In October 2025, Samsung and SK hynix entered strategic partnerships with OpenAI to scale memory production toward a projected 900,000 DRAM wafer starts per month—roughly 40 percent of estimated global capacity. The figure is a production target based on projected demand, not current consumption, but OpenAI and Samsung both confirm the intended scale.

By July 2026, the other end of the market looked increasingly absurd. AMD introduced a new entry-level Radeon with just 4 GB of memory. The following morning, a Finnish retailer listed a Gigabyte RTX 5090 at €5,599.90, warning that scarce stock would be released for sale only once per day.

These are not three unrelated curiosities. Together, they show a market being reorganized around buyers who can shape productive capacity at global scale: compromised capability at the bottom, boutique scarcity at the top, and the industrial center of gravity moving away from ordinary ownership. They are early symptoms of what I will call the Metered Intelligence Complex.

Core thesis

The argument is not simply that computing hardware is becoming scarce. It is that several independent forces are pointing toward the same political-economic settlement: intelligence becomes a metered utility, controlled by strategic corporations and governed in partnership with states that increasingly depend on them.

Each vector has its own immediate cause:

  • concentrated manufacturers protect margins by limiting oversupply;
  • hyperscalers reserve the inputs required to convert compute into intelligence;
  • economic inequality directs scarce capacity toward institutions rather than households;
  • utility delivery replaces owned capability with conditional, revocable access;
  • concentrated AI capability lets institutions influence the public through its own coordination channels;
  • governments subsidize and protect the firms they need for economic, military and administrative power;
  • regulators avoid fights that might weaken strategically important national champions;
  • surveillance makes resistance to the resulting allocation more legible and costly.

The sum matters more than the motive behind any one part. Memory manufacturers do not need to want state surveillance. Police do not need to care who receives GPUs. Cloud providers do not need to design an unequal economy. Each actor can pursue a narrow, locally rational objective while moving the whole system in the same direction.

The resulting relationships look like this:

flowchart TD
    investors["Investors"]
    manufacturers["Manufacturers"]
    hyperscalers["Hyperscalers"]
    government["Government"]
    regulators["Regulators"]
    platforms["Platforms"]
    police["Police"]
    public["Public / households"]

    investors -->|"reward capacity discipline"| manufacturers
    investors -->|"fund compute expansion"| hyperscalers
    hyperscalers -->|"reserve capacity with long-term contracts"| manufacturers
    manufacturers -->|"allocate scarce chips and memory"| hyperscalers
    manufacturers -->|"offer residual supply at higher prices"| public

    government -->|"subsidize and protect as national champions"| manufacturers
    government -->|"subsidize, contract with and protect"| hyperscalers
    hyperscalers -->|"provide cloud, AI and intelligence infrastructure"| government
    regulators -.->|"avoid difficult enforcement against"| manufacturers
    regulators -.->|"avoid difficult enforcement against"| hyperscalers

    hyperscalers -->|"meter intelligence as revocable service"| public
    public -->|"pay rent for access"| hyperscalers
    government -->|"authorize public-order enforcement"| police
    platforms -->|"supply searchable data"| police
    platforms -->|"collect and retain activity"| public
    police -->|"observe, classify and deter"| public
    public -.->|"attempt collective correction"| government

This state-corporatist outcome is the Metered Intelligence Complex, or MIC. Corporations own and operate the infrastructure that converts compute into intelligence. The state treats that infrastructure as strategically indispensable, helps finance and secure it, and increasingly organizes policy around its requirements. Individuals encounter the resulting capability as a utility: something requested, metered, monitored and withdrawn by institutions they do not control.

The complex behaves as though it were coordinated toward that outcome, despite nobody needing to articulate the whole plan.

The Metered Intelligence Complex is state corporatism organized around intelligence as a utility. No conspiracy is required.

1. The preconditions

The mechanism becomes possible when several conditions coincide.

Concentrated supply

A small number of firms control something difficult and slow to reproduce: memory fabrication, advanced packaging, leading-edge wafers, cloud infrastructure, energy, payment systems.

In a consumer graphics card, concentration appears at two stacked layers. At the first, a buyer seeking a high-performance discrete GPU is effectively choosing between Nvidia’s GeForce and AMD’s Radeon. Intel remains a small third entrant rather than a counterweight: Nvidia ended 2025 with 94 percent of desktop add-in-board shipments, AMD with 5 percent, and Intel without a measurable share.

The processor then needs memory supplied by an even tighter market. In the third quarter of 2025, SK hynix held 33.2 percent of global DRAM revenue, Samsung 32.6 percent and Micron 25.7 percent: more than 91 percent between three firms. GDDR for consumer graphics and HBM for AI accelerators are not interchangeable products, but the expertise, investment and production decisions sit inside the same three suppliers.

This market has crossed the line from structural coordination into literal conspiracy before. Between 1998 and 2002, DRAM manufacturers fixed prices charged to major computer makers. Samsung later paid a $300 million criminal fine and Hynix $185 million. Micron’s own regulatory filing acknowledged evidence of price-fixing by its employees and competitors; it avoided prosecution through the Justice Department’s corporate-leniency programme. This history is not evidence that the three firms are colluding now. Indeed, a later case alleging coordinated restraint during the 2016–2017 upcycle was dismissed because the same behaviour was more plausibly explained as lawful parallel conduct. The history establishes something narrower: this is not a hypothetical market structure, and its boundary between explicit agreement and mutually understood restraint has mattered before.

This matters because competitors can easily observe one another. They do not need secret communications to understand that aggressive expansion would collapse everyone’s margins.

Unequal purchasing power

Consumers may intensely want hardware, housing, energy or medical care. But desire is not economically decisive. The decisive question is who can place the largest credible bid.

A hyperscaler can reserve years of production because the purchase becomes productive infrastructure. A consumer buys one GPU from disposable income.

These are not remotely equivalent forms of “demand.”

The cryptocurrency boom demonstrated the difference between buying a GPU as a consumer good and buying one as an income-producing asset. A gamer compares the price with disposable income and the value of playing games. A miner compares it with expected revenue and the time required for the card to repay its purchase price. When the second calculation remained positive, the conventional consumer price ceiling disappeared.

In December 2021, when mining profitability was still driving demand, new RTX 3080 cards sold on eBay for an average of $1,693 against a $700 launch price. Current-generation Nvidia cards averaged 114 percent above MSRP and AMD cards 92 percent above it. Prices moved with cryptocurrency values and mining returns because buyers were capitalizing future income into the hardware’s present price.

AI infrastructure applies the same logic with vastly larger balance sheets. A GPU that generates saleable tokens, cloud revenue or strategic capability is not priced against what a household can afford. It is priced against the revenue or institutional advantage its owner expects to extract.

A recent trauma of oversupply

Manufacturers remember what happens when they expand too quickly:

  • inventories accumulate;
  • prices collapse;
  • margins disappear;
  • factories become financial liabilities.

The pandemic memory cycle supplied a recent demonstration. Lockdown demand, component shortages and defensive over-ordering encouraged the supply chain to prepare for continued growth. When purchases of PCs, phones and other consumer electronics slowed, inventories remained. In August 2022, TrendForce projected DRAM supply growth of 14.1 percent against demand growth of only 8.3 percent. By the first quarter of 2023, average DRAM selling prices had fallen 20 percent in three months, with supply still outpacing demand even after Micron and SK hynix began cutting production.

The lesson management learns is not merely “forecast better.”

It is:

Never allow supply to run far enough ahead of demand that buyers regain power.

This logic is not hidden. In July 2026, ADATA chairman Simon Chen argued that the DRAM shortage could persist for another decade. More revealing than the duration of his forecast was its rationale: the major suppliers had learned from earlier downturns and would expand cautiously rather than repeat the disorderly capacity races that collapsed prices.

Chen is not a neutral observer. ADATA had accumulated more than NT$30 billion in chip inventory, so continued price increases directly improve the value of its holdings. That makes the ten-year prediction weak evidence for how long the shortage will actually last. It makes the reasoning behind it unusually clear evidence of how “responsible” capacity management is understood inside the industry.

The shortage does not have to be fabricated for this mechanism to operate. It can be materially real and still be managed by institutions that benefit from preventing it from becoming a glut.

A new strategic buyer

AI demand gives manufacturers customers willing to:

  • prepay;
  • sign multiyear contracts;
  • accept take-or-pay commitments;
  • reserve capacity before it exists;
  • tolerate extraordinary margins.

That allows producers to stop behaving like commodity vendors and start behaving like infrastructure providers.

2. Turning compute into intelligence infrastructure

The rush becomes easier to understand if AI is treated not as a website or a piece of software, but as an industrial conversion process.

Training combines data, computation, memory bandwidth and energy to produce a model. Inference combines that model with new context and more computation to produce tokens: text, code, images, classifications, plans and control signals. “Intelligence” here is not a metaphysical claim. It is shorthand for output that a buyer can use to augment or replace some economically valuable cognitive work.

Training: data + compute → model

Inference: model + context + compute → usable output

AI can therefore be understood economically as a compute-to-intelligence converter. If its output can be sold by the token, embedded in a product, used to operate infrastructure or substituted for paid labour, then the hardware running it is productive capital. A GPU and its memory are no longer priced only as components. They are inputs into a machine expected to manufacture a continuing stream of saleable or strategically useful capability.

When the output is new knowledge

By July 2026, “usable output” included mathematical discoveries that had eluded humans for generations. A counterexample credited to Claude Fable 5 refuted the Jacobian conjecture in three and higher dimensions, settling in those dimensions a problem open since 1939. Two months earlier, a general-purpose OpenAI reasoning model disproved Erdős’s conjectured upper bound for the planar unit-distance problem, which had stood since 1946. Mathematicians subsequently checked both results.

In these narrow circumstances, the models produced results that almost no human could produce and that the mathematical community had failed to find across decades of expert effort. That is enough. Intelligence need not be universal to be economically decisive; it only has to exceed available human capability on a valuable task.

A converter does not have to succeed on every attempt to become valuable capital. If more capacity buys more attempts at otherwise unobtainable knowledge, corporations and states can rationally value a compute cluster as a portfolio of potential discoveries.

The cryptocurrency boom was a narrow preview of this calculation. Miners paid according to the income a GPU might produce. AI extends the same logic across a much larger imagined field: software, advertising, logistics, research, administration, weapons, surveillance and the automation of intellectual labour. The prospective return is no longer one coin. It is a claim on a portion of the economy.

The rush does not require every claim about AI to be true. A corporation or government only has to believe that exclusion would be more dangerous than overinvestment. If competitors secure the conversion capacity first, a latecomer may have to rent intelligence from them indefinitely. Capacity therefore has option value even before a profitable use is known. Defensive purchases help create the shortage they anticipate, and the race becomes self-reinforcing.

From an open commodity to allocated capacity

The memory chip may remain technically standardized. The important transformation occurs upstream. Instead of producing large quantities and selling them into a competitive market, manufacturers increasingly:

  1. reserve future capacity for large customers;
  2. prioritize customized, high-margin products;
  3. require long-term commitments;
  4. expand only against relatively secure demand;
  5. leave everyone else competing for residual supply.

The commodity has not disappeared. The open market has been demoted.

Price is no longer primarily determined by the cost of producing another unit. It reflects the value of gaining access to constrained production capacity.

That price does not confront all buyers equally. A hyperscaler has large pools of capital and can turn the same unit of compute into subscription revenue, advertising leverage, platform lock-in, market share, valuation and further financing. Households have far less money, far less access to credit and no comparable way to recover the expense. Suppliers therefore rationally redirect capacity toward institutional customers even while consumer desire remains enormous.

Consumers are not losing because they stopped wanting the product. They are losing because purchasing power has concentrated elsewhere.

Calling this “demand destruction” hides what is happening. Demand is being hierarchically rationed.

This is the transition from:

“How cheaply can we manufacture memory?”

to:

“How much will you pay to guarantee that memory exists for you?”

From ownership to permission

For most of the personal-computing era, ownership was itself an allocation mechanism. An individual bought a general-purpose machine and ran software locally. Once acquired, the machine could continue to compute without a provider approving each use, charging for each operation or necessarily retaining a record of the interaction. Personal computers never eliminated dependence on manufacturers and software vendors, but they created a meaningful domain of capability under the user’s control.

Rented compute reverses that default. The individual no longer owns the machine performing the computation; they submit requests to an institution that owns it. The provider can meter, modify, monitor, ration or terminate access. What looks like a change in deployment architecture is therefore also a transfer of power.

Two paid Claude users have reported losing access without an identified triggering interaction. [1] [2] The particulars matter less than the authority the incidents expose.

Anthropic’s consumer terms describe access as a contractual right rather than ownership, permit suspension or termination, and state that the right to use the service ends immediately when access is terminated. Correct enforcement and false enforcement have the same architecture: the provider makes a remote decision and the capability vanishes. A locally stored model running on owned hardware cannot be withdrawn in the same way.

Personal computing distributes capability through ownership. Utility computing centralizes capability and distributes permission.

3. The K-shaped result

The allocation contest is the expression of an inequality that already exists. In ordinary language, demand means wanting something. In economics, it means wanting it and having the money to pay. A person who needs a computer but cannot afford one registers in the market not as unmet need, but as an absence of demand.

Purchasing power has become extraordinarily concentrated. The World Inequality Report 2026 estimates that the top 10% receive 53% of global income and own 75% of global wealth, while the bottom half receive 8% of income and own just 2% of wealth. The imbalance is still growing at the extreme: the wealthiest 0.001% increased their share of global wealth from almost 4% in 1995 to more than 6% in 2025.

The practical meaning is visible in household accounts. In the Federal Reserve’s 2025 survey of U.S. household finances, 12% of adults said they could not pay a $400 emergency expense by any means. Because of price increases, 60% had used less or stopped using products, 46% had delayed a major purchase, 41% had reduced savings and 16% had increased borrowing. At a certain point, saying that consumers “will not pay” becomes a euphemism. They cannot pay.

A market cannot distinguish that condition from a change in preference. A household that postpones a GPU because housing, food and energy have consumed its income appears to the supplier simply as weaker demand. The supplier follows the money that remains.

Late-July reporting provided an unusually literal example. Taiwan's Economic Daily News, citing BenchLife, reported that Nvidia had notified board partners of a 20–30% increase in the price of GPU packages containing both the chip and graphics memory. Nvidia had not announced a retail-price increase, and the expected pass-through to consumers remained a forecast. More revealing was MSI chairman Joseph Hsu's description of the market: consumer shipments were expected to fall 10–20%, average prices to rise, and competitors to protect gross margins rather than cut prices. Falling volume and rising prices are not treated as a failure if margin discipline holds.

That produces the K-shape. The upper arm owns financial assets, controls corporate budgets and can borrow against future returns. It receives guaranteed allocation, financing and subsidies; the compute it purchases becomes another productive asset, and scarcity can increase its strategic value. The lower arm buys from wages and depleted savings. It receives higher prices, longer replacement cycles, reduced specifications and subscriptions in place of ownership.

Inequality is therefore both an input and an output of the allocation system. Concentrated purchasing power tells suppliers to build for institutions and the wealthy. Preferential access then gives those buyers more productive capital, more income and more leverage, while everyone else pays rent for capabilities they can no longer afford to own.

The upper arm owns the scarcity. The lower arm is priced out of ownership and pays rent to access it.

This is why aggregate indicators can remain strong while ordinary experience deteriorates. Semiconductor revenue, capital expenditure and GDP can rise even as personal computing becomes less accessible. The money has not disappeared from the economy. It has accumulated among actors whose purchases count for more.

This may be a stable economic equilibrium. It is not automatically a stable political one.

When a small group controls productive assets while a much larger group experiences declining access, the larger group has a rational response available: stop treating the rules of allocation as fixed. Through elections, unions, antitrust, taxation, public provision, strikes or revolt, people can try to rewrite the structures that produced the distribution. Once scarcity is understood as an institutional outcome rather than a natural fact, the object of resistance shifts from the price to the system that sets it.

Evidence for that response comes from one of its opponents. In a July 2026 Fox News interview, Foundation for Economic Education president Diogo Costa called it “affordability crisis socialism”: support among young people for rent freezes, free public transport, public provision and price controls. Costa presents this as “anti-economics” produced by inadequate economic education. That is advocacy, not survey evidence. But the diagnosis is revealing: when markets cease to deliver basic affordability, people stop treating market allocation as legitimate and demand political control over prices and provision. The dispute is over whether that response is wise, not whether the structural pressure exists.

The K-shape therefore creates its own destabilizing pressure. The first set of structural mechanics explains how capacity and rents move upward. It does not explain why the people on the lower arm continue to accept those mechanics. For the arrangement to persist, a second set must prevent, redirect, fragment or absorb attempts to change it.

Some of these stabilizers are soft: they describe scarcity as inevitable, offer selective concessions, or narrow the range of reforms considered realistic. Others make collective action more difficult or personally expensive.

An unequal market equilibrium requires a political stabilization system.

4. The Metered Intelligence Complex

The political form of the Metered Intelligence Complex is a state-corporatist settlement. In the sense used here, this does not mean that corporations simply issue orders to government. It means that the operating needs of large firms become requirements of the state, while the material needs of people become claims to be balanced, deferred or policed.

The state has the authority to rewrite the allocation regime, but it depends on the same corporations it might otherwise regulate. That dependency is asymmetric. Strategic firms can withhold investment, relocate production, interrupt essential services, cut employment or deny access to infrastructure. Most citizens, already positioned on the lower arm of the K, control little that the state cannot afford to lose. Their countervailing power has to be created through coordination—elections, unions, protest and strikes—and surveillance makes that coordination more legible, predictable and personally costly.

The military-industrial complex, generalized

This is the military-industrial complex of the twenty-first century. The twentieth-century complex joined permanent geopolitical threat, public procurement and private weapons manufacturing into a self-reinforcing system. The state needed capabilities that only a small number of firms could supply. Those firms grew on public money, became indispensable to national power and gained greater influence over the policies that determined what the state needed next.

The reuse of MIC is deliberate.

Compute and AI reproduce that structure, but generalize it from armaments to intelligence. The same privately owned infrastructure can design weapons, process intelligence, operate public administration, automate corporate work, conduct scientific research and provide cognitive assistance to ordinary people. Every new use makes the suppliers more strategically important. Their commercial expansion becomes a national-security objective, while geopolitical competition makes restraint look like unilateral disarmament.

The state finances fabrication capacity, energy infrastructure, data centers, research and procurement without necessarily acquiring public ownership of what it helps build. Private firms retain the models, hardware and operating expertise; public institutions become customers. As those systems enter military, intelligence and administrative workflows, replacing or regulating their providers becomes progressively more expensive.

The older complex primarily supplied the state’s capacity to wage war. This one also mediates the population’s capacity to think, work, communicate and organize. Its products do not remain in arsenals. They become the utility layer through which both state power and ordinary cognition operate.

The twentieth-century complex privatized the production of military power. The twenty-first-century complex privatizes the production of intelligence itself.

Governments need major technology firms for:

  • AI competitiveness;
  • military and intelligence systems;
  • cloud services;
  • domestic manufacturing;
  • critical infrastructure;
  • employment and investment;
  • geopolitical competition.

That creates conflicting priorities.

The government may want lower prices and competition, but it also wants its strategic suppliers highly profitable, domestically invested and capable of outspending foreign rivals.

Antitrust enforcement therefore collides with industrial policy.

The implicit question becomes:

Why destabilize a strategically important national champion merely because it is extracting rents from consumers?

No explicit immunity needs to be granted. Investigations can simply become slower, narrower and less institutionally attractive.

This does not require a government to decide that it will suppress the many on behalf of the few. Each institution can choose a locally defensible form of stability: protect investment, preserve employment, avoid disruption, maintain strategic capacity. Career and financial incentives do the remaining work. People who make incompatible decisions are denied promotion, funding, access, credibility or future employment. The complex does not need loyal conspirators. It needs replaceable functionaries who understand what is expected. The aggregate effect is to insulate the allocation regime from political correction.

Sometimes the mechanism is proposed almost verbatim. In July 2026, Dean Ball—who had just become OpenAI’s head of strategic futures after serving as a White House AI policy adviser—described the prospect of AI becoming state-provided digital infrastructure rather than a market product as dystopian. He then suggested that the administration could discourage enterprise use of open-weight Chinese models without banning them: agencies could issue advisory guidance raising the possibility of backdoors even if the warning was not well substantiated, creating enough regulatory fear, uncertainty and doubt that regulated companies backed away. Ball also argued that sufficiently capable open models pose genuine risks. What matters here is that his proposed mechanism did not depend on demonstrating those risks. The post was a proposal, not evidence that either OpenAI or the administration had adopted the policy. But it states the structural logic plainly: a corporate policy insider describes how state authority could make an alternative provisioning model institutionally unusable without formally prohibiting it.

The ownership distinction becomes especially stark beside OpenAI CEO Sam Altman’s description of the same endpoint. At BlackRock’s March 2026 Infrastructure Summit, Altman said that model providers would sell tokens and that intelligence would become a utility like electricity or water, bought from companies such as OpenAI on a meter.

This is not a two-way choice between corporate and public provision. It is a three-way struggle among personal ownership, corporate rental and public provision. The legacy personal-computing model allowed the individual to provide their own computation. The other two make an institution the provider of intelligence. They differ profoundly in ownership, accountability and public purpose, but they share one structural feature: the individual receives conditional access to capability controlled elsewhere.

Ball and Altman do not disagree. Together their statements describe a coherent institutional preference: intelligence may become essential infrastructure, but access to it should remain a privately metered market product. Corporate delivery of metered intelligence is presented as abundance; public delivery of digital infrastructure is, in Ball’s account, a dystopian endpoint. What disappears from both framings is the individual who owns enough hardware to provide intelligence for themselves.

Once locally capable hardware becomes unaffordable or too underprovisioned, personal ownership ceases to be a practical option. The apparent contest between corporation and state then begins only after the more consequential transfer has occurred: the individual no longer owns the means of computation.

Strategic dependency can extend to the systems that make a population legible. EU Perspectives describes a “Palantir paradox”: European governments invoke digital sovereignty while police, intelligence, military and public-service workflows become dependent on an American data-integration company. France’s domestic intelligence service adopted Palantir after the 2015 Paris attacks as a temporary expedient, then renewed the contract three times because a sovereign replacement was not ready. France has since selected ChapsVision to replace it, but the transition is expected to take years, with the latest Palantir contract still providing the bridge. Meanwhile, NATO has integrated Palantir’s Maven system into Allied Command Operations.

Palantir does not present this relationship as reluctant procurement. In its own 22-point summary of The Technological Republic, a book co-authored by Palantir CEO Alexander Karp, the company says that “Silicon Valley owes a moral debt” to the country that enabled its rise, that its engineering elite has an affirmative obligation to participate in national defence, and that modern hard power will be built on software. It treats military AI not as a policy choice but as an inevitability whose only open question is who builds it. This is a public theory of state-corporate fusion: private technology firms supply the state’s coercive capacity, while national security supplies those firms with purpose, priority and protection.

This is more than hypocrisy. Emergency procurement creates operational dependency; dependency turns the supplier into strategic infrastructure; and strategic infrastructure becomes costly to replace, regulate or expose. A European-owned substitute may reduce geopolitical dependence while preserving the same capacity to combine data, identify patterns and act on populations. The question is therefore not only who owns the system, but what the system allows institutions to do.

The direction of enforcement follows the distribution of leverage. Corporate demands arrive as constraints on national capacity; popular needs arrive as claims from people whose individual noncooperation has little immediate effect. The state can therefore remain formally representative while becoming operationally preoccupied with securing corporate cooperation and managing the political consequences for everyone else.

Private economic power sets the conditions of stability. Public authority enforces them. Surveillance raises the cost of replacing them.

5. Why surveillance belongs in the Metered Intelligence Complex

Dependence, legitimacy and selective concessions can redirect pressure, but they cannot eliminate the possibility of collective action. Surveillance adds another stabilizing mechanism: it raises the expected personal cost of organizing that action.

Again, this does not require surveillance systems to have been created specifically for suppressing protest. They may be justified through:

  • child protection;
  • terrorism;
  • public safety;
  • fraud prevention;
  • misinformation;
  • border enforcement;
  • crime reduction.

The child-protection example is not hypothetical. In July 2026, the EU reinstated a temporary exception to its ePrivacy rules that allows online providers to voluntarily detect, report and remove child sexual abuse material. The parliamentary path was unusual: 314 MEPs voted to reject the Council’s position, but a second-reading rejection required an absolute majority of 360. Parliament instead adopted narrower amendments, and the Council then gave the measure final approval. The result is not a mandate to scan every message: participation is voluntary, and end-to-end encrypted interpersonal communications are excluded. But a capability introduced in 2021 as a short-term exception has now been restored until April 2028 while a permanent regime is negotiated. Temporary status has become part of the argument for keeping it alive.

Private speech can become legible through consumer hardware as well as communications platforms. LG’s television terms controversy makes this concrete: the device can listen inside the home. LG’s own terms state that nearby third-party voices may be recorded and analysed, and make the owner responsible for warning household members and guests. Whether LG currently transmits or stores those conversations is not the structural question. Once the microphone and processing path exist, privacy depends on every actor capable of accessing or altering that path remaining trustworthy. One malicious insider, compromised update or coercive authority is enough to turn capability into surveillance.

A 2022 incident shows why the intended audience stops being the only relevant audience once private speech passes through a machine. Aditya Verma sent a bomb-threat joke to five friends in a private Snapchat group before a flight from Gatwick to Menorca. British security services somehow obtained the message, alerted Spain, and a fighter jet was scrambled to escort the aircraft. A Spanish court later acquitted Verma, finding that the message was sent in a strictly private setting and that he could not have expected it to be intercepted. The method of interception was never established; claims that Gatwick’s Wi-Fi exposed the message remain speculation.

The parallel is that private speech becomes governable once a device or platform can capture and process it. The intended audience is no longer the only possible audience. Who else can listen is determined by software, permissions and institutional access that the people being heard may neither see nor control.

But once the infrastructure exists, it makes collective action more legible.

The public-safety example supplies the other half. Flock’s automated license-plate readers photograph vehicles on public roads and make their recent movements searchable across participating police networks. CNN reported in July 2026 that officers had resigned or been arrested in at least two dozen cases involving alleged use of the system to stalk romantic interests. One Milwaukee officer secretly searched for his partner 124 times and the partner’s ex 55 times. An internal-affairs detective who investigated that case was later himself arrested for allegedly misusing the same system. Flock says abuse is rare and has introduced alerts for abnormal searches. The point is not that the system was built for stalking. It is that once the database and access path exist, illegitimate tracking becomes cheap enough to be an ordinary problem of permissions, auditing and enforcement.

Institutions can potentially identify:

  • who attended;
  • who communicated with whom;
  • who funded an organization;
  • who travelled where;
  • who works for which employer;
  • which networks connect activists, unions and journalists.

The chilling effect does not require mass arrests. It only requires uncertainty:

“This may be legal now, but the record will exist permanently.”

That increases the expected personal cost of protest, strikes, occupations, sabotage, leaks and other materially disruptive action.

The political significance is clearest in the asymmetry between institutional and popular coordination.

A small network of large nodes connected directly by thick private conduits above a much larger network whose thin transparent connections pass through shared chokepoints

Corporations and states begin as organized collective actors, with hierarchy, money, institutional memory and private channels. The public begins as separate individuals. To create countervailing power, those individuals must find one another and become visible to the platforms, employers and authorities through which they communicate. The act of constructing opposition produces a map of the opposition before it produces comparable power.

The asymmetry also runs in the other direction. A small institution can address millions of people without those people coordinating first. With superior AI, population-scale data and access to delivery channels, it can model reactions and adapt its influence continuously. The public uses shared platforms to organize; concentrated actors use the same platforms to observe and influence the public.

Surveillance therefore acts on the process by which political opposition becomes possible. It need not prohibit collective action. It makes the formation of collective power more legible and personally costly while leaving institutional power already assembled.

Capital enters politics already organized. The public must organize itself under observation.

6. How the Metered Intelligence Complex reproduces itself

Scarcity loop 

flowchart TD
    capacity["Restricted capacity"] --> prices["Higher prices"]
    prices --> margins["Higher margins"]
    margins --> discipline["Investors reward discipline"]
    discipline --> capacity

Strategic-allocation loop 

flowchart TD
    reserve["Hyperscalers reserve capacity"] --> supply["Less open-market supply"]
    supply --> competition["Smaller firms lose competitiveness"]
    competition --> capital["Hyperscalers gain revenue and capital"]
    capital --> reserve

Political-dependence loop 

flowchart TD
    race["AI framed as a geopolitical and military race"] --> funding["Public funding and procurement"]
    funding --> capacity["Private capacity expands"]
    capacity --> dependence["State dependence deepens"]
    dependence --> regulation["Regulation looks strategically dangerous"]
    regulation --> race

Surveillance loop 

flowchart TD
    pressure["Economic pressure"] --> unrest["Unrest increases"]
    unrest --> monitoring["Monitoring and public-order powers expand"]
    monitoring --> risk["Organizing becomes riskier"]
    risk --> resistance["Effective resistance weakens"]
    resistance --> extraction["Extraction continues"]
    extraction --> pressure

Influence-asymmetry loop 

flowchart TD
    concentration["Compute concentration"] --> capability["Superior AI capability"]
    capability --> channels["Public information channels become easier to shape"]
    channels --> narratives["Policies and narratives favor compute owners"]
    narratives --> protection["More capital and protection"]
    protection --> concentration

Narrative loop 

flowchart TD
    labels["Respectable names for local actions"] --> unnamed["The complete system remains unnamed"]
    unnamed --> inevitable["The outcome appears inevitable"]
    inevitable --> opposition["Opposition becomes harder to articulate"]
    opposition --> choices["The underlying choices continue"]
    choices --> labels

7. Observable predictions

This account of the Metered Intelligence Complex predicts:

  • more long-term capacity-reservation agreements;
  • expansion financed or guaranteed by major customers;
  • less speculative expansion ahead of demand;
  • consumer products with increasingly compromised specifications;
  • higher prices presented as a permanent structural reset;
  • consumer capability moving from local ownership to rented cloud access;
  • industrial subsidies without corresponding affordability conditions;
  • AI procurement and infrastructure subsidies increasingly justified through military and geopolitical competition;
  • antitrust cases focusing narrowly on explicit communication rather than structural coordination;
  • surveillance powers repeatedly introduced under narrow emergency justifications and later generalized;
  • expanding use of AI to generate, target and continuously adapt institutional messaging through mass communication platforms;
  • strong corporate investment statistics alongside declining household purchasing power;
  • “temporary” scarcity becoming the basis of a permanent business model.

The diagnosis would be weakened by:

  • aggressive capacity expansion not backed by advance contracts;
  • sustained price wars between the dominant manufacturers;
  • consumer supply growing faster than strategic demand;
  • serious structural antitrust remedies;
  • subsidies tied to open-market supply and affordability;
  • governments materially restricting surveillance of lawful political activity.

8. The sum of the vectors

The projected reservation of 40% of global DRAM capacity, a new 4 GB graphics card and a €5,600 consumer GPU do not prove a hidden agreement. They are coordinates in the same transformation. One vector concentrates physical supply. Another turns compute into productive intelligence. Another concentrates the money capable of buying it. Another makes states dependent on its private providers. Surveillance and AI-amplified influence reduce the ability of everyone else to change the arrangement.

The point is not any single vector. It is their sum. The resultant points toward the Metered Intelligence Complex: a state-corporatist system in which intelligence becomes essential infrastructure, its means of production remain private, and access is distributed as a monitored and revocable service.

A conspiracy would, in one sense, be easier to oppose. It would have members, a plan and an agreement that could be exposed or prohibited. The MIC is assembled from decisions that each institution is already authorized and rewarded to make: a capacity contract, a subsidy, a procurement rule, a pricing model, a safety policy, a promotion decision. There is no single room to raid and no master agreement to discover. Replace the people without changing the incentives and their successors will reproduce the same structure.

The absence of conspiracy does not mean the absence of choice or responsibility. Manufacturers choose how quickly to expand. Hyperscalers choose what to reserve and how to sell access. Governments choose whether subsidies purchase public rights or private dependency. Regulators choose which concentrations of power to tolerate. Platforms and police choose what to collect, retain and make searchable. A system can be emergent without being inevitable.

Nor does industrialized intelligence have to take the form of a private utility. It can be distributed through personally owned hardware, supplied as accountable public infrastructure, or governed as a cooperative or commons. But those alternatives become harder with every long-term capacity reservation, closed model, unaffordable local device and public workflow built around a private provider. Infrastructure decisions made now determine which ownership arrangements remain practical later.

The central contest of the AI era is therefore not merely who builds the most capable model. It is who owns the machines that convert compute into intelligence, who is permitted to use them, and whether the public retains enough economic and political power to choose a different arrangement.

The defining question is not whether machines become intelligent. It is whether intelligence, once industrialized, remains something people can possess—or becomes something institutions permit.