Executive Summary
The personal computer is being quietly demoted to a screen. Over the past three years, the locus of meaningful computation has migrated from devices owned and controlled by users to a small set of hyperscale data centers operated by five American firms and one Taiwanese fabricator. By April 2026, every flagship consumer AI experience — Microsoft Copilot, Apple Intelligence’s “complex requests,” Gemini, ChatGPT, Claude — runs primarily in someone else’s data center, even on hardware marketed as an “AI PC.” The combined Big Three cloud providers hold 63% of global cloud infrastructure and rising. NVIDIA holds roughly 80% of AI accelerator revenue. TSMC fabricates roughly 90% of leading-edge logic at facilities concentrated within 110 miles of the Taiwan Strait. The top five hyperscalers will spend roughly $600–700 billion on capex in 2026, three-quarters of it on AI infrastructure — a sum approaching 2.2% of US GDP. None of this is accidental, and none of it is reversible by market forces alone.
This paper argues that the transition from local to centralized computation is the most consequential structural transformation in the human relationship to machines since the displacement of artisanal labor by the factory system, and that it carries five distinct civilizational risks that together compound into a single threat: the loss of computational sovereignty for individuals, communities, and nations. Those risks are economic concentration and rent extraction at a scale exceeding the railroad and oil trusts; surveillance and coercion infrastructure on a scope that makes East German Stasi capability look quaint; supply-chain fragility centered on a single fab on a contested island; the documented atrophy of cognitive capacity in populations that offload reasoning to remote systems they cannot inspect; and the disappearance of the off- grid-capable personal computer as a civilizational fallback. The window for shaping this trajectory is closing faster than the policy apparatus is moving. The Digital Markets Act, current US antitrust cases, and existing right-to-repair laws — even taken together — do not reach the structural problem. They regulate the symptoms of platform power while the underlying enclosure of compute proceeds unimpeded.
The recommendations that follow are structural, not cosmetic. They include a federal local-computation guarantee for any device sold for general use; a public option for compute analogous to the government-owned, contractor-operated (GOCO) model used in defense; mandatory openness and portability standards for capability-class AI model weights; tax treatment that ceases to subsidize the conversion of perpetual ownership into recurring rent; right-to-repair and right-to-modify legislation extended to the AI/computing stack; federal procurement preferences for open and locally- modifiable systems; and an education policy that explicitly preserves the development of unassisted reasoning. The premise of this paper is that compute is now a fundamental infrastructure of cognition, citizenship, and continuity, and that allowing it to be enclosed into a handful of private balance sheets while a generation grows up unable to reason without remote assistance is a civilizational error of the same magnitude as letting the railroads write their own rate schedules — but with consequences orders of magnitude harder to reverse.
I. The transformation: where we are in April 2026 The trajectory of personal computing tells a story in three acts. The first, from roughly 1977 to 1995, was the emancipation of computation: the Apple II, the IBM PC, the Mac, and finally the commodity x86 box made the most powerful intellectual tool ever devised available for purchase, modification, and ownership by individual people. The second act, from 2007 to 2022, was the cloud era: networked services hosted on someone else’s machines came to dominate most consumer applications, but the device on the desk still booted, computed, and stored data locally. The third act, which began with the public release of ChatGPT in November 2022 and accelerated through the launch of Apple Intelligence (2024), Microsoft Copilot+ PCs (May 2024), and Project Stargate (January 2025), is the cognitive recentralization: the relocation of the most economically and intellectually consequential computation — large-language-model inference and the agentic workflows built atop it — to a small number of hyperscale facilities owned by a small number of firms.
The architecture of this transition is now visible in detail. Apple Intelligence on a current iPhone or Mac ships a roughly 3-billion-parameter on-device model (with novel 2-bit quantization-aware training and KV-cache sharing) for latency-sensitive utility tasks: notification summaries, smart replies, Genmoji, light writing assistance. Anything more substantial is routed to Apple’s Private Cloud Compute — a server-side mixture- of-experts model running on Apple Silicon servers — or, if the user accepts a confirmation prompt, to OpenAI’s ChatGPT entirely outside Apple’s trust boundary. Apple’s Private Cloud Compute is the most cryptographically sophisticated privacy architecture deployed at consumer scale: stateless computation, signed images, public transparency logs, no privileged runtime access, OHTTP relay between client identity and request payload. It is also, fundamentally, a cloud architecture. The data leaves the device. The computation does not run locally. Whatever the cryptographic guarantees, the device is a thin client for the part of the experience that users find most useful.
Microsoft’s strategy is more nakedly cloud-first. The Copilot+ PC specification — a neural processing unit of at least 40 trillion operations per second, 16 gigabytes of memory, and a Microsoft-curated silicon list — is genuine engineering, but the headline Copilot key on the keyboard launches an Azure-hosted GPT-class model. Microsoft 365 Copilot is almost entirely cloud: its orchestrator queries the Microsoft Graph for tenant data, grounds the prompt in a tenant-wide semantic index, and sends the package to Azure OpenAI, where, since March 2026, GPT-5.4- class models handle the inference. The genuinely local features on a Copilot+ PC — Recall, Studio Effects, Live Caption translation, Cocreator in Paint — are narrow utilities. Microsoft Recall itself, the flagship local feature, has spent 18 months in unresolved security controversy: pulled in June 2024 after researchers demonstrated unencrypted SQLite extraction, iTnews relaunched April 2025 with virtualization-based-security enclaves and Hello-gated decryption, iTnews and in March 2026 again shown to be extractable by user-context malware. As of Recall’s first anniversary in production, fewer than ten percent of Windows 11 PCs can run it.
Google’s Gemini Nano runs on more than 140 million Android devices through the AICore service, but Google’s flagship AI experiences — Gemini Live, Project Astra, Deep Research, Gemini in Workspace — are cloud-served, and the trajectory of consumer Android places voice assistants, agentic flows, and the Gemini app itself on Google’s servers. ChromeOS is the limit case: a browser plus a kernel, with computation in the browser tab and the model in the data center.
The economics make local inference progressively less competitive in absolute terms even as on-device hardware improves. Per, cost per token for equivalent capability has fallen between nine-fold and nine-hundred-fold per year
since 2023; the entire cloud API price stack fell roughly 80% between early 2025 and early 2026. Meanwhile, frontier models have grown into the 400-billion-to-1.5-trillion-parameter range as sparse mixtures of experts. A consumer RTX 5090 with 32 gigabytes of GDDR7 — the most capable consumer GPU available, when it can be found, at a street price of $2,500 to $4,000 — can run a quantized 30-billion-parameter dense model comfortably. It cannot, by orders of magnitude, run a frontier closed model. A Mac Studio with 512 gigabytes of unified memory (about $9,500) can hold DeepSeek V3’s 671 billion parameters at four- bit quantization, but at roughly six tokens per second — adequate for hobbyist work, not for the agentic workflows that closed frontier models support. The gap between what fits on a desk and what runs in a data center is widening, not narrowing, on the dimensions that matter most for capability.
The open-weight counter-current is real and important — and it is largely Chinese. The April 2026 leaderboards are dominated by DeepSeek V3.2, Kimi K2 from Moonshot, Qwen 3.5/3.6 from Alibaba, GLM-5.1 from Zhipu (which now beats Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro), Mistral Large 3, Meta’s Llama 4 series, and OpenAI’s belated GPT-OSS open-weight release. Quantization research — most strikingly Microsoft Research’s BitNet b1.58 ternary-weight architecture, which compresses models to roughly 1.58 bits per weight at near-FP16 quality — points toward a future in which capable models do fit on personal hardware. But the consumer-runnable open tier remains a generation behind the closed frontier on the hardest reasoning and agentic tasks, and the gap is sustained by a roughly $600 billion annual capex flow that no individual or community can match.
The result, in April 2026, is a hybrid architecture in which the visible intelligence lives in the cloud and the device is reduced to a sensor, a screen, and a small auxiliary brain. This is not a transient configuration. It is the equilibrium toward which the entire industry is converging.
II. The frame: this is structural, not preferential The standard response to the foregoing is that markets reflect preferences. Users prefer cloud AI because it works better. They prefer subscriptions because they smooth costs. They prefer locked-down devices because the devices are more secure. The recentralization is the consumer’s choice, and policy intervention is paternalism. This response is wrong, and to see why requires a frame older than computing.
Pierre-Joseph Proudhon, writing in 1840, distinguished possession — the right in a thing, grounded in actual occupation and use — from property — the absentee right of the thing, grounded in title and exercised through rent, withholding, or revocation. A farmer who works the land he stands on holds the land in possession; a landlord who has never seen the land but receives rent from it holds the land as property.
Proudhon’s claim, controversial in 1840 and almost forgotten in the twentieth century, was that the difference is not technical but moral and political: possession constitutes a relationship between a person and a thing, while property constitutes a relationship of power between persons mediated by a thing. The shift from one to the other is a downgrade in the structure of rights even when the user does not notice it.
Mapping this onto computation is straightforward and damning. A 1995 Pentium running locally compiled software was held in possession: the owner could open it, modify it, repurpose it, run it offline, sell it, hand it to a child. A 2026 iPhone running Apple Intelligence with a ChatGPT fallback is held — at best — in licensed tenancy. Apple controls the bootloader; Microsoft controls the firmware (via Pluton on AMD, Intel, and Qualcomm SoCs and via Windows Update on the Pluton firmware itself); Adobe controls whether yesterday’s purchased software runs tomorrow; OpenAI controls whether the model that wrote this morning’s email exists this evening; Amazon, after February 2025, no longer permits even the download-and-USB-transfer escape hatch for purchased Kindle books. The user is a tenant, and the lease is at-will. This is a regression from possession to property of precisely the kind Proudhon described, executed at civilizational scale and at the speed of software updates.
Ivan Illich, writing in 1973, gave us the second necessary frame. Tools for Conviviality distinguishes convivial tools — autonomy-enhancing, transparent, repairable, accessible to non-experts, “permitting ample latitude in use” — from manipulative tools that require institutional or expert mediation and dis-able the user. Illich identified two thresholds in the development of any tool: the first at which it genuinely solves a problem, the second past which it “impairs the goals it was meant to serve.” His core warning was about radical monopoly: not the dominance of one brand within a category, but the dominance of one type of product that excludes alternatives, as the automobile did not by outcompeting other transit but by reshaping cities so that walking and cycling became impossible. Centralized AI is the radical monopoly of the next two decades. The question is not whether GPT-5 is better than DeepSeek V3 but whether the entire built environment of work, education, and civic life will be reorganized on the assumption of continuous high-bandwidth access to a frontier model in a hyperscaler data center, at which point the local alternative is not slower; it is simply outside the assumed infrastructure of life.
Lewis Mumford, in his 1964 Technology and Culture essay “Authoritarian and Democratic Technics,” gave the third frame. Two technologies have coexisted since the dawn of civilization: democratic technics, small-scale, locally controlled, distributed, durable, b2o “resourceful but relatively weak,” and authoritarian technics, large- scale, centralized, hierarchical, “system-centered, immensely powerful, but inherently unstable.” Mumford warned that authoritarian technics had finally overcome its historical weakness — its dependence on “resistant, sometimes actively disobedient servo-mechanisms, still human enough to harbor purposes that do not always coincide with those of the system.” The hyperscale data center, the frontier model trained on a thousand-GPU cluster at $100 million per run, and the locked-down terminal in the user’s hand are precisely the technical architecture Mumford predicted. He called the architects of such systems “the pyramid builders of our own age.” Stargate, with its $500 billion of announced infrastructure, eight-figure permits, and dedicated nuclear baseload, is the literal pyramid.
These three frames — Proudhonian possession-vs-property, Illichian conviviality-vs- radical-monopoly, Mumfordian democratic-vs-authoritarian-technics — are not decorative. They name a structural pattern that markets, by themselves, do not correct, because the pattern operates by capturing the very space in which alternatives would emerge. To these frames, recent decades add Jaron Lanier’s siren server (the asymmetric computational node that takes in vastly more information than it gives out, externalizes risk, and presents itself as free); Cory Doctorow’s enshittification (the three-stage decay by which platforms first court users, then exploit them in service of business customers, then exploit business customers for shareholder rents) remio and his earlier war on general-purpose computing (the prediction, made at 28C3 in December 2011, that the copyright wars were a skirmish and the real fight was over whether general-purpose computers would be replaced by appliances that obey the manufacturer rather than the owner); Shoshana Zuboff’s instrumentarian power (knowing and shaping behavior through ubiquitous computational architecture); and James C. Scott’s legibility (the simplification of complex local realities into forms a central authority can read and govern). Each names a facet of the same shape. Centralized AI is the synthesis of all of them: the siren server with cognition added, the enshittification engine with no exit, the legibility infrastructure that can read not only what you do but what you are about to think.
III. The five threat vectors
1. Economic Concentration and the New Rentier Class
The economic transformation underway is the most rapid and concentrated wealth- capture event in modern history, and its scale is not yet appreciated outside the firms executing it. The Big Three cloud providers — AWS, Azure, Cloud — held 63% of the global cloud infrastructure market in Q3 2025, up from 61% two years earlier. Concentration is rising, not falling. NVIDIA’s share of AI accelerator revenue peaked near 87% in 2024 and is projected to settle around 75% only because the market is doubling, not because competitors are gaining ground; CUDA lock-in, twenty years deep, is the moat. The five US hyperscalers will spend roughly $600–700 billion on capital expenditure in 2026,
of which roughly 75% — about $450 billion — is AI-specific. They have raised $108 billion in new debt in 2025 alone, with Morgan Stanley and JPMorgan projecting up to $1.5 trillion in further sector debt issuance.
Capital intensity at these firms now reaches 45–57% of revenue, a ratio
describes as “historically unthinkable.” Goldman Sachs projects cumulative 2025–2027 hyperscaler capex of $1.15 trillion, more than twice the
2022–2024 figure. McKinsey estimates $7 trillion of data-center investment required by 2030. OpenAI raised in April 2025 at $300 billion post-money, sold secondaries in October at $500 billion, and by April 2026 traded at implied valuations near $850 billion on. Anthropic, valued at $61.5 billion in March 2025, closed a Series G at $380 billion in February 2026; secondary markets imply trillion-dollar marks. Per Ramp, 35.2% of US businesses paid for OpenAI and 30.6% for Anthropic by March 2026 — a near-duopoly that did not exist eighteen months earlier. The frontier-lab oligopoly is now five firms (OpenAI, Anthropic,, Meta, xAI) capturing essentially all enterprise AI spend. Microsoft’s commercial remaining performance obligation reached $392 billion in mid- 2025, a multi-year locked-in revenue backlog that exceeds the GDP of all but about thirty countries.
The historical analog is unmistakable but, as analog, understated. The railroad trust era of the 1870s and 1880s produced the Interstate Commerce Act of 1887 and, ultimately, the antimonopoly politics that culminated in the Standard Oil dissolution of 1911. The transcontinentals, in Richard White’s reading, were the first corporate behemoths, sustained by government largesse, generating new forms of corruption, and imposing discriminatory rates on a captive shipper base. The hyperscaler-AI complex of 2026 reproduces every feature of that pattern at vastly greater scale: dependence on government largesse (the CHIPS Act, defense contracts, $1 billion DOE loans for nuclear restart, hyperscaler-utility bilateral deals that shift grid costs to ratepayers), discriminatory pricing (compute access, API tiers, capability gating), and an emergent antimonopoly politics — except that the antimonopoly politics is, so far, losing.
The subscription transformation is the rent-extraction mechanism that converts this concentration into recurring household and business outlays. Adobe completed the perpetual-to-subscription transition in 2013 and reached $21.5 billion in revenue in fiscal 2024, a sustained annuity flow. Microsoft 365 Personal at $99.99 a year compounds, over a decade, to roughly $1,000 against a one-time $250 for the equivalent perpetual Office Home & Business 2024. 4sysops Autodesk’s elimination of perpetual licenses in 2016 produced a ten-year total cost of ownership for AutoCAD of roughly $19,000 against the $4,000–$5,000 of the prior model. ChatGPT Plus, Pro, Enterprise; Claude Pro, Max; Gemini Advanced; Copilot Pro; GitHub Copilot — each adds another monthly debit. The aggregate effect is the conversion of household and small-business IT spend from periodic capital purchase to continuous tribute, with the corresponding shift in power: the buyer who has paid in full and walked away is replaced by the tenant whose access to professional capability depends on remaining current with five or six remote landlords.
The historical analog that best captures this is not the railroad or the oil trust but the enclosure movement that converted the English commons to private property between roughly 1604 and 1914. Enclosure, in E. P. Thompson’s reading, was not principally a market phenomenon; it was a legal-political event — Parliamentary Enclosure Acts, the 1723 Black Act making customary forest uses capital crimes — that re- categorized a shared resource as private property and the people who had used it as trespassers. The cloud-AI transition is the enclosure of compute, of model weights, of the digital commons of personal data, and ultimately of cognition itself. The historical pattern is that enclosure, once accomplished, is not reversed by markets, only by politics.
2. State Coercion and the Surveillance Substrate
Every keystroke routed to a cloud API is a record. Every prompt to a chatbot is, at the operator’s discretion, a stored and indexable artifact. Every agentic workflow that drafts an email, plans a trip, researches a topic, or explores an idea is an entry in a database that — under existing US law — is generally accessible to law enforcement with a subpoena or, under Section 702 of the FISA Amendments Act, in many cases without one. The legal architecture surrounding remote-stored content has been built around the third-party doctrine of Smith v. Maryland (1979) and the diminished Fourth Amendment expectation in data voluntarily disclosed to a service provider. The Stored Communications Act provides modest protections, but the trajectory of compelled- process volume — Apple,, Microsoft, and Meta collectively respond to hundreds of thousands of US government requests per year — points in one direction.
The qualitative novelty of cloud AI is that the thinking adjacent to writing now passes through the same surveillance substrate as the writing itself. An author’s drafts, a journalist’s source-development notes, a dissident’s policy explorations, a lawyer’s hypothetical case theory, a researcher’s intellectual experiments — all of which were previously held in the privacy of an unmonitored local machine — are now, by default, processed in a remote system whose logs are subject to compelled production, internal access, employee misuse, and the compliance regime of whichever jurisdiction the data center happens to sit in. The East German Stasi at peak deployed roughly 90,000 full- time officers and 170,000 unofficial collaborators against a population of seventeen million. Their primary technical limitation was the physical impossibility of reading and indexing what they collected. That limitation is now solved. A frontier large- language model can ingest, classify, and cross-reference a population’s intellectual output at marginal cost approaching zero. The infrastructure to do so is the same infrastructure on which Microsoft Copilot, Apple Intelligence, and Gemini run.
Apple’s Private Cloud Compute, with its cryptographic transparency logs and stateless computation, represents the most determined effort in industry to provide architectural privacy guarantees at scale. It is also a single firm’s design, subject to a single firm’s continued willingness to maintain it, and entirely outside the scope of the ChatGPT fallback that Apple Intelligence routes to OpenAI when its own models fall short. The trajectory of state pressure on encrypted services — the United Kingdom’s Investigatory Powers Act amendments of 2024, the European Union’s chat control proposals, the recurring Senate efforts to constrain end-to-end encryption — does not suggest that voluntary corporate cryptographic guarantees will survive a sufficiently severe political event. The plausible deniability of “I was just thinking out loud at my desk” is being replaced by the auditability of every thought-adjacent action that touches an LLM, and the auditability does not require a state to act; it requires only a state to choose to.
This is the surveillance capitalism that Shoshana Zuboff described, plus a state interface. It is the legibility architecture that James C. Scott warned about, made operational. The chilling effect on dissent, journalism, organizing, legal defense, and the simple human act of exploring an unconventional idea is the kind of harm that does not show up in any single incident but reshapes the distribution of what people permit themselves to think.
3. Supply Chain and Sovereignty
The cognitive infrastructure of the United States — and increasingly of every advanced economy — depends on advanced logic chips fabricated principally at TSMC’s facilities in Hsinchu and Tainan, with critical packaging and HBM memory supplied by SK Hynix and Samsung in South Korea. TSMC fabricates roughly 90% of the world’s leading- edge logic at 3 nanometer and below, and entered volume production at 2 nanometer in the fourth quarter of 2025 at 90,000-to-100,000 wafers per month against Samsung’s targeted 21,000 by end of 2026 and Intel 18A’s continued pre- production status. The Arizona Fab 21 expansion, raised from $65 billion to $165 billion in March 2025 with the planned fab count expanded from three to six, has Phase 1 in mass production at 4 nanometer; Phase 2, accelerated to 3 nanometer in the second half of 2027; Phase 3 targeting 2 nanometer late this decade. Today, no 2-nanometer chips are fabricated in the United States. Apple’s M-series, NVIDIA’s Hopper and Blackwell and forthcoming Rubin generations, AMD’s MI300/MI350/MI400 series, and most hyperscaler custom silicon are fabricated in Taiwan.
Bloomberg Economics’ January 2024 modeling, updated February 2026, estimates that a Chinese invasion of Taiwan that draws the United States into conflict would produce a $10 trillion loss in global GDP in the first year, roughly 10% of world output — exceeding the combined impact of COVID-19, the global financial crisis, and the war in Ukraine. A blockade short of invasion would still cost about 5% of world GDP. The top twenty TSMC customers carry roughly $7.4 trillion in market capitalization. Roughly 5.6% of global value-added comes from sectors that use chips as direct inputs.
These are not estimates of a tail risk that policy can defer. The Chinese Communist Party has moved its public timetable for cross-strait reunification readiness to 2027; the Pentagon’s Davidson window is well known. US export controls have tightened in October 2022, October 2023, October 2024 (multilateral
GAA), December 2024 (country-wide HBM and FDPR expansion), January 2025 (the global AI diffusion rule),
Congress.gov only to be partially reversed in late 2025 with H20 and H200 sales licensed to Chinese firms — a policy oscillation that signals neither a coherent strategy nor a settled commitment. China’s domestic capacity at SMIC remains stuck at 7 nanometer with reported delays to 5 nanometer;
Huawei Ascend 910C performs at roughly 60% of H100 capability. Council on Foreign Relations CSIS and CFR analysts project Huawei will not match H200 until the Ascend 960 in late 2027 or 2028.
The sovereignty problem is not principally about Taiwan, however. It is about the deeper asymmetry that a civilization whose intelligence infrastructure depends on a single fab on a contested island, three hyperscaler clouds, one accelerator vendor, and five frontier model labs has reduced itself to a level of strategic dependence with no historical precedent for a peer competitor in industrial capability. The defense industrial base has the same structural problem in slow motion: publicly traded firms whose fiduciary obligations push them toward the highest-margin niches at the expense of the surge capacity the Republic actually needs in a crisis. The cloud-AI complex is the same problem, occurring faster, in a sector where the capital intensity is even greater and the consolidation more complete. The case for government-owned, contractor-operated computing infrastructure, parallel to the GOCO model used in defense, is that the strategic logic that justifies sovereign munitions production capacity at Lake City and Radford applies a fortiori to sovereign compute capacity in a world where compute is the commanding height of every other industry.
The geographic concentration within the United States compounds the problem. Northern Virginia’s Loudoun County hosts more than 200 data centers; the Northern Virginia complex carries somewhere between 30% and 70% of global internet traffic depending on which methodology one trusts, and exceeds the combined capacity of Dublin, London, Frankfurt, Amsterdam, Singapore, and Sydney. The Pacific Northwest, Phoenix, and Atlanta clusters are similar. More than 70% of new hyperscale builds are concentrated in fewer than twelve metropolitan areas worldwide. The grid stress is now binding: Loudoun has issued de facto permitting halts through 2028; PJM capacity prices have spiked roughly tenfold; Microsoft signed a 20-year power purchase agreement with Constellation for the restart of Three Mile Island Unit 1 (now scheduled for 2027); Amazon and Talen restructured a 1.92-gigawatt deal at Susquehanna; Meta locked in 1.1 gigawatts at Constellation’s Clinton plant. The IEA’s Energy and AI report projects global data-center power demand to roughly double from 415 terawatt- hours in 2024 to between 945 and 1,100 terawatt-hours by 2026, with US data centers reaching 6.7% to 12% of national electricity by 2028. The civilizational cognitive substrate now shares its grid interconnection points with a small number of nuclear reactors and a small number of high-voltage substations. A single prolonged outage — natural, accidental, or adversarial — at any of a few dozen locations would degrade the cognitive function of the modern economy in ways the public has not yet imagined.
4. Epistemic and cognitive dependency: the atrophy of the process of knowing
The fourth threat vector is the one for which the empirical evidence is thinnest, the temporal horizon longest, and — for that reason — the policy response most likely to be deferred. It is also the one with the highest plausible upside and the deepest implications for the nature of citizenship.
Begin with what is well-established. Cognitive offloading — Risko and Gilbert’s 2016 Trends in Cognitive Sciences canonical definition, “the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand” — is a general adaptive strategy with measurable trade-offs. The trade-offs are not uniformly negative. The calculator literature is the strongest empirical counter- case to alarmism: Hembree and Dessart’s 1986 meta-analysis of 79 studies, NCTM Publications and Ellington’s 2003 follow-up of 54, found that calculator use, when integrated into instruction and assessment, improved computational and problem-solving skills with no skill erosion. The conditional matters. Calculators were curricularly integrated; their use was structured; the underlying mathematical operations remained legible and could be checked. AI use, at present, is none of these.
The GPS literature points the other way. Dahmani and Bohbot’s 2020 Scientific Reports paper, drawing on a three-year longitudinal cohort, found that habitual GPS use predicted steeper hippocampal-dependent spatial-memory decline. Javadi and colleagues’ 2017 Nature Communications fMRI work showed that hippocampal and prefrontal activity tracking route options collapsed under satnav use compared with self-navigation. Bohbot’s quotable line — “Getting lost is good!” — encodes a real finding: the engagement of effortful spatial reasoning maintains the hippocampal substrate, and its disuse correlates with measurable atrophy. Sparrow, Liu, and Wegner’s 2011 Science paper on the effect found that participants who expected information to be saved recalled less of the content and more of the location Semantic Scholar — the substrate of memory shifted from substance to address. Science Subsequent meta-analysis confirms the where-versus-what shift, with smaller pooled effects than the original.
The AI-specific literature is younger and noisier, but points consistently. Lee and colleagues’ Microsoft Research / Carnegie Mellon survey of 319 knowledge workers, presented at CHI 2025, found that higher confidence in generative AI predicted less critical-thinking enactment; workers shifted from problem-solving to “AI response integration” and from execution to “task stewardship.” Campus Technology The paper concluded that knowledge workers “often refrain from critical thinking when they lack the skills to inspect, improve, and guide AI-generated responses” — the precise dependency loop in which the use of the tool degrades the capacity to evaluate the tool. Kosmyna and colleagues’ 2025 MIT Media Lab EEG study (n=54, preprint, treat as suggestive) found that subjects writing essays with ChatGPT showed the lowest neural connectivity across alpha, theta, beta, and delta bands; could not quote their own essays back; progressively disengaged across sessions; and, when later asked to write unaided, exhibited persisting reduced connectivity the authors termed “cognitive debt.” Gerlich’s 2025 mixed-methods study of 666 participants in Societies found a significant negative correlation between frequent AI use and critical-thinking scores, mediated by cognitive offloading, with younger participants showing higher dependence and lower scores than older.
The student-use data is unambiguous about scale. The UK Higher Education Policy Institute reports student generative-AI use rose from 66% in 2024 to 92% in 2025, with assessment-specific use rising from 53% to 88%. Education Council’s 2024 global survey found 86% of higher-education students use AI in studies. Pew’s January 2025 survey of US teens 13–17 found 26% used ChatGPT for schoolwork in 2024, doubling from 13% in 2023. Walton Family Foundation’s 2024 work with Gallup reports 79% of Generation Z using generative AI, 47% weekly. The College Board’s 2025 figures put roughly half of US high-schoolers in the same category. Whatever the precise causal effects on critical- thinking development — and the empirical base is genuinely too thin to be confident — a generation is now passing through the formative years of intellectual development with an LLM as a constant cognitive prosthesis, and the prosthesis lives in someone else’s data center.
The framing that captures this most precisely comes from the educational-philosophical tradition. Call it philepistemicism — love of the process of knowing. The tradition runs from Socrates, who argued in the Phaedrus that writing would weaken memory because it externalized recollection; to Montaigne, who insisted that “to know by heart is not to know”; to Cardinal Newman’s Idea of a University and its insistence on “the formation of a habit of mind”; to Hannah Arendt’s distinction between thinking, knowing, and the dangerous comfort of received conclusions. The shared premise is that the act of working through a problem is not merely a means to the conclusion but constitutive of the kind of intellectual being that can hold the conclusion. To outsource the process is, over time, to lose the capacity. The student who has never solved a problem unaided does not merely lack the answer; she lacks the formed capacity to know whether the answer she has been given is right.
This is not a luddite argument against tools. It is a structural argument about the conditions under which tools enhance rather than replace cognition, which returns us to Illich. Convivial tools — the calculator integrated into a curriculum, the dictionary used after a guess — sit alongside human capacity and amplify it. Manipulative tools dis-able the user. The cloud-AI assistant, in its current configuration, is closer to the second than the first: opaque, unmodifiable, optimized for engagement and adoption rather than for the user’s intellectual development, and increasingly indispensable to the social and professional infrastructure in which the user must operate. An epistemically dependent populace is not a populace that can practice republican self-government in any robust sense. The political-economic implication is that a citizenry whose reasoning runs through a small number of remote systems is a citizenry whose reasoning is — at the limit — an artifact of those systems’ training data, alignment choices, and content policies.
5. Resilience and Continuity
Hurricane Harvey dropped 60 inches of rain on the Texas coast in late August 2017. Power was out across Houston for days; mobile networks were saturated where they functioned at all; the parts of the city that worked were the ones with local capacity — generators, fuel, cached supplies, paper records, radios. The systems that depended on continuous network connectivity to a remote service did not work. This is not a hypothetical. It is the recurring lesson of every regional disaster, every ransomware event, every fiber cut, every deliberate disruption.
The trajectory of cloud-AI computing is to make this failure mode the default. Microsoft 365 Copilot does not function offline. ChatGPT does not function offline. Workspace’s AI features do not function offline. The Apple Intelligence on-device model handles small tasks; anything substantive routes to the cloud; the ChatGPT fallback requires connectivity. A modern Tesla cannot receive the bulk of its features without periodic check-ins with the mothership; a John Deere tractor’s Service ADVISOR diagnostics, until the FTC’s January 2025 case and the June 2025 class-action settlement that forced ten years of farmer-accessible diagnostic tools, were dealer-only and remote-keyed. Sonos’s May 2024 forced app update bricked older speakers across the installed base; CEO Patrick Spence was terminated in January 2025 after a roughly $500 million market-cap loss, but the speakers remained bricked. Spotify deactivated 700,000 Car Thing devices on December 9, 2024. Amazon, on February 26, 2025, removed the download-and-USB-transfer feature for Kindle books
— eliminating the only sanctioned offline-backup path for purchased reading material. Each of these is a small case. The aggregate is a civilization in which the reasonable expectation that a paid-for device will continue to function in the absence of the manufacturer’s continued goodwill has been quietly retired.
Compound this with the geographic concentration of cloud infrastructure described above, and the resilience picture is unflattering. A coordinated cyberattack on three of the dozen largest data-center clusters; a Carrington-class geomagnetic storm; a Pacific cable cut at the right two locations; a Taiwan Strait kinetic event; a sufficiently sophisticated firmware-level attack on the Pluton or T2 trust roots — any of these would degrade civilizational cognitive function in ways that are not adequately modeled in any existing continuity-of-operations plan. The just-in-time supply chain analogy is exact: the optimization of average-case efficiency at the cost of worst-case resilience, executed at a scale where the worst case is no longer a tail event but a foreseeable scenario on a multi-decade horizon.
The personal computer, in its 1995–2015 form, was a resilience asset. It booted without a network. It computed on local data. It could be repaired with screwdrivers and replacement parts available at retail. It could be backed up to media the owner physically held. A homelab with a few rack units of network-attached storage, a modest GPU, an open-source LLM running on llama.cpp, a Meshtastic LoRa node for off-grid communication, an SDR for spectrum awareness, and a self-hosted media stack is a contemporary expression of the same principle: continuity does not depend on the willingness of a remote third party to keep your tools functioning. The number of households and firms maintaining such capability is small. The trajectory of the industry is to make it smaller. This is the disappearance of the off-grid-capable personal computer as a category, and it is happening at the same time as climate-driven extreme weather, peer-competitor cyber capability, and grid stress are making continuity more — not less — important.
IV. Why current policy responses do not reach the core The policy apparatus is not idle, and it would be unfair to suggest otherwise. The European Union’s Digital Markets Act, in force since November 2022 and applicable to gatekeepers since March 2024, has produced the first €500 million Apple fine for anti-steering breaches in April 2025, the first €200 million Meta fine for the consent-or-pay model, and 23 designated core platform services across seven gatekeepers. The US v. (Search) liability finding of August 2024 produced a remedies judgment in December 2025 — a six-year framework Winston & Strawn with shared search-index metadata for qualified competitors Winston & Strawn and a prohibition on exclusive defaults, though without divestiture of Chrome or Android. The US v. (Ad Tech) liability ruling of April 2025 has remedies pending. The DOJ-Apple case survived motion to dismiss in June 2025. The Epic v. Apple contempt ruling of April 2025 — Judge Gonzalez Rogers finding Apple executives “outright lied” — was largely affirmed by the Ninth Circuit in December 2025, with Fortnite returning to the US App Store in May 2025 and the EU Core Technology Fee restructured into a Core Technology Commission in June 2025. Right- to-repair statutes have passed in New York (2022), Minnesota (2023), California (2023), Oregon (2024, the first US ban on parts pairing), and Colorado (2024), with the EU Right to Repair Directive adopted in 2024 and member-state transposition due July 2026. The John Deere FTC case was filed in January 2025, and the parallel class action settled in June for $99 million and ten years of farmer-accessible diagnostic tools.
These are real wins. They are also structurally inadequate to the threat described in this paper. Three observations:
First, every active enforcement action targets a legacy market — search defaults, ad- tech, app-store commissions, repair access — while the new market of compute-and- frontier-model concentration deepens. There is no active US antitrust case against cloud infrastructure concentration, against NVIDIA’s accelerator-and-CUDA bundle, or against the vertical Microsoft–OpenAI, Amazon–Anthropic, and Google–DeepMind compute-and-model bundles, despite the FTC’s January 2024 6(b) orders Federal Trade Commission flagging precisely these concerns and the January 2025 staff report documenting them. The FTC v. Meta loss in November 2025 — the court rejected the narrow “personal social networking services” market definition given TikTok and YouTube competition — illustrates the broader problem: the doctrines are hard to apply to fast-moving technical markets, and judicial appetite for structural remedies is uneven. Antitrust as currently practiced is fighting the last war.
Second, the right-to-repair movement, for all its progress, does not yet reach software, model weights, or cloud-dependent functionality. Oregon’s parts- pairing ban applies to physical components. Colorado’s 2024 law is the broadest, but it applies to “digital electronic equipment” in the conventional sense. None of the state laws creates a right to run a different operating system on hardware one owns; none creates a right to local execution as an alternative to cloud routing; none addresses the Tivoization problem Stallman identified in 2006. Directive bans hardware or software obstacles to repair, which is a step, but its product scope (Annex II Ecodesign categories) excludes most of what matters for the AI stack.
Third, the DMA’s interoperability and sideloading provisions reach the symptoms of platform power without reaching the underlying compute concentration. A user is more able, in April 2026, to install a third-party app store on an iPhone in Berlin than at any time in the platform’s history. The user is no more able to run a frontier-class AI model on local hardware that she owns and controls. The DMA does not require Apple to permit alternative on-device language models; it does not require Microsoft to make Recall’s vector index portable to a competing search service; it does not require to expose Gemini Nano’s weights, even at modest capability tiers. The Digital Services Act, AI Act, and Data Act layer additional process requirements on top of the existing platforms but do not alter the underlying architecture in which the meaningful computation happens in the gatekeeper’s data center.
The pattern is the same across jurisdictions and across instruments. Existing policy treats the platforms as the unit of regulation. The threat described in this paper is that the unit of computation itself has migrated to a place where existing policy cannot easily reach. To reach it requires policy designed around the structure of the threat.
The recommendations that follow are intended to be structurally adequate to the threat, in the sense that — taken together — they would alter the trajectory described in Sections I through III rather than merely manage its symptoms. Each is grounded in an existing precedent and an existing policy lever; none is invented from whole cloth.
A federal local-computation guarantee
Every device sold in the United States for general-purpose computing — phones, tablets, laptops, desktops, in-vehicle compute, smart-home hubs — should be required by federal statute to provide a local-computation option for any consumer-facing AI feature whose primary function does not intrinsically require remote data. A search of the user’s own files should be performable locally. A summarization of the user’s own document should be performable locally. A drafting assistant should have a local mode. The standard need not require parity with frontier cloud capability; it must require that a degraded but functional local alternative exists and is selectable by the user. The standard should be technology-neutral and capability-tiered: small models that meet a defined utility threshold qualify; opaque cloud-only implementations do not. This is the AI analog of the analog-passthrough requirements that allowed PCs to function across operating system transitions, and of the FCC’s Part 68 rules that opened the customer- premises equipment market in the 1970s. It is the minimum floor of computational sovereignty for individual citizens.
A public option for compute
The defense industrial base relies on a small number of government-owned, contractor-operated facilities — Lake City, Radford, Holston — for surge capacity in munitions and propellants that the publicly traded prime contractors will not maintain because the fiduciary economics do not support it. The cognitive infrastructure of the Republic now has the same structural problem. The recommendation is direct: the federal government should fund and own a tier of compute infrastructure — perhaps fifteen to thirty gigawatts of capacity over the next decade — operated under contract by a mix of incumbent providers and new entrants, available on non-extractive terms to (i) federal and state agencies, (ii) qualifying public-interest researchers and institutions, (iii) public libraries and land-grant universities for citizen access, and (iv) strategic surge use during emergencies. The model is partly Tennessee Valley Authority, partly Federal Reserve discount window, partly land-grant extension service. The aim is not to displace private capacity but to provide a non-extractive baseline that disciplines the private market and guarantees civic access. Scale matters: ten gigawatts is large enough to matter to the market, small enough that the federal balance sheet can absorb it, and consonant with the order of magnitude of the Constellation–Microsoft and Talen–Amazon deals already underway.
Mandatory openness for capability-class model weights
For frontier large language models above a defined capability threshold — measured by training compute, by performance on standardized benchmarks, or by some combination — the federal government should require that model weights be deposited in a publicly accessible, sovereign repository within a defined window after deployment, on terms analogous to compulsory licensing in pharmaceuticals or to the Bayh-Dole Act’s march-in rights for federally funded research. The repository need not be free; it should be available on non-discriminatory terms to qualified domestic users including academic, public-sector, and small-business actors. The mechanism is necessary because closed frontier weights, held privately and revocably, are an unprecedented form of intellectual enclosure of capability that, in many cases, was developed using publicly funded research, public data, or content scraped without compensation. The European Union’s AI Act creates a partial template; this recommendation goes further by establishing portability and access rights as an affirmative obligation.
Tax treatment that ceases to subsidize the conversion of ownership into rent
The current US tax code treats software-as-a-service expenditures as fully deductible operating expenses while requiring the capitalization and amortization of perpetual software purchases over multi-year horizons, and — under Section 174’s 2022 amendments to require capitalization of research and development expenditures — has further tilted the treatment in ways that disadvantage in-house, locally-controlled software development. The cumulative effect of the tax code is to subsidize the conversion of capital purchase to recurring rent. The recommendation is to neutralize this distortion: equalize the deductibility of capital and subscription expenditures for software and computing services; restore Section 174 expensing for in- house development; consider a modest excise on subscription-only software-and- service revenue above a threshold to fund the public compute option above. The principle is that tax policy should not have a thumb on the scale against ownership.
Right-to-repair and right-to-modify
State right-to-repair statutes should be extended at the federal level and broadened in scope to cover (i) the right to install alternative operating systems on hardware one owns, with documented security implications disclosed but not prohibited; (ii) the right to local execution as an alternative to cloud routing for consumer-facing AI features; (iii) a federal ban on parts pairing analogous to Oregon’s SB 1596 but applicable nationally and to a broader product set; (iv) a prohibition on remote disablement of features that were operational at point of sale absent a documented safety or legal cause; and (v) mandatory escrow of activation servers, license servers, and cloud-dependent firmware so that the cessation of a manufacturer’s support does not brick existing devices. The Sonos, Spotify Car Thing, Stadia, and Amazon Kindle download cases provide ample precedent for the harm. The Stallman framework — without adopting movement orthodoxy — provides the rights vocabulary: the four freedoms become the floor, not the ceiling.
Federal procurement preferences
The federal government is the largest single buyer of computing services in the United States. Federal procurement should systematically prefer systems that are open- weight, locally executable, and modifiable by the agency operating them, with cloud-only and closed-weight systems requiring affirmative justification rather than the current default acceptance. The model is the existing Buy American provisions and the security-tiered acquisition framework under FedRAMP. The aim is not to ban cloud or closed-weight systems but to ensure that the federal apparatus does not, by default, deepen the dependencies described in this paper. Adjacent: the Department of Defense and Intelligence Community should treat sovereign compute as a national-security category in its own right, with explicit budget lines and continuity-of-operations requirements.
Strategic compute reserves
Parallel to the Strategic Petroleum Reserve and the National Defense Stockpile, the United States should establish a Strategic Compute Reserve: a defined inventory of pre-positioned, sovereign-controlled GPU and accelerator capacity, geographically distributed, with model-weight escrow and operational continuity for defined civic, research, and emergency uses. Initial scale: 500,000 to 1,000,000 high-end accelerator-equivalents, distributed across at least eight geographic regions, with full redundancy. The investment is on the order of $30–50 billion, modest relative to the $600 billion annual private capex flow and trivial relative to the $10 trillion exposure modeled for a Taiwan Strait event.
Antitrust enforcement targeted at compute concentration
The active enforcement docket should be expanded to cover (i) the Microsoft–OpenAI, Amazon–Anthropic, and Google–DeepMind vertical bundles, with structural remedies on the table; (ii) NVIDIA’s CUDA-and-accelerator bundle, with mandatory interoperability standards for compiled CUDA workloads on competing hardware; (iii) cloud infrastructure concentration, with consideration of the regulated-utility model that governed AT&T from 1913 to 1984 and the divestiture model that ended that arrangement; (iv) the geographic concentration of data centers, treated as a separate competition harm given grid and continuity implications. The doctrinal basis exists; the appetite must be created politically.
Education policy that preserves unassisted reasoning
Schools and universities should be required to maintain and assess substantial unassisted reasoning capability as a curricular and assessment objective, with specific protected categories of work in which AI assistance is prohibited and with assessment formats — oral examinations, in-class written work under controlled conditions, supervised practica — that can verify the formation of the underlying capacity. The calculator literature provides the template: tools are integrated where they enhance learning, restricted where they substitute for the formation of the capacity the curriculum is meant to develop. The aim is not luddism. It is the preservation of the kind of intellectual being who can use AI as a tool rather than be used by it. Philepistemicism — the love of the process of knowing — is a curricular commitment, not a slogan.
Funding for open-source model development as public good The federal government should fund — through NSF, DARPA, the Department of Energy’s national laboratories, and a new dedicated entity if necessary — the development of frontier-capable open-weight models on the order of $10–20 billion per year, sustained for a decade, with the explicit objective of maintaining a publicly available capability frontier within roughly one generation of the closed-weight frontier. The investment is large in nominal terms and small relative to the comparable private investment. The justification is the same as the justification for federally funded basic research: the social return on knowledge that can be widely used exceeds the private return on knowledge that can be privately captured. The Chinese open-weight ecosystem — DeepSeek, Qwen, Kimi, GLM — already provides a frontier counter- current; the strategic case for an American (and allied European, Japanese, Korean, Taiwanese, Indian) public counterpart is overwhelming.
VI. Conclusion: the stakes at civilizational scale The argument of this paper is that the relocation of computation from owned, modifiable, locally controlled hardware to a small set of remote, opaque, irrevocable services controlled by five firms and dependent on one fab is a structural transformation of the same order as the enclosure of the English commons, the conversion of artisanal labor into wage labor, and the centralization of broadcast media in the twentieth century. Each of those transformations was, in its time, defended on the grounds of efficiency and consumer preference. Each, in retrospect, is recognized as having reorganized the distribution of power between individuals and institutions in ways that took generations to partly reverse and that, in important respects, were never reversed.
The cloud-AI transformation has features that distinguish it from all three precedents and make it more dangerous. It is faster: the relevant transition is occurring within a decade, not a century. It is more concentrated: five firms, not five hundred. It penetrates deeper: into cognition, not merely production or communication. It is harder to exit: the alternative to a Word document was a typewriter; the alternative to a Cloud LLM is, increasingly, nothing the average citizen can readily reach. And it is reinforced by a financial architecture — $600 billion in annual capex, $1.5 trillion in projected debt issuance, $7 trillion in projected investment by 2030 — that creates path dependencies measured in decades.
The window in which policy can shape this trajectory is narrow. Within roughly five years, the consumer-runnable open-weight tier may either close the gap with the closed frontier (the BitNet, distillation, and Mixture-of-Experts trajectories suggest this is technically possible) or fall further behind (the capex curve suggests this is the default outcome). Within ten years, the generation now passing through formative education will enter professional life having developed its intellectual habits in a world where a cloud LLM is a constant cognitive prosthesis, and the question of whether they retain the capacity for unassisted reasoning will be largely answered. Within twenty years — by 2046 — the geographic and infrastructural facts on the ground will be either a recognizably plural ecosystem of public, private, sovereign, and personal compute, or a hyperscaler oligopoly with a thin civic veneer. The choices that determine which of these futures arrives are being made now, in board rooms, in capitols, in the design decisions of operating-system vendors, and in the willingness or unwillingness of citizens to demand that the devices they buy continue to belong to them.
There is a final consideration that the political economy of the moment tends to obscure. The personal computer revolution of 1977–1995 was not principally a market phenomenon. It was a civic phenomenon: the conviction, held by a generation of engineers, hobbyists, educators, and writers, that the most powerful intellectual tool ever devised should be available for ownership, modification, and use by ordinary people, and that the alternative — computing as a service rendered by a priestly caste in glass houses — was unworthy of a free people. That conviction built the industry that the firms now consolidating it inherited. It has not yet been refuted; it has merely been quietly retired in favor of a more profitable arrangement. The policy task of the next two decades is to insist that the conviction was correct, that its implications still bind, and that the architecture of computation in 2046 should be one a citizen of 1995 — or of 1789 — would recognize as compatible with the kind of self-governing republic on which everything else depends.
The personal computer was always, at its best, a possession, in Proudhon’s sense: a thing held in actual use by the person standing in front of it. It can be again. But not by accident, and not by markets alone, and not for much longer if the trajectory described in this paper continues unchecked. The recommendations above are a beginning. The stakes are the kind that are usually recognized only after the window for action has closed. There is still time, but not a great deal of it.