follows a predictable two-tier path to the owners of capital and the owners of infrastructure, not to labor.

Observation four: Application-layer solutions cannot remedy infrastructure-layer dependencies. Encryption, privacy tools, data portability, and regulation operate within infrastructure they do not control. They mitigate symptoms; they cannot address the underlying condition.

Observation five: The human cost of this transition is not being managed. AI displacement is wage arbitrage applied to cognitive work, the same dynamic that hollowed out manufacturing communities, now executing at ten times the speed, against a population living in a subscription economy designed for income stability they're about to lose. The governance frameworks required to manage this transition do not exist. The people who will bear the cost are not in the rooms where the decisions are being made. This is the condition that produced forty years of inadequate response to the last transition, and that transition moved slowly enough to allow forty years of delay. This one does not.

Extension: AI concentration determines economic topology. The same infrastructure control that creates epistemic dependency creates economic dependency. The publicly traded corporation's legal architecture guarantees it will deploy AI to optimize shareholder return. Centralized compute produces a platform economy where AI automates labor and surplus accrues to platform operators. Distributed compute enables a peer-to-peer economy where AI amplifies individual capability and value accrues to the operator.

Logical inference: If compute is the foundational layer of the digital economy, and AI has made compute both an epistemic infrastructure that mediates understanding and an economic infrastructure that determines how value is created and distributed, and application-layer solutions are insufficient to ensure autonomy, then sovereignty at the compute layer is a prerequisite for both digital autonomy and economic agency. This isn't a preference. It's a deduction.

The choice before us isn't sovereignty versus dependency in the abstract. It's a choice between two concrete futures. In one future, AI consolidates economic power in platforms that automate labor and extract rent. The individual becomes a dependent of systems they do not control, working on terms they did not set, for a share of value determined by an algorithm they cannot inspect. In the other future, AI distributes economic capability to individuals who transact directly with each other: sovereign operators in a peer-to-peer economy, choosing their own clients, setting their own terms, retaining the value they create.

The defensive case and the positive case are two sides of the same argument. The defensive case says: avoid dependency, because the entity that controls your infrastructure controls your autonomy. The positive case says: build sovereignty, because the individual who controls their infrastructure can participate in an economy that is only now becoming possible, a peer-to-peer economy of AI-augmented sovereign operators that is the alternative to platform consolidation.

The individual response is to build sovereign AI infrastructure now, while the window of technological feasibility and regulatory permissiveness remains open. The components exist: open-weight models, capable consumer hardware, open-source orchestration tools. What's required is awareness, effort, and the willingness to prioritize ownership over convenience.

The collective response is to build the incentive structures that convert individual sovereignty into a self-sustaining decentralized network, and into the connective tissue of a peer-to-peer economy. The Eno Project is one such effort: designed on open principles, published for public scrutiny, and built on the premise that structural problems require structural solutions.

The reader who disagrees with this argument is invited to identify the specific premise that is false: that compute is concentrated, that AI mediates epistemology and economic structure, that the enterprise arms race is already in motion, that application-layer solutions are insufficient, that the human cost of AI displacement is being absorbed without institutional management, or that infrastructure sovereignty is the logical remedy. If all premises hold, the conclusion follows.

If the conclusion follows, the question isn't whether to act. It's when. And the answer, given the dynamics of increasing concentration, the closing window of opportunity, and the pace of a displacement that doesn't allow forty years for a delayed response, is now.

Sovereign Compute

The Eno Project, theenoproject.com

Version 1.0, June 2026

Notes and References¶

[^1]: J.M. Neeson, Commoners: Common Right, Enclosure and Social Change in England, 1700–1820, Cambridge University Press, 1993; Robert C. Allen, Enclosure and the Yeoman: The Agricultural Development of the South Midlands, 1450–1850, Clarendon Press, 1992. Enclosure acreage: UK Parliament, "Enclosing the land," https://www.parliament.uk/about/living-heritage/transformingsociety/towncountry/landscape/overview/enclosingland/.

[^2]: Daron Acemoglu, Simon Johnson, and James A. Robinson, "The Colonial Origins of Comparative Development: An Empirical Investigation," American Economic Review 91(5):1369–1401, 2001, https://www.aeaweb.org/articles?id=10.1257/aer.91.5.1369.

[^3]: Ron Chernow, Titan: The Life of John D. Rockefeller, Sr., Random House, 1998. Standard Oil market share (≈91% of refining, 85% of sales, 1904): U.S. Bureau of Corporations, Report on the Petroleum Industry, 1906–07; 1911 dissolution into 34 companies: Standard Oil Co. of New Jersey v. United States, 221 U.S. 1 (1911), https://civics.supremecourthistory.org/article/standard-oil-company-v-united-states/.

[^4]: Forrest McDonald, Insull, University of Chicago Press, 1962; Richard Munson, From Edison to Enron: The Business of Power and What It Means for the Future of Electricity, Praeger, 2005.

[^5]: Synergy Research Group, "The Big Three Together Hold 63% of the Cloud Market," November 2025, https://www.srgresearch.com/articles/cloud-market-share-trends-big-three-together-hold-63-while-oracle-and-the-neoclouds-inch-higher (Q3 2025 market value $106.9B). IaaS shares (top three ≈71%, AWS ≈38%): Gartner, "Worldwide IaaS Public Cloud Services Market Grew 22.5% in 2024," August 6, 2025, https://www.gartner.com/en/newsroom/press-releases/2025-08-06-gartner-says-worldwide-iaas-public-cloud-services-market-grew-22-point-5-percent-in-2024.

[^6]: Synergy Research Group, quarterly cloud market data, November 2025 (Big Three combined share rose 61%→63% over eight quarters), https://www.srgresearch.com/articles/cloud-market-share-trends-big-three-together-hold-63-while-oracle-and-the-neoclouds-inch-higher.

[^7]: Full-year 2025 cloud infrastructure revenue (~$419B): Synergy Research Group, 2026, https://thetechcapital.com/2025-cloud-spending-hit-419-billion-as-growth-rates-soar/; public cloud end-user spending forecast ($723.4B, 2025): Gartner, November 19, 2024, https://www.gartner.com/en/newsroom/press-releases/2024-11-19-gartner-forecasts-worldwide-public-cloud-end-user-spending-to-total-723-billion-dollars-in-2025.

[^8]: "Charted: The Battle for AI Data Center Revenue (2021–2025)," Visual Capitalist (data: Bloomberg and company filings), https://www.visualcapitalist.com/charted-the-battle-for-ai-data-center-revenue-2021-2025/.

[^9]: NVIDIA Corporation, Third Quarter Fiscal 2026 results (data-center revenue $51.2B; gross margin 73.4% GAAP / 73.6% non-GAAP), SEC filing, https://www.sec.gov/Archives/edgar/data/0001045810/000104581025000228/q3fy26pr.htm; Fiscal 2025 full-year revenue ($130.5B), https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2025.

[^10]: AMD AI-chip share and hyperscaler custom-ASIC projections: Silicon Analysts, "AMD vs NVIDIA AI GPU Market Share 2026," 2026, https://siliconanalysts.com/analysis/amd-vs-nvidia-ai-gpu-market-share-2026 (analyst estimates; ranges vary by methodology).

[^11]: Four-hyperscaler 2026 capital-expenditure consensus (~$700–725B): CNBC, February 6, 2026, https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html; Tom's Hardware, "Big tech's AI spending plans reach $725 billion," 2026, https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion.

[^12]: Per-firm 2026 capex guidance: company earnings calls, Q4 2025–Q1 2026, summarized by CNBC (February 6, 2026). Microsoft capacity/power constraints and commercial backlog (remaining performance obligations ≈$625B, Q3 FY2026): https://www.globaldatacenterhub.com/p/microsoft-q3-fy2026-the-190b-capex.

[^13]: Hyperscaler debt issuance ($121B in 2025, of which >$90B in Q4; Morgan Stanley and JPMorgan projection of up to ~$1.5T in new technology-sector debt): Mellon Investments Corporation, "Record-Breaking AI-Related Debt Issuance in 2025," 2025, https://www.mellon.com/insights/insights-articles/record-breaking-ai-related-debt-issuance-in-2025.html.

[^14]: Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, and David Owen, "The Rising Costs of Training Frontier AI Models," arXiv:2405.21015 (Epoch AI), 2024, https://arxiv.org/abs/2405.21015; cost detail (GPT-4 ≈$40M amortized, ≈$800M cluster; >$1B runs by 2027): Epoch AI, "How Much Does It Cost to Train Frontier AI Models?", https://epoch.ai/publications/how-much-does-it-cost-to-train-frontier-ai-models.

[^15]: Josh You, "How Does OpenAI Allocate Its Compute?", Epoch AI, 2025 (of ~$5B in 2024 R&D compute, under $1B went to final released-model training), https://epochai.substack.com/p/how-does-openai-allocate-its-compute.

[^16]: "Announcing The Stargate Project," OpenAI, January 21, 2025, https://openai.com/index/announcing-the-stargate-project/ (joint venture of OpenAI, SoftBank, Oracle, and MGX; $500B over four years, $100B initial).

[^17]: "Gartner Predicts 40% of Enterprise Applications Will Feature Task-Specific AI Agents by 2026," Gartner, August 26, 2025, https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-applications-will-feature-task-specific-ai-agents-by-2026. Enterprise AI-in-production figure (72%) traces to McKinsey, The State of AI survey, compiled in MedhaCloud, "AI Adoption Statistics for 2026," March 2026, https://www.medhacloud.com/blog/ai-adoption-statistics-2026.

[^18]: Goldman Sachs Research (Joseph Briggs and Devesh Kodnani), "The Potentially Large Effects of Artificial Intelligence on Economic Growth," March 2023 (≈300 million jobs globally exposed to automation; 6–7% of the US workforce at displacement risk), reported by CNBC, March 28, 2023, https://www.cnbc.com/2023/03/28/ai-automation-could-impact-300-million-jobs-heres-which-ones.html; reaffirmed in Goldman Sachs, "How Will AI Affect the US Labor Market?", Goldman Sachs Insights, March 2026.

[^19]: Daron Acemoglu and Pascual Restrepo, "Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity," NBER Working Paper No. 32536, 2024 (rev. 2025), https://doi.org/10.3386/w32536.

[^20]: David Autor and B.N. Kausik, "Resolving the Automation Paradox: Falling Labor Share, Rising Wages," LSE Programme on Innovation and Diffusion (POID) Working Paper No. 140, February 9, 2026, https://poid.lse.ac.uk/; see also Goldman Sachs Research, August 2025.

[^21]: Dodge v. Ford Motor Co., 204 Mich. 459, 170 N.W. 668 (1919); Milton Friedman, "The Social Responsibility of Business Is to Increase Its Profits," The New York Times Magazine, September 13, 1970. (Stakeholder-capitalism frameworks contest this as a normative matter; none suggest firms should forgo AI-driven efficiency gains to preserve employment at competitive cost.)

[^22]: Author's analysis of SaaS market-segment reporting, 2025–2026, noting observable displacement of point-solution vendors across legal-tech, content, and customer-success categories. Illustrative rather than a single attributable source.

[^23]: AI 2027, https://ai-2027.com, April 2025 (Daniel Kokotajlo et al.; authors include former OpenAI researchers).

[^24]: David H. Autor, David Dorn, and Gordon H. Hanson, "The China Syndrome: Local Labor Market Effects of Import Competition in the United States," American Economic Review 103(6):2121–2168, 2013, https://www.aeaweb.org/articles?id=10.1257/aer.103.6.2121; Anne Case and Angus Deaton, Deaths of Despair and the Future of Capitalism, Princeton University Press, 2020 (ISBN 9780691190785).

[^25]: Acemoglu and Restrepo, "Automation and Rent Dissipation," NBER Working Paper No. 32536, 2024 (rev. 2025), https://doi.org/10.3386/w32536; Goldman Sachs Research, March 2026.

[^26]: C+R Research, "Subscription Service Statistics and Costs," 2024 ($219/month actual spend vs $86 self-estimate; 74% say charges are easy to forget), https://www.crresearch.com/blog/subscription-service-statistics-and-costs/; West Monroe, "Americans Are Spending More on Subscriptions and Are Less Aware of Spending" ($273/month), 2021, https://www.westmonroe.com/press-releases/americans-are-spending-more-on-subscriptions-and-are-less-aware-of-spending. Adobe Creative Cloud transition (2013): https://www.dpreview.com/articles/3716254152.

[^27]: Mrinank Sharma, public resignation statement, X, February 9, 2026, reported by eWeek, https://www.eweek.com/news/ai-safety-leader-resigns-anthropic-global-risks/, and The Hill, https://thehill.com/policy/technology/5735767-anthropic-researcher-quits-ai-crises-ads/; Zoë Hitzig, "OpenAI Is Making the Mistakes Facebook Made. I Quit," The New York Times, February 2026; OpenAI mission-alignment team dissolution: Platformer, February 2026, https://www.platformer.news/openai-mission-alignment-team-joshua-achiam/, and TechCrunch, February 11, 2026, https://techcrunch.com/2026/02/11/openai-disbands-mission-alignment-team-which-focused-on-safe-and-trustworthy-ai-development/.

[^28]: Jan Leike, resignation statement, X, May 17, 2024 ("safety culture and processes have taken a backseat to shiny products"), reported by CBS News, https://www.cbsnews.com/sanfrancisco/news/openai-exec-jan-leike-resigns-says-safety-has-taken-a-backseat/; part of a broader pattern of OpenAI safety-team departures, 2024–2026.

[^29]: Anthropic, statement on its Department of Defense contract, anthropic.com, February 27, 2026; CNBC, February 27, 2026, https://www.cnbc.com/2026/02/27/trump-anthropic-ai-pentagon.html; NPR, "OpenAI announces Pentagon deal after Trump bans Anthropic," February 28, 2026, https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban; Electronic Frontier Foundation, "The Anthropic-DoD Conflict: Privacy Protections Shouldn't Depend on the Decisions of a Few Powerful People," March 6, 2026, https://www.eff.org/deeplinks/2026/03/anthropic-dod-conflict-privacy-protections-shouldnt-depend-decisions-few-powerful; The Intercept, "OpenAI on Surveillance and Autonomous Killings: You're Going to Have to Trust Us," March 8, 2026, https://theintercept.com/2026/03/08/openai-anthropic-military-contract-ethics-surveillance/.

[^30]: Nick Lichtenberg, "Sam Altman admits AI is killing the labor-capital balance—and says nobody knows what to do about it," Fortune, March 12, 2026, https://fortune.com/2026/03/12/sam-altman-ai-labor-capital-jobs-nobody-knows/ (remarks delivered at the BlackRock Infrastructure Summit, Washington, D.C., March 11, 2026).

[^31]: Amazon Web Services termination of Parler hosting, effective January 10, 2021, reported by The Washington Post, https://www.washingtonpost.com/technology/2021/01/09/amazon-parler-suspension/, CNBC, and the Associated Press. Signal's reliance on AWS infrastructure confirmed publicly in October 2025.

[^32]: Open-model capability comparisons: LMSYS Chatbot Arena (LMArena) leaderboard, 2026, https://lmarena.ai/; Epoch AI capability index, https://epoch.ai/. Open-weight model releases: Llama 3 (Meta), DeepSeek V3 (DeepSeek), Qwen 2.5 (Alibaba). Apple Silicon local-inference performance and Hetzner cloud pricing per vendor documentation, 2026.