The Sovereignty Stack

For about thirty years after the Cold War, the question “who is in charge of the world?” had a default answer. It no longer does, and nothing single has replaced it. What we have instead is several kinds of power stacked on top of one another: armies and borders at the bottom, trade and money above them, digital networks above those, and at the top, individuals who can now do with AI what used to take an institution.

This essay is a way of reading that stack. Its thesis is that power now goes to whoever can combine energy, computing power, people and legitimacy across those layers faster than rivals can capture or constrain them. The outcome isn’t settled. Two forces are racing each other. One is diffusion: capability spreading outward to individuals and networks. The other is capture: states and concentrated capital taking hold of the infrastructure that capability depends on. A third possibility, which I think is the likeliest for the next decade, is a hybrid I call managed diffusion: the tools spread, and the infrastructure stays captured. Most of the variables that decide which force wins can be named and measured, and they are in play now.

One definition carries the whole essay. Sovereignty here means practical freedom to decide under outside pressure. Every layer below is judged against that standard, not against flags or legal status.

A caution before the map. The four layers are not the same kind of thing. The first is a form of government, the second a way of coordinating, the third a mix of infrastructure and community, the fourth a class of actor. A tidier scheme would sort everything by type of control: coercive, economic, informational, cognitive. I have kept the untidy version because it is easier to hold in the head, and I flag the places where the mixing causes trouble.

The post-Cold War order was a set of arrangements: the financial institutions designed at Bretton Woods in 1944, American security guarantees, and a single integrated global economy. It did not die on one day. It failed in stages over two decades, and each shock exposed a deeper layer of the same fracture.

Four of the shocks are the ones the West remembers. They looked different from Beijing, Moscow or Lagos.

September 11, 2001 exposed a security model that assumed threats came from states. It had no answer to violence organized as a network and driven by ideology. In much of the developing world the event registered less as a shock to the system than as confirmation that the “rules-based order” applied selectively, bending around the interests of the states that wrote the rules.

The financial crisis of 2008 exposed the plumbing. Interdependence had been sold as resilience through diversification; it turned out to be a route for contagion. For much of the world this was the moment American economic stewardship lost its presumption of competence.

COVID-19 in 2020 exposed supply chains with no spare capacity and institutions that had already lost public trust. At home, it sped up the collapse of shared ground on basic facts: people could not agree on what was true, let alone on what to do. The pandemic didn’t create the trust deficit. It made it impossible to ignore.

Ukraine in 2022 exposed the military and alliance architecture. The belief, revived after the Cold War, that economic interdependence had made a war of conquest in Europe unthinkable proved false, as it had in 1914. Security guarantees turned out to be conditional. For non-aligned states the lesson was plain: great powers will set the rules aside the moment the rules constrain them.

Two more fractures belong in the sequence. Crimea in 2014 was the dress rehearsal: the first forcible annexation of territory in Europe since 1945, met with a response calibrated to avoid disruption rather than to enforce the norm. Revisionist powers could read the message. Brexit and the 2016 US election were a collapse of legitimacy inside the order’s own core. The populations underwriting the system voted to leave it, or at least to renegotiate the terms. That fracture came from within.

None of these was a new crisis. Each was the same crisis showing itself at a deeper level: an order whose guarantees turned out to be conditional. By the time the last crack came, the system had already lost its vitality. February 24, 2022 was not the death of the old order. It was the funeral everyone finally attended.

The claim: the old order is gone. The question is what replaces it, and for whom.

Nothing single will replace it. Historical transitions rarely produce one winner. They produce overlapping arrangements that coexist, compete and constrain one another. The system of sovereign territorial states conventionally dated to the Peace of Westphalia in 1648 did not replace empires overnight; the two lived side by side for centuries. Westphalia was also the settlement of a broken epistemic order, a point Section III returns to. So the useful question isn’t which model wins. It is how several models interact, and who gets to choose which one they live under.

I find it clearest to picture the result as a stack: several kinds of authority operating at once, at different scales.

Dependence runs up the stack. Constraint runs down it, with a delay.

The network analogy, and where it breaks

Engineers describe computer networks in layers, from physical cables at the bottom to applications at the top. Each layer has its own rules, lower layers constrain upper ones, and you understand the system by studying how the layers meet. That much carries over.

One thing does not, and it matters. In a computer network the layers cooperate. In the Sovereignty Stack they are adversaries. They share no common rulebook. Each tries to absorb, constrain or override the others.

The friction this creates I call jurisdictional drag: the cost the lower layers impose on anything done above them. Suppose an individual uses a decentralized network to move money between trade blocs. The state underneath does not politely pass the transaction along. It sends enforcement to a physical address. Everything done in the upper layers pays a toll to the layers below, and the toll can rise to confiscation without notice.

Dependence runs upward. Individuals depend on networks, networks on economic infrastructure, and all of them on the tolerance of whoever holds physical force. Upper layers can tunnel through lower ones: a VPN, software that disguises where a connection is coming from, gets around censorship; satellite internet gets around ground-based control; digital tokens pegged to the dollar (stablecoins, discussed below) get around the banking system. But a workaround isn’t sovereignty. An AI-equipped individual with no passport is just a very online refugee.

Feedback runs downward, with a delay. Lower layers are not immune to what happens above them. Financial markets discipline governments. Platforms tilt elections. Capital flight narrows a state’s options. The upper layers can’t command the lower ones, but they can make the use of force ruinously expensive. Russia’s 2022 invasion showed this in real time. Other states, working through the upper layers, imposed sanctions, exclusion from finance, denial of semiconductors and the withdrawal of software. Together these imposed costs that Moscow did not anticipate and has been absorbing ever since at a heavy price to its economy and armed forces.

So the dependence runs in both directions, unequally. Material power flows up, constraint flows down, and the delay between an action and its consequences is itself something strategists can exploit. (Section X adds an important qualification: the delay mostly disappears when a state acts against its own companies and citizens.)

Layer 1 · Nation-states

Layer 1 endures because physical force isn’t a metaphor. Territory has to be administered and infrastructure maintained. Violence has to be monopolized, or it spreads.

The veto, qualified. Layer 1 can override every layer above it, when it can call on the rest of the stack. That condition matters, and the invasion of Ukraine, the event this framework started from, shows why. In 2022 Russia had enormous physical force: conscripts, armor, a nuclear arsenal, vast territory to fall back on. It bled anyway, for want of semiconductors, precision munitions, financial access, logistics software and good intelligence. Force without competence in the upper layers produced a war of attrition instead of the swift override that history’s great mobilizations suggest.

Put carefully: an aircraft carrier and its escorts do not end your independence in thirty minutes. A carrier group whose supply chain still works, whose missiles are still being built, whose software is still being updated and whose financing is still solvent does. The veto is real, but it is latent. No power in 2026 can exercise it entirely from its own resources. Even the United States relies on Taiwanese chips, allied bases and the global financial system to project force at scale.

Plate 01Force / dependence

Read the diagram in text

A state holds a physical veto, but exercising it at scale requires inputs from the rest of the stack. Administrative orders can interrupt access to those inputs.

History supports both halves of that. The Qin state unified China in 221 BCE by force of arms, but only after mastering the agriculture, administration and logistics that fed its armies. (The dynasty fell within fifteen years. The unified state outlasted it.) Rome’s legions rested on economic integration and a culture that provincials wanted to join. The lesson of those cases isn’t that force is supreme. It is that force succeeds when it is integrated with everything else, and fails, often spectacularly, when it is not.

The bad-news problem. States also suffer a decay in what they know. The higher you go in any centralized hierarchy, the worse the information gets, because the middle ranks filter out bad news to protect themselves. Every hierarchy has this problem. Centralized systems have it worst, because they lack the correctives that open societies rely on: a free press, independent audit, competitive markets. The state has the heavy weapons, but its decision cycle is increasingly choked by its own distorted readings. An individual at the top of the stack has almost no force but may see the ground truth far more accurately.

Part of the race between diffusion and capture can be restated this way: it is a contest between the state’s power to coerce and the individual’s ability to see clearly. A state whose own reporting is corrupted can crush you and still not know where you are, what you are doing, or whether its own systems are telling it the truth. Surveillance, taken up below, is its attempt to close that gap.

The state also hires the upper layers. The tidy picture has Layer 1 threatened from above and striking back. But states use upper-layer actors too. Ransomware gangs that enjoy state shelter, digital privateers and “patriotic” hackers are force by proxy. The precedent is the privateer, licensed by a letter of marque to raid an enemy’s commerce, with one difference: the old license was a public legal document, and today’s arrangement is deniable. North Korea funds weapons programs through cryptocurrency theft. Russia tolerates cybercriminals so long as they aim abroad. Iran projects force through proxy cyber units.

So Layer 4 is not only individuals seeking autonomy. It is also digital warlords working under state cover. The relationship between the bottom and top of the stack is employer and mercenary as often as sovereign and subject.

Geography now extends upward. Satellite constellations, space-based surveillance and the emerging competition over the space between Earth and the Moon push Layer 1 into a domain that borders don’t cover. Whoever controls low orbit holds the communications that can’t be cut on the ground and much of the surveillance for every layer above, and potentially the delivery of weapons. This is already operational: Starlink’s role in Ukraine, China’s anti-satellite tests and the spread of commercial imaging satellites are facts of 2026.

Layer 2 · Economic blocs

The picture of “the West against China, Russia and the Global South” is already out of date. The blocs of 2026 are messier, more fluid and more transactional than Cold War analogies suggest.

The American energy lever. The United States supplied about 57 percent of the European Union’s liquefied natural gas in 2025, nearly four times the volume it shipped in 2021. That isn’t alliance solidarity. It is dependence, and Washington uses it. Europe has swapped reliance on Russian energy for reliance on American energy while describing the change as strategic autonomy. The vulnerability is of the same kind, though not the same degree: a tanker can be redirected in a way a pipeline cannot.

India’s multi-alignment. India takes part in BRICS, the grouping founded by Brazil, Russia, India, China and South Africa, buys American gas, keeps defense ties with Russia, competes with China for manufacturing and runs its own space program. It doesn’t fit in a bloc; it trades between them. This is more than opportunism. It is a doctrine: keep enough options open that no bloc can compel alignment, and build domestic capability at every layer meanwhile. This is what sovereignty at Layer 2 looks like in practice.

Gulf capital on every side. Saudi Arabia, the United Arab Emirates and Qatar invest their sovereign wealth in Chinese AI, American technology, European energy and African mining at once. They are not choosing a side. They are becoming market makers, the dealers who profit from standing between buyers and sellers, and what they earn on is the friction between blocs.

“The Global South” is not one actor. Nigeria’s young population and fast digital adoption create different dynamics from Brazil’s dependence on commodities or Indonesia’s manufacturing ambitions. The African Union is developing strategic agency of its own, shaping the terms on which both Chinese and Western capital arrive. Treating all of this as one bloc hides the differences and repeats the old habit of defining everyone else in relation to the West.

Institutional thickness. Real cohesion takes more than adding up economies. NATO has lasted because of more than seventy-five years of joint exercises, shared classification systems, compatible equipment and officers who have known each other for whole careers. That depth is a kind of capital that economic analysis misses. It also explains why BRICS, for all its combined output, has far less ability to act together under pressure.

Layer 2.5: stablecoins and parallel finance. The retail era of cryptocurrency produced speculation, fraud and concentrated control behind a show of decentralization. The institutional phase that followed built something structurally different: Bitcoin funds traded on ordinary stock exchanges and holding more than $100 billion, government bonds issued as digital tokens, and stablecoins, digital tokens pegged to the dollar, moving volumes that rival mid-sized payment networks. Stablecoins may be the most successful upper-layer technology operating at Layer 2. They are a parallel dollar payment system that extends American monetary reach while also allowing transfers outside the banks. That is neither simple disruption nor simple capture. It is a zone, which I label 2.5, where traditional and crypto-native finance coexist, compete and increasingly connect. (A half-numbered layer is a warning that the taxonomy is under strain. I’ve kept the label because the thing it names is real.)

Layer 3 · Digital networks

The best-known version of this layer is the network state, the technologist Balaji Srinivasan’s term for an online community that organizes first and seeks territory and recognition later. Others include decentralized autonomous organizations (groups whose rules are enforced by software on a blockchain, a shared public record of transactions) and looser transnational communities. In theory they can operate across borders with governance that travels and identity that goes with the member.

Layer 3 is the most promising layer in theory and the most fragile in practice, because its independence rests on infrastructure it doesn’t control.

The transparency trap. If a network runs on a public blockchain, it is an open book to state surveillance. Real independence requires privacy-preserving technology such as zero-knowledge proofs, which let someone prove a fact (that they are over eighteen, that a payment is valid) without revealing the underlying data. The mathematics exists and is in commercial use; the question is adoption, and tolerance. Tornado Cash, the first privacy tool of this kind to reach wide use, became the first target of sanctions (Section X). Without such tools, Layer 3 “sovereignty” is a fishbowl: visible, catalogued and revocable. The same choice faces central banks designing their own digital currencies. Whether those are built with real privacy or with full visibility of every transaction is among the most consequential either-or decisions now being made at Layer 1.

Who actually holds this layer. The network state and the DAO are the versions of Layer 3 that want sovereignty. They are not the versions that have power. The digital networks that shape what billions of people see and say are corporate platforms: Meta’s apps, X, Telegram, YouTube, and the Apple and Google app stores that decide what software reaches a phone at all. They govern without territory and write and enforce their own rules, which is this layer’s definition. They are also owned by Layer 2 capital and answer to Layer 1 courts. That makes them the captured half of the layer, and the rails that Hypothesis C says stay captured.

So Layer 3 is two things at once. Its platform half is the most consolidated part of the stack. Its polity half, the protocols, charter zones and commons that try to govern themselves, is the fragile part this section describes. The fragility claim applies to the second half only. Inferred.

Three cases from the polity half.

Ethereum: code was law until people overrode it. In 2016 an attacker drained about 3.6 million ether from The DAO, an investment fund run entirely by software on Ethereum. The code permitted the withdrawal. On July 20 the community hard-forked the chain to reverse it, after an on-chain vote in which about 5.5 percent of the supply took part. The minority that refused kept running the original chain, now Ethereum Classic. The most successful polity in this layer settled its founding crisis by social consensus overriding the code, which is the argument Section VI makes against “code is law”. Known.

Próspera: a charter zone that took a state to arbitration. Próspera is a privately governed zone on the Honduran island of Roatán, with its own civil law and arbitration as the default for disputes, created under the country’s ZEDE law. Honduras’s Congress repealed that law in 2022, and its Supreme Court declared the framework unconstitutional in 2024. The company answered with an investor-state claim under the CAFTA-DR trade agreement, first estimating damages at up to $10.7 billion. Honduras left the World Bank’s arbitration body, ICSID, in 2024 and rejoined in August 2026 under a new government; the United States filed its own submission in the case in October 2026. This is the stack in miniature: a Layer 3 project, using a Layer 2 instrument, against a Layer 1 state, with a second Layer 1 state taking part. Known; outcome open.

Telegram: a network with one founder. Telegram operates across borders and is run in practice by its founder, Pavel Durov. France arrested him in August 2024 and placed him under formal investigation over crimes committed on the platform; the case is still open. In July 2026 Russia charged him in absentia with aiding terrorism and moved to place him on an international wanted list. A network with no territory still has a founder with a body. Telegram belongs here as a platform and returns in Layer 4 as a human anchor. Known.

Layer 4 · Augmented individuals

This is where the democratization thesis lives or dies. The claim: individuals with AI tools that used to require an institution can reach real independence of decision by moving among the layers beneath them.

The counterclaim: Layer 4 sovereignty is a luxury good, available to those with the capital, credentials and geographic luck to exploit it. For most of humanity Layer 1 decides how life goes, and Layer 4 is theater.

Both claims have evidence today. Which one holds depends on the variables in the next three sections.

Who is actually in this layer. Name the people who best fit the description and a pattern appears. They sort into four kinds.

• The solo founder. Pieter Levels runs a set of web businesses, among them Nomad List, Remote OK and Photo AI, with no employees, and reports multi-million-dollar annual revenue from them while choosing where he lives. This is the democratization thesis at its cleanest. It also started from a European passport, fluent English and a laptop: Layer 1 luck.
• The orchestrator. Elon Musk spans every layer at once: a satellite constellation doing work once reserved for states, companies with bloc-level economic weight, a platform, and a period inside the US government. In September 2022 he declined Ukraine’s emergency request to extend Starlink coverage to Sevastopol for a drone attack on the Russian fleet. One person set the edge of a military operation. He is the strongest case for the orchestration thesis in Section IX and the strongest case for the luxury-good objection, because none of it was available without enormous capital.
• The anchored. Roman Storm, Pavel Durov, Ross Ulbricht and Edward Snowden each built or exposed something that operated above Layer 1, and each was reached through the body. Snowden’s passport was cancelled while he was in transit in 2013, leaving him in a Moscow airport transit zone for 39 days: the “very online refugee” above, literally. Durov holds citizenship in Russia, France, the UAE and Saint Kitts and Nevis, and was still arrested by one of those states and charged by another; a second passport is a hedge, not a shield. Ulbricht, serving a life sentence, was freed by presidential pardon in January 2025, which fits the rule in Section X that diffusion wins when part of a lower layer sponsors it.
• The unfound. Satoshi Nakamoto published the design for Bitcoin in 2008, stopped corresponding around 2011, and has never been identified.

What the roster shows. The visible members of Layer 4 are either rich or reached. That is not bad luck in the sampling. It is the legibility penalty (Section X) acting as a filter: a Layer 4 actor I can name is one a state can find, so any roster of named individuals over-counts capture and under-counts quiet success. The most sovereign actor in the layer is the one nobody can name. Inferred; speculative as a general rule.

Two consequences follow. The luxury-good objection is stronger than the February text allowed, because even the success stories began with capital, credentials or a protecting state. And the evidence for quiet success has to come from aggregates, such as the dashboard’s hardware prices and share of AI run locally, not from biographies. Inferred.

Plate 02Layers / actors

Read the diagram in text

Layer 4: Levels (founder), Musk (orchestrator), Satoshi Nakamoto (never identified), and Storm, Durov, Ulbricht and Snowden (reached by a state). Layer 3: platforms such as Meta, X, Telegram, YouTube and the app stores (the captured half), and polities such as Bitcoin, Ethereum, Próspera, Wikipedia and Linux (the fragile half). Layer 2: NATO, BRICS, the EU single market, India, the Gulf funds, SWIFT against CIPS, and stablecoins. Layer 1: the United States, China, Russia and Japan, with Somalia and Libya as failures and the Lazarus Group as a state proxy. Layer 1 reaches Layer 4 through the human anchor.

What sorts actors into winners and losers across the stack? Three material variables: energy, AI capability and demographics. One philosophical variable: legitimacy, which I treat in Section VI under the name of the Logos problem. And two forces that cut across every layer: the information environment and climate. The material variables are constraints. Legitimacy decides whether upper-layer sovereignty is coherent or just organized noise. The cross-cutting forces reshape the ground the others stand on. AI capability is the subject of Section IV; the rest follow here.

Energy: the Jevons constraint

In 1865 the economist William Stanley Jevons noticed that more efficient steam engines did not reduce Britain’s coal consumption. They increased it, because cheaper power found new uses. The pattern has held through coal, oil, electricity and computing: demand grows to consume the supply and then invents new kinds of demand. On that record I don’t expect energy to become abundant in the sense that matters. Demand has always caught up with supply, whatever the source.

But Jevons doesn’t rule out democratization at the margin. The question that matters isn’t whether energy becomes abundant. It is whether the cost of useful AI, in energy and in hardware, falls fast enough that access spreads widely before centralized powers can fence it off. The two costs fall on different people. Running a model at home takes little electricity; what limits the individual is the price of chips. Energy gates the frontier, where training the largest models takes power on the scale of a city.

The smartphone precedent. The world spends more on computing every year, exactly as Jevons would predict. Yet the cost of useful computation for one person fell by orders of magnitude. Scarcity in aggregate coexisted with abundance at the margin. The hopeful case for AI depends on the same thing happening again. The precedent cuts both ways, though. A smartphone is a cheap device running on app stores, identity systems and payment networks its owner doesn’t control, which is managed diffusion in miniature.

The risk. If the gap stays wide between frontier AI, meaning the most capable models at any given moment (energy-hungry, concentrated, available to states and the largest firms), and accessible AI (efficient, distributed, available to individuals), the result is a two-tier economy of intelligence that reproduces existing hierarchies.

China’s solar position: dominant and brittle

China’s solar build-out mirrors what the United States did with oil in the twentieth century. It accounts for roughly 85 to 97 percent of world output at each stage of solar manufacturing, from polysilicon to finished panels. It dominates the processing of lithium, cobalt and rare earths. And its Belt and Road program supplies the distribution network.

The paradox. Sunlight is the most widely distributed energy source there is. The supply chain for capturing it is the most concentrated on earth. Oil had OPEC, a dozen or so producers with competing interests. Solar manufacturing is, in effect, one country. China has achieved in manufacturing what OPEC never managed in extraction. The technology that should do most to spread energy independence sits behind the most centralized supply chain in history. The oil analogy has a limit: oil is burned and must be bought again, while a panel keeps producing for decades once installed. China’s leverage is over new capacity, replacement and price, not over the power already flowing.

The brittleness. The dominance is real and it is fragile. It rests on a shrinking workforce (China’s working-age population peaked around 2011), an insolvent property sector that limits what the state can spend, and the bad-news problem of command economies, where managers optimize for political survival over operational truth. China’s four largest panel makers lost about $1.5 billion between them in the first half of 2025, the price of overcapacity. Dominance and brittleness coexist, and the open question is which shows first at scale. There is a wild card in this. If overcapacity forces China to sell panels abroad below cost, it may speed up energy independence in exactly the regions where it hoped to build dependence.

Four energy futures

1. Chinese solar hegemony, the current default. Solar becomes the new oil and Chinese manufacturing the new OPEC, only more concentrated. Developing countries align with Beijing as their infrastructure supplier, and blocs harden around energy supply chains.
2. Competitive diversification, the active counterstrategy. The United States, India and the European Union build their own panel and battery industries. Viable, but slow: a decade or more to reach competitive scale against an entrenched cost advantage, and vulnerable to political reversal. Much of the American program was cut back in 2025.
3. Technological disruption, the wild card. A source China doesn’t control, such as small modular nuclear reactors or fusion, scales fast enough to strand China’s solar position.
4. Hydrocarbon extension, the American play. US gas and oil prolong the fossil era, buying time for the second or third future while building a dependence of their own. Not a long-term answer, but a powerful medium-term lever, as Europe’s position shows.

Assessment. The first is the default absent a deliberate counterstrategy. The fourth is what is actually happening. The second and third need sustained investment over periods that may outlast political attention. China’s brittleness is the key uncertainty. If it shows before diversification matures, the default breaks open.

Demographics: the population constraint

A country’s working-age population limits its position in the stack as surely as its energy supply does. Demographics also move on a timescale that policy barely touches. The workers of 2045 have already been born. That makes this the most predictable of the three material variables and the hardest to change.

The world is splitting into two demographic realities.

Aging and shrinking. Japan is furthest along, with about 37 percent of its people projected to be over 65 by 2050. South Korea, China and much of Europe follow. Decline of this kind squeezes military manpower, economic dynamism and the capacity for innovation all at once. China’s case matters most, because it is the first major power to grow old before it has grown rich.

Young and growing. Sub-Saharan Africa is projected to add several hundred million working-age people by 2040, and its total population to grow by about two-thirds by 2050, to roughly 2.1 billion. That means heavy pressure on governments, infrastructure and migration routes. It also means the chance to skip legacy systems altogether, as the region did with mobile phones, and a demographic dividend that will produce either development or instability depending on governance, schooling and jobs.

How this reshapes the stack:

• At Layer 1, military manpower tracks the working-age population. Japan and South Korea already struggle to fill their ranks, and China’s military modernization is racing its own demographic clock. Countries with many young people face the opposite problem: large numbers of unemployed young men are the classic raw material of instability.
• At Layer 2, pensions and the ratio of dependents to workers eat into what states can spend. A 2019 study by the Chinese Academy of Social Sciences projected that China’s main state pension fund would run dry by 2035. Meanwhile countries with growing workforces become attractive for manufacturing, if their governance and infrastructure can absorb it.
• At Layers 3 and 4, innovation and adoption skew young. Aging societies have capital and infrastructure but fewer of the risk-tolerant, technically adaptable people who drive adoption. Young societies may lack infrastructure but have the appetite for experiment that the upper layers need.

Assessment. Demographics shift every probability in the scenarios below. How far depends on migration policy, birth rates and the degree to which AI and automation can stand in for missing workers, which is one of the most consequential unknowns in this framework.

The information environment: narrative as terrain

Who controls what populations believe is happening now matters as much as who controls energy or computing power. The information environment is not a fifth layer. It is the medium through which every layer projects power, and a weapon that states use to shape conditions for all the others.

Earlier drafts of this framework underweighted it, and the omission is telling. It repeated the blind spot of the post-Cold War order itself: the assumption that more connectivity would bring the world toward openness and shared truth. The opposite happened.

The precedent: epistemic oligarchies

Shared truth at scale has a history, and it is not a history of openness. For most of the last millennium, what large populations accepted as authoritative was certified by a few institutions that controlled a scarce input. I call these epistemic oligarchies. The medieval Church controlled Latin literacy, the copying of texts and the right to interpret scripture. Twentieth-century broadcasting rested on licensed spectrum; in the United States three networks dominated it, operating under federal licenses and, from 1949, the Fairness Doctrine. Universities control credentials and peer review. Each was a captured information rail in this essay’s sense. Anyone could speak. Few could certify.

That changes how to read the present. The collapse of shared ground on basic facts, described in Section I, is less the decay of a natural baseline than the breakup of the rail that manufactured it. Mass consensus was an artifact of oligarchy. Its loss is a real loss, but repairing it does not mean returning to a state of nature. It means deciding who, if anyone, certifies next.

The three oligarchies failed in three different ways, and the difference tells you what to watch for in the next one.

• Rupture. The Church’s monopoly broke when a distribution technology, printing, met political sponsors: German princes with their own reasons to shelter a heretic.
• Fragmentation. Broadcast authority dissolved through deregulation and channel proliferation. The FCC repealed the Fairness Doctrine in 1987, cable multiplied the channels, and the internet multiplied them past counting. The algorithmic curation described below is what grew in the space.
• Drain. Academia’s forms persist (degrees, journals, tenure) while public trust drains out of them. It is the same failure this essay assigns to economic blocs.

Plate 03Information / authority

Read the diagram in text

Church: Latin literacy and interpretation; failure by rupture. Broadcasting: licensed spectrum and networks; failure by fragmentation. Academia: credentials and peer review; public trust drains. Frontier AI: compute, competing labs and default answers; open models provide an alternative.

The rupture case carries the warning. Breaking the Church’s monopoly did not hand interpretive authority to individuals. It fragmented that authority among princes, and after more than a century of confessional war it was settled by territorializing belief: first at Augsburg in 1555, then at Westphalia in 1648, on the principle later summarized as cuius regio, eius religio, whoever rules the land sets its religion. The Westphalian order this essay uses as its anchor for overlapping sovereignties was the settlement of the last epistemic oligarchy’s breakup. The first of the four regimes below is that settlement again, with packet inspection.

Four regimes

Four information regimes now coexist.

• State enclosures. China’s Great Firewall, Russia’s “sovereign internet” and similar national boundaries produce populations whose picture of events is managed from above. This goes beyond censorship. It is the construction of an alternative reality.
• Reflexive control. This is the Soviet and Russian doctrine of steering an opponent’s decisions by shaping the information he acts on. It is used against government legitimacy (election interference), against bloc cohesion (energy disinformation) and against trust in digital spaces (flooding them with noise).
• Algorithmic curation. In open societies, feeds tuned for engagement fragment the public without any censor. People in the same country inhabit different realities depending on what the algorithm serves them. No conspiracy is required. It is what an economy built on attention produces.
• AI-generated environments. The newest regime and the least understood. As tools that generate convincing text, images, audio and video spread, the cost of manufacturing a false story approaches zero while the cost of checking one stays high. (So far the same tools have cheapened forgery faster than detection. That could change.) In the information domain that asymmetry favors capture: polluting the common pool of knowledge is cheaper than cleaning it.

A new oligarchy in formation. The four regimes describe a medium that has lost its certifiers. AI may be supplying new ones. A handful of frontier labs now produce the default answer to a growing share of everyday questions, and an answer engine collapses many sources into one voice, which puts it closer to the priest than to the library. The scarce input is frontier computing power, playing the part that spectrum and the scriptorium once played. The analogy breaks at exit. There was no open-weights Church: open models trail the frontier by months (Section IV), and the labs compete with one another instead of coordinating. If this is an epistemic oligarchy, it is the leakiest one on record. Which way it goes is the race between diffusion and capture, run in the information domain. Section X records the American state treating access to a frontier model as a licensable export. A licensed AI oligarchy would be the broadcast model again, with the state holding the license. Metric 2 on the dashboard, the gap between frontier and consumer AI, is the gauge of how strong such an oligarchy is.

Every layer depends on this medium. A state can’t keep its legitimacy at home if it can’t at least influence the story. A bloc holds together through a shared sense of threat, which is an information product. A network needs members who see roughly the same reality. And an individual’s autonomy consists of decisions, which are only as good as the information behind them. If the inputs are polluted, Layer 4 “sovereignty” is confident decision-making on garbage.

What to watch. The ratio I’d most like to measure is narrative velocity against the speed of events: how fast an information operation can reshape what people believe happened, compared with how fast things actually happen. When the story outruns the event, the information environment becomes the main battlefield and physical events become props in a script already written. My judgment is that this ratio has been rising for two decades. I don’t yet have a clean way to measure it, and Section VIII says so.

Assessment. The information environment doesn’t decide outcomes directly, as energy or demographics do. It is the medium through which every other variable is perceived and acted on. A framework that models the hardware and ignores it is trusting sensors that may be lying.

Climate and water: the forcing function

The energy transition isn’t happening in a vacuum. It is driven by a physical force, climate change and its effects on water, that destabilizes every layer on the same fifteen-year horizon as the scenarios below.

This is not a climate essay, and it takes no position on policy. But it has to register that climate changes the ground under everything else.

• At Layer 1, water stress will strain borders and legitimacy in the Middle East, Central Asia and the India–Pakistan corridor. By UN estimates, about half the world’s people already face severe water scarcity for at least part of the year. Migration driven by climate will dwarf today’s migration debates, and volatile harvests will test states already under stress.
• At Layer 2, extreme weather is already a measurable cost to supply chains, and a rising one. Coastal infrastructure, including undersea cables and data centers, exposes Layer 3 to physical risk as well.
• At Layer 4, where a person can live starts to matter as much as where they would choose to.

The Jevons pattern is likely to hold for water as well: efficiency will unlock new demand. And desalination at scale takes enormous energy, which ties the water question directly to the energy question.

Assessment. Climate is neither a layer nor a sorting function. It is the ground shifting under the whole stack. Any scenario that leaves it out is incomplete on a fifteen-year horizon.

The central question is not whether AI will transform how societies are governed. It will. The question is whether the transformation spreads power or concentrates it. I set out three hypotheses, because a two-way choice hides the likeliest answer.

Hypothesis A: diffusion wins

AI follows the path of the smartphone. The cost of running a good-enough model falls to consumer prices. Open-weight models stay legal. Energy stops being a gate. Layer 3 and Layer 4 sovereignty become achievable for a large share of the world’s people.

Signs that favor it:

• Consumer hardware that can run a capable model at home falls below $500
• Open-weight models stay close behind the frontier; by the research group Epoch AI’s measure they trailed by about a year in late 2024 and by about four months in 2026
• More AI runs on personal devices and less on company servers
• Privacy-preserving technology goes mainstream
• The energy cost of running AI keeps falling
• More home energy systems can run independently of the grid

What it has going for it. Innovation has tended to favor open systems. Attempts to hold information technology under central control have usually leaked or failed over time, from the printing press onward, and digital goods drift toward wide distribution once the cost of a copy nears zero.

Where it is weak. It needs Layer 1 to tolerate open tools and Layer 2 to keep economic infrastructure open. It is the most conditional of the three.

Hypothesis B: capture wins

AI capability concentrates in a few frontier labs and the states and corporations that fund them. Regulators classify advanced models, or the computing power to train them, as controlled technology. Energy costs keep the frontier behind enormous capital outlays. States use AI for surveillance and social control, and allow the appearance of upper-layer access as a pressure valve.

Signs that favor it:

• The gap between frontier and consumer AI widens or holds
• Major jurisdictions restrict or export-control open-weight models above a capability threshold
• Nine-tenths or more of consumer AI features run on company servers by 2028
• Central bank digital currencies arrive without meaningful privacy
• Freedom House’s record of global democratic decline continues. Its 2026 report counted the twentieth consecutive year
• AI surveillance is deployed faster than countermeasures
• Compute thresholds in regulation that in practice bar independent developers from training models

What it has going for it. States have guns. Capital has inertia. Regulatory capture is a well-understood and effective strategy. And authoritarian systems are proving adaptable in using AI for control: China’s surveillance and social-credit systems, Russia’s managed information space, the Gulf monarchies’ technology-heavy rule. These are not dinosaurs. They are a rival and currently working model of AI-era government.

Where it is weak. The bad-news problem gets worse as the pace of change rises. Centralized systems reward loyalty over competence, and the penalty on innovation compounds. The feedback loops that fast adaptation requires are the ones authoritarian systems suppress.

Hypothesis C: managed diffusion

This is the outcome I think most likely over the medium term, and the one the two-way framing was hiding: capability spreads at the level of tools, while the rails stay captured. By rails I mean the shared infrastructure everything else has to run on: identity, payment, computing, connectivity and, increasingly, the default answer. Open models exist and anyone can get them. But what you need in order to use them at scale, identity verification, payment processing, cloud computing and network access, stays centralized, monitored and revocable.

In practice: you can run a language model at home, but your internet connection is registered to your verified identity. You can make things, but platforms decide who sees them. You can hold cryptocurrency, but the exchanges that turn it into ordinary money check your identity first. You can ask anything, but the answer most people get comes from a few providers whose defaults are set upstream. The tools of sovereignty are available. The infrastructure of sovereignty is not. The result is a population that feels empowered and remains structurally dependent, which makes it the most stable form of control over the medium term, because it produces real satisfaction instead of resentment.

Plate 04Tools / rails

Signs that favor it:

• Open-weight models stay legal, but the products most people actually use are built to need a cloud connection
• AI capability spreads while the infrastructure around it (identity, payment, connectivity) consolidates
• Everyday questions are answered by a few assistants even as open models multiply
• Jurisdictions adopt “permissioned openness”: tools are available, and use is monitored and revocable
• Surveys show people highly satisfied with AI and largely unaware of how dependent the infrastructure makes them

What it has going for it. It is the most politically sustainable outcome, because it meets the demand for access while keeping the levers of control. It avoids the instability of full diffusion, which threatens the state’s legitimacy, and the innovation penalty of full capture. It is the path of least resistance for incumbents.

Where it is weak. Over long periods it is unstable. Tools of autonomy, once widely held, create demand for the infrastructure of autonomy, and each generation of users is more capable of building alternatives. Managed diffusion is a holding pattern. In time it either opens toward diffusion or closes toward capture.

The asymmetry in time

Capture holds the medium-term advantage, to roughly 2035. Surveillance and control scale faster than the infrastructure of independence. A regulatory capture can be completed in one legislative session; diversifying an energy supply chain takes a decade. Managed diffusion is the likely holding pattern for that period. Diffusion’s advantages are long-term: the innovation penalty and the bad-news problem of centralized systems compound, but slowly.

Four races decide the outcome:

• how fast AI access spreads, against how fast regulation and infrastructure control close around it
• how fast energy supply chains diversify, against how fast manufacturing and resources concentrate
• how fast authoritarian systems lose touch with reality, against how fast they deploy surveillance
• how fast workforces shrink, against how fast automation replaces the missing labor

No one of the three hypotheses is “correct”. All three dynamics are real and operating at once. The outcome depends on relative speeds, on sequence and on geography, and different regions may live under different hypotheses at the same time.

Japan is the most instructive case I know, for two reasons. It faces pressure at every layer while living through the sharpest version of the demographic constraint. And its culture offers the most suggestive example I know of a way to organize the upper layers that is neither libertarian nor authoritarian.

Layer 1: capable and constrained. Japan has one of the world’s most technologically advanced militaries, inside a constitution (Article 9) that restricts its use. The American alliance supplies the nuclear umbrella and the reach that Japan’s constitution and long-standing policy keep it from building alone. That is Layer 1 sovereignty outsourced to an ally. It works as long as the ally’s commitment is believed, which is exactly what the progressive collapse has put in doubt.

Layer 2: leverage and exposure. Japan holds critical positions in the equipment and materials used to make semiconductors, chokepoints that give it more leverage than its economic weight alone would. It is also among the most import-dependent of the large economies for energy. Since the Fukushima accident of 2011, with less nuclear capacity, heavy reliance on imported gas and cautious adoption of renewables, it has been acutely exposed to the energy variable.

Layers 3 and 4: capability, friction and a possible answer. Japan has the technology for the upper layers. What it doesn’t share are the assumptions about privacy, obligation and the individual’s relation to the group that underlie either Silicon Valley’s libertarianism or China’s state-directed model.

That friction may prove valuable. Japanese institutions put a high value on wa, harmony treated as something a society works at, on long time horizons, and on collective restraint held as a positive good. That is a generalization about a complicated country, and I offer it as a pointer, not a finding. It may be the closest existing prototype for upper-layer governance that rejects both the atomized individual and the command hierarchy. It suggests that sovereignty at the top of the stack could be organized around discipline, mutual obligation and patience.

The open question is whether any of this travels. Japan developed it in an unusually homogeneous society, and one that is now shrinking. Whether the principles can be separated from the culture, whether wa is a philosophy of governance or a feature of one place, is unresolved.

The lesson. Japan shows that the unit of analysis is the interaction between layers, not any single layer. Its semiconductor leverage offsets its military constraints. Its energy exposure limits its autonomy despite its technology. Its demographics reshape the power available at every level. No single-layer reading captures Japan’s position. The stack has to be read as a system.

Cross the two main material variables, energy and AI, and four end states appear. Demographics and the information environment modify each one. The Logos decides whether upper-layer sovereignty is real or nominal wherever it is possible at all.

The four quadrants are attractors: states the system tends toward, not snapshots of any moment. The likeliest condition for 2026 to 2035 is none of them cleanly. It is the transitional hybrid of Hypothesis C, showing features of several quadrants at once, with different regions in different places.

I. Network Renaissance (AI diffuses, energy diversifies). Accessible AI and diversified energy make real Layer 3 and Layer 4 sovereignty possible. Innovation speeds up and experiments in distributed governance multiply. States persist but lose relative power. Likelihood over fifteen years: low. It needs three uncertain things at once: open-weight AI surviving legally, energy supply chains diversifying at scale, and a workable Logos. What would raise it: capable home AI hardware under $500 by 2028; the legal fights over open weights resolved in favor of openness; modular nuclear reaching commercial scale; a governance model for the upper layers that works.

II. Asymmetric Empowerment (AI diffuses, energy concentrates). The tools are everywhere, but concentrated energy creates geographic winners and losers. Individuals in energy-rich regions thrive. Elsewhere people have the tools and not the infrastructure. Likelihood: low to medium. Working remotely across borders becomes a survival strategy, and talent moves toward energy-rich jurisdictions. Young populations in energy-poor regions are hit hardest. What would move it: Chinese overcapacity flooding developing countries with cheap panels (toward empowerment); regulation that keeps off-grid systems tied to central control (away); better batteries making small independent grids practical (toward).

III. Neo-Westphalian Blocs (AI is captured, energy diversifies). Frontier AI sits in labs tied to states, but diversified energy stops any one bloc from dominating. The result is a multipolar world of capable blocs competing on familiar great-power lines. Layers 1 and 2 dominate, and the upper layers survive as managed spaces inside bloc boundaries. Likelihood: medium to high. It is also the outcome the last breakup of an epistemic oligarchy produced (Section III): fragmentation, then territorial settlement. This is spheres of influence for the twenty-first century, with three or four poles in place of two. AI export controls spread and regional AI ecosystems diverge. What would raise it: export controls extended to open-weight models; the EU achieving real AI capability and not only AI regulation; BRICS acquiring institutional depth; demographic divergence forcing each bloc into its own immigration and labor strategy.

IV. Techno-Feudalism (AI is captured, energy concentrates). The name is borrowed from writers such as Yanis Varoufakis. Frontier AI and energy both end up with states and the largest corporations. Upper-layer sovereignty is nominal: the tools exist, and access is gated, watched and revocable. The two-tier economy of intelligence arrives in full. Likelihood: low to medium. It needs both captures to succeed together, so it is likelier as the fate of particular authoritarian regions than of the world. What would raise it: central bank digital currencies without privacy; a war over Taiwan wrecking semiconductor diversification; consumer AI moving almost entirely onto company servers; continued global democratic decline.

Crossing between quadrants. The system moves. If open-weight models lose their legal status in the United States and Europe, the world shifts from III toward IV. A breakthrough in modular nuclear power shifts it from IV toward III, or from III toward I. A major war in the Taiwan Strait or the South China Sea resets everything to raw Layer 1 competition and suspends the upper-layer dynamics for its duration, whichever quadrant we were in.

A framework like this is only useful if it can be watched and proved wrong. This section says what to watch and what would count as failure.

What to track

Some of the names: the IEA is the International Energy Agency and Ember an energy research group; the BIS is the Bank for International Settlements, the central banks’ own bank; IEEFA and Bruegel are energy and economics institutes; ASPI is the Australian Strategic Policy Institute; DFRLab is the Atlantic Council’s Digital Forensic Research Lab; and TSMC is the Taiwanese company that manufactures most of the world’s advanced chips. Metric 15 is the weakest on the list, and I keep it as a placeholder for something that needs a better instrument.

What would prove it wrong

Hypothesis A (diffusion) is dead if:

• by 2028 Apple, Microsoft and Google have moved nine-tenths or more of their AI features onto their own servers
• major jurisdictions export-control open-weight models above a capability threshold
• the price of capable home AI hardware stalls above $2,000 while the frontier accelerates
• privacy-preserving technology stays marginal, under 5 percent of Layer 3 transactions
• regulation forces 95 percent or more of new local energy systems to stay tied to the grid with a central off-switch

China’s October 2026 draft controls (Section X, revision 7) build the mechanism for the second condition without yet meeting it.

Hypothesis B (capture) is weakened if:

• consumer AI shows professional-grade ability on hardware under $500
• open-weight models stay within a year of the frontier despite regulatory pressure, as they have so far
• the energy cost of running AI falls far enough for wide use in developing countries
• the democratic decline reverses for two consecutive years
• home battery storage halves in cost, making off-grid life practical

Hypothesis C (managed diffusion) is confirmed if:

• open models stay legal, but the products most people use need a cloud connection tied to verified identity
• AI capability spreads while payment, identity and connectivity consolidate
• people stay highly satisfied with AI tools and largely unaware of their dependence
• “permissioned openness” rules spread across major jurisdictions

Hypothesis C is wrong if tools and rails stop diverging: if open-weight models are restricted along with the infrastructure, which is capture, or if identity, payment and connectivity open up along with the tools, which is diffusion. I have not yet given this test dates and thresholds as strict as those for Hypothesis A. Until I do, the hypothesis I favor is the least testable of the three, and a reader should hold that against it.

The thesis of Chinese solar hegemony breaks if:

• demographic or fiscal brittleness disrupts China’s supply chain before others diversify
• dumped Chinese panels build energy independence in developing countries faster than they build dependence
• an energy technology China doesn’t control reaches cost parity with solar before 2035

The whole framework needs rebuilding if:

• a war in the Taiwan Strait or the South China Sea collapses the upper layers into pure Layer 1 competition
• AI capability jumps so sharply, to general intelligence or its equivalent, that the race between diffusion and capture stops mattering
• a new model of governance solves the Logos problem in a way this framework doesn’t foresee
• climate disruption (wars over water, mass migration, failed harvests) overwhelms every layer before the scenarios mature

Partial failure. Complex systems usually fail by degrees. This framework may turn out to be right in direction and wrong in timing, by five years either way. It may hold for the Indo-Pacific and the Atlantic and fail to describe Africa, Latin America or Central Asia. It may reach the right outcome by the wrong mechanism, or miss a variable that proves decisive. I expect some of these. They should lead to revision, not abandonment.

Structural forces make the landscape. People still move through it. The stack doesn’t move itself. It is moved by decisions taken by fallible people in Politburo sessions in Beijing, boardrooms in Tokyo, committees in Brussels, campuses in Silicon Valley and garages in Murrieta, California.

What states can do

• Export controls on chips and AI models. The most direct instrument of capture. The United States uses them against China; their scope and effectiveness is a primary swing variable.
• Energy policy. Whether to build solar manufacturing at home, restart nuclear programs or extend fossil infrastructure reshapes the energy variable over a decade.
• AI rules. The EU’s AI Act, American executive action and Chinese regulation draw the boundaries that decide which hypothesis prevails in each jurisdiction.
• The design of central bank digital currencies. Privacy or full visibility: an either-or choice with large consequences for Layer 3.
• Immigration. Given the demographic split, the ability to attract and integrate working-age talent is a first-order strategic variable. States that can’t will feel it at every layer.

How a state might read the dashboard. If metrics 2 and 9 improve together (the AI gap narrowing while open weights stay legal), diffusion is gaining, and openness will attract Layer 4 talent. If metric 7 keeps falling and metric 10 accelerates, capture is dominant, and states that choose openness will feel growing competitive pressure from authoritarian efficiency.

What firms can do

• Open or closed release. Each major lab’s decision to publish its weights or gate them moves the trajectory of diffusion.
• Device or cloud. Where the model runs decides whether Layer 4 is viable.
• Infrastructure. Satellite internet and mesh networks get around state control of information, but they still depend on state tolerance. Satellites can be jammed or shot down, and companies can be sanctioned. Tunneling is a capability. It isn’t sovereignty.

How a firm might read it. If metric 4 rises while metric 3 falls, managed diffusion is setting in: invest in both cloud and device. If metric 12 spikes on geopolitical tension, diversify chip suppliers at once.

What individuals can do

• Legal presence in more than one country, as a hedge against any single state.
• Local computing: hardware that runs AI without a cloud connection.
• Skill positioning. As the stack grows more complex, the ability to coordinate across layers becomes worth more than expertise inside any one of them.
• Energy independence at small scale: solar, storage and efficient computing at home.
• Discipline about information. In a degraded information environment, the ability to weigh sources, resist manipulation and keep an accurate picture of events is itself a form of Layer 4 sovereignty, perhaps the most basic one.

Most of that list costs money, which is the luxury-good objection from Section II in practical form. The last item doesn’t.

How an individual might read it. If metric 2 narrows and metric 9 holds, conditions favor Layer 4 autonomy: invest in local AI capability and second residency. If metric 6 worsens where you live, use the tunnels and reduce the points where the state has leverage over you. If metrics 2 and 9 deteriorate together, the window is closing: fall back to a Layer 2 position, meaning institutional employment and a career aligned with your bloc.

These playbooks assume cleaner signals than the dashboard will actually give. They illustrate how to reason from the metrics. They are not instructions.

The orchestration thesis, and its limits

At every level, the actors best placed for what is coming are not the strongest within one layer. They are the ones most able to work across layers: the state that can use economic blocs while keeping its digital infrastructure open, the firm that can handle several regulatory regimes without losing technical edge, the individual who can coordinate among physical, economic, digital and AI-assisted ways of working.

That is the orchestration principle: in a system of many layers, the ability to translate between them is worth more than dominance in any one.

History offers a counter-pattern, and it has to be faced. Rome, the Mongol Empire, the British Empire and the postwar United States did not rise by orchestrating between layers that others controlled. Each was a single power that came to dominate every layer at once, with force as the foundation. That is hegemony, not orchestration. Orchestration may describe something new about this transition: a world in which no single power can achieve the dominance at every layer that earlier hegemons enjoyed. Or it may describe the passing condition of a system between hegemons, which ends when a new one imposes order.

The test is simple to state. If either the United States or China achieves dominance at every layer within this framework’s horizon, the orchestration thesis fails and we are back to hegemony, with the other layers as features of the hegemon’s design. The evidence so far, mutual constraint between the two, no actor self-sufficient at every layer, and the progressive collapse of the last hegemon’s order, suggests orchestration is more than an interlude. The evidence isn’t conclusive, and I would sooner hold the uncertainty than resolve it early.

The sections above were written in February 2026. In June 2026 the United States government forced a commercial frontier AI model offline for nineteen days, and I went back to test the framework against the record. This section reports what changed and introduces the instrument I now use to keep the framework honest: a dated ledger of clashes between the layers.

This was not the first shot

It is tempting to call the Fable shutdown the shot heard round the world. That makes a good story and bad analysis. The conflict between the layers did not begin with the shutdown of Fable 5, the most capable AI model the American lab Anthropic had released, and Mythos 5, a restricted version of the same model. It began in 2022, with two strikes by Layer 1 against the layers above it within weeks of each other. In August, the US Treasury sanctioned Tornado Cash, a “mixer”: autonomous software on a public blockchain that pools cryptocurrency from many users and pays it out again so that the trail is lost. In October, the Commerce Department imposed export controls on advanced semiconductors bound for China.

Fable is the first time Washington has forced a commercial frontier AI model off the air. That is a new phase, not the opening. The ledger below starts in 2022 on purpose, to resist the bias that makes the most vivid event feel like the origin.

What did not change is worth saying first. The framework’s central call was that capture would run through export controls and through control of cloud access, and that managed diffusion would lead in the medium term. That call is holding against live events. The seven revisions below correct the misses. Each carries a confidence mark: known, inferred or speculative.

Seven revisions

1. The delay disappears at home. Section II treated the lag between action and consequence as a structural feature of the stack. It is a feature of distance. Against a company or citizen within its own reach, Layer 1 acts almost without friction. The Fable directive, a non-public letter from the Commerce Department requiring a license for access by any foreign national, was issued on June 12, 2026, three days after the model launched, and published by Bloomberg four days later. The worldwide shutoff followed within hours. The delay applies only when enforcement has to cross a border the state doesn’t control. Known.

2. The veto is paperwork first. Physical force is the floor of the state’s veto, not its working instrument. The working instrument is administrative: an export-control directive, a secret technical order, a sanctions listing, a license. And the state rarely seizes the thing itself. It cuts access to the rail. The model, the code and the weights (the numbers that make up a trained model) still exist; what is withdrawn is the ability to reach them. This is managed diffusion made literal. The instrument is old. The Church answered the printing press with licenses and the Index of Prohibited Books, first issued in 1559: control the permission, not the press. There is a cost, too. A veto used arbitrarily spends legitimacy and invites a challenge on process, from the target and from rival states. Anthropic did not respond to the directive by denying the model’s capability. It attacked the process, saying such action should be “transparent, fair, clear, and grounded in technical facts”. When Britain ordered Apple to weaken its encryption, the United States objected on civil-liberties grounds: one state contesting another state’s veto. Known for the first two points; inferred for the third.

3. The manager has a name, and it is not coherent. Hypothesis C left a question open: who manages the captured rails? The first real answer is that it is no single mind optimizing for security. It is a contested apparatus that can blacklist a firm, be sued by it, depend on it and regulate it all at once, and that may act on information supplied by the firm’s commercial rivals. The same administration had designated Anthropic a supply-chain risk in February 2026 and was sued over it. In September a federal appeals court upheld the designation two to one, while a separate district-court ruling against the government still stands. By published accounts it went on using the company’s models where it had no alternative, and it issued the Fable directive after another technology company reportedly told officials that its researchers had broken the model’s safeguards, and after an attempt to have the launch paused had failed. Decisions to capture come out of bureaucratic and lobbying conflict, not strategy. “Managed” diffusion is less managed, and more contingent, than the name suggests. Inferred, from one richly reported event.

4. The legibility penalty. A capability that is demonstrated in public invites the veto. Visibility is exposure. The more legible a capability is, meaning the easier it is for an outsider to read, through published benchmarks, disclosed offensive potential or advertised usefulness, the easier its owner is to capture. The models’ disclosed ability to find flaws in operating systems and browsers was the lever used to pull them. Tornado Cash’s visible usefulness, including to North Korea’s Lazarus hacking group, was the hook for sanctions. In both cases what made the capability valuable made it a target. For the askesis hypothesis this adds a corollary: discipline means more than carrying extra burdens. It includes managing what you show. Inferred, from two cases.

History suggests a refinement. Jan Hus and Martin Luther were both summoned to answer for heresy under an imperial safe-conduct. Hus was burned at Constance in 1415. Luther came to Worms in 1521 as the most printed author of his generation, with a sponsor, Frederick the Wise of Saxony, who hid him afterward. Visibility, joined to a constituency and a protector, made him too costly to kill. The penalty may rise with visibility only up to a point and then reverse: exposure at middling scale, protection at large scale. Speculative, from one historical pair.

5. The human anchor. When the state can’t capture an autonomous artifact, whether unchangeable code, weights already distributed or a protocol with no operator, it captures the person attached to it: the developer, the operator, the founder. The code may be out of reach. The coder has a physical address. This is the clearest demonstration in the whole ledger. The sanctions on the Tornado Cash software and the prosecution of its developer ran side by side; he was indicted in 2023. When a federal appeals court put the software beyond the reach of sanctions, the case against the person went ahead regardless, and ended in a conviction for conspiring to operate an unlicensed money-transmitting business. The pattern is old. In 1428 the Church had John Wycliffe’s remains dug up and burned, four decades after his death, his writings having proved harder to destroy than he had been. Layer 4’s deepest dependency is that it is tied to a body inside a jurisdiction. The code can be sovereign while the coder is not. Known.

Plate 05Code / jurisdiction

Read the diagram in text

A distributed artifact may be hard to withdraw. Its developer or operator remains a person living inside a jurisdiction. The two have different exposure to legal pressure.

6. The partition trap. When a state demands selective compliance (block only foreigners, surveil only domestic users) and the technology can’t make that cut safely, the rational response is to withdraw the capability from everyone. So capture overshoots. A narrow demand produces a broad loss. Anthropic shut the models off for all users, its own foreign staff included, because it could not sort them by nationality in real time. Apple withdrew its strongest encryption from an entire country sooner than build a back door that anyone might later find. Capture is cruder than the scenario matrix implies, and it routinely breaks more than it means to control. Known as a pattern; inferred as a general rule.

7. The first threshold on open weights came from Beijing. The framework assumed that if open weights were ever fenced, Washington would build the fence. On October 16, 2026, China’s commerce and science ministries published draft export controls that name model weights, source code and training methods as controlled technology once a model passes any of three thresholds: 10^26 training operations, two trillion parameters, or a global top-ten ranking. Weights already public are exempt, and a future open-source release would receive automatic licensing, so nothing now downloadable is withdrawn. A temporary pre-export registration requirement took effect the same day. Washington has moved the other way: its voluntary pre-deployment testing regime of June 2026 exempted open-weight models, and the Biden-era controls on closed-model weights have not been enforced since May 2025. The day before Beijing’s draft, Meta released the weights of an 816-billion-parameter model, reversing its own move to closed weights earlier in the year. This is permissioned openness, the sign Hypothesis C predicts, in its plainest form: release stays legal and the state keeps the permit. It does not trip the Hypothesis A kill condition in Section VIII, which requires export controls on open-weight models; it builds the mechanism for one. Known for the rule; inferred for its meaning.

The ledger

Each entry records who acted against whom, by what mechanism, and how that round came out. Results are scored by round, not by war, and several remain open.

What each entry means:

• Tornado Cash sanctions. The first attempt to treat autonomous code as something that can be sanctioned, and aimed at a tool built on the very privacy mathematics Section II calls the upper layers’ best defense. The founding test of state against network.
• Chip controls. The template that every later capture of an AI rail has followed, Fable included. Export control is the state’s instrument of choice against computing power.
• Próspera. A Layer 3 project fighting a Layer 1 state with a Layer 2 instrument. The clearest case on the board of the layers being used against each other.
• Durov in France. The human anchor applied to a platform: the state could not reach Telegram’s servers, so it reached the founder at an airport.
• Van Loon. A hard legal ceiling on capturing autonomous code as code. The most important legal win for diffusion on the board.
• DeepSeek. If near-frontier capability is open and cheap, gating one model achieves little. This is the standing counterweight to every capture entry.
• Ulbricht. The only release of an anchored individual on the board, and it came from inside Layer 1. Diffusion by sponsorship, the Luther pattern.
• Apple and Britain. The partition trap in its purest civilian form, and a live clash between two states over one state’s veto. In August 2025 American officials said Britain had dropped the demand; within weeks a new order was reported, limited to British users. Apple filed a fresh challenge to the narrower order in mid-2026; a related challenge by Privacy International is listed for a substantive hearing in December 2026.
• The GENIUS Act. The subtle capture. The escape route for money was not suppressed. It was licensed and turned into an extension of the dollar. This is what managed diffusion looks like when the captured side prefers being captured.
• Storm. The human anchor. When the code is out of reach, the coder is not. The case is unfinished: a motion to set aside the conviction is undecided, and prosecutors intend to retry the counts on which the jury deadlocked.
• Anthropic designation. A procurement tool used against a domestic frontier lab. Capture by exclusion, not by seizure.
• Fable. The first forced takedown of a commercial frontier model. The restricted version returned to vetted American organizations after two weeks. The controls were lifted on June 30 and public access resumed the next day, once a screening system tested by a Commerce Department institute was in place. I score the round a draw: the rail came back, on terms the state set. The legal question of whether access to a model is an export is unresolved. Watch whether the precedent extends to other providers.
• Durov in Russia. Two states now hold open cases against one founder. Multiple citizenship did not prevent either.
• Meta’s release. The counterweight to Beijing’s draft, one day earlier. A license with commercial limits is still open weights.
• China’s draft controls. The first permit regime for model weights anywhere. See revision 7.

Reading the board

Capture leads in the fights over rails, without sweeping them. Computing power is captured, with leakage. Encryption is contested and unresolved. The AI frontier was held for nineteen days and released on conditions. In each case the instrument is the same: administrative orders, not force. The open-weight rail now has its first permit regime, drafted in Beijing (revision 7).

Capture has been checked on code-as-property. Van Loon set a real limit on sanctioning autonomous code, and the state’s case against the human anchor went on. The protocol has a partial legal shield. The people around it do not. The shield came from Layer 1 itself, in the form of a court. That fits the oldest pattern in the record: Luther survived because a prince found it useful to protect him. Diffusion tends to win when some part of a lower layer sponsors it. Inferred.

The money rail was annexed without a fight. The cleanest capture on the board, and the quietest.

Diffusion’s live wins are open weights, Meta’s return to them, the Van Loon ceiling and Ulbricht’s pardon. Thin, but real.

The human anchor is the busiest instrument on the board. Three entries are actions against a named person: Storm, and Durov twice. When the state cannot reach the code, it reaches the founder. The one release on the board, Ulbricht’s pardon, came from another part of the state. Inferred.

The verdict so far: Hypothesis C is leading. Tools spread and stay legal. Rails stay captured, watched and revocable. This is what the February text named as the likeliest medium-term pattern, and it is the part of the framework to defend.

The standing objection. The ledger is kept by the person who built the framework, choosing events that fit it. The danger is motivated pattern-matching: every administrative act reads as “capture” when some are ordinary regulation, posturing or litigation tactics in national-security dress. The Fable episode is the test case. Access came back in under three weeks, and one of the ledger’s own rules says that rails restored within days mark posturing, not structure. A single reversal doesn’t settle that. It does count against reading the episode as a clean capture, and I have scored it accordingly. The rules below exist to fight this bias. They are not there to decorate the thesis.

How the ledger is kept

• Inclusion. An event enters only if it is a concrete, dated action, by an identifiable actor at one layer, against or escaping another layer, by a mechanism that can be named. Opinion pieces, forecasts and “concerns” don’t qualify.
• Symmetry. Wins for diffusion and failures of capture are logged as carefully as wins for capture. After three capture entries in a row, I go looking for counter-evidence before adding a fourth. A ledger that only grows toward its thesis is a campaign poster.
• Scoring. Every entry carries a round result: capture, diffusion, draw or open. Open entries are revisited on a schedule, and closing one is itself an entry.
• What would overturn the model. If captured rails are routinely restored within days under pressure, capture is tactical posturing and the thesis should be downgraded. If open weights stay level with the frontier through repeated capture attempts, Hypothesis A gains. If prosecutions of developers keep failing in court, the human anchor is weaker than I have claimed.
• Confidence. Every entry and every revision carries its mark, and the prose is not allowed to sound surer than the mark.

The post-Cold War order failed in stages across six shocks: 9/11, the 2008 financial crisis, Crimea, Brexit and the 2016 US election, COVID-19 and Ukraine. Each exposed a deeper layer of the same fracture. The old order is gone. What replaces it is a layered stack of nation-states, economic blocs, digital networks and augmented individuals, operating at once, as adversaries, with dependence running up the structure and constraint running down.

Who gains and who loses across that stack depends on three material variables (energy, AI capability and demographics), one philosophical variable (the Logos, the unsolved question of what gives the upper layers coherence and legitimacy) and two cross-cutting forces (the information environment and climate).

Energy stays scarce in aggregate, but access can still spread at the margin if the cost of AI follows the smartphone’s path. China’s hold on the solar supply chain is real and brittle. America’s fossil-fuel lever is powerful and temporary. Demographics split the world into the aging and the young. The information environment is degrading under every kind of regime. Shared truth at scale was historically manufactured by epistemic oligarchies, and the breakup of the last great one ended not in individual sovereignty but in territorialized belief. The frontier AI labs may be the next such oligarchy, and the leakiest on record.

Three hypotheses compete: diffusion, capture and managed diffusion, in which tools spread while infrastructure stays captured. Time favors capture and managed diffusion over the next decade and diffusion beyond it. The dated record since 2022 supports that reading: managed diffusion is leading, though the most dramatic capture on the record lasted nineteen days.

The Logos problem is the deepest open question. Technology alone won’t solve it. The askesis direction, sovereignty as a discipline practiced and not a luxury purchased, is the most promising hypothesis I have, and it is unproven.

The state keeps its veto, but the veto is latent. It needs the rest of the stack to be exercised, it works mainly through paperwork, and it costs legitimacy when used crudely. Japan shows that the interactions between layers, not the layers one by one, are what must be analyzed.

The most consequential decisions shaping this landscape are being made now, by people who may not fully understand the system they are shaping. This framework is an attempt to make that system legible. It does not predict the outcome. It identifies the variables that decide it and commits to tracking them honestly. It is a map of the territory as well as I can chart it, with the uncharted regions marked as such.

A note on method. This essay makes three kinds of claim. Empirical claims are sourced in the dashboard or in the notes below and can be checked. Structural claims reason from history or from how systems behave; they can be tested, but only over years. Judgments of likelihood are informed opinion, stated as ordinary words (low, medium, high) and not as percentages, to avoid false precision. No claim here should be given more confidence than its kind deserves. The framework was built from a Western vantage point and carries that inheritance; later versions should engage Chinese strategic thought, Indian multi-alignment doctrine and African Union frameworks on their own terms.

First written in February 2026, as the invasion of Ukraine reached its fourth anniversary, and revised in June, October and November 2026. Developed by John Fite with Claude, and critiqued in draft by Claude Code, Grok, Gemini and Codex. It is a living document.

First published June 2026 · Revised November 2026