What Will Happen to $84 Trillion of American Wealth Three economic scenarios, cyber risk and the contradiction between power over people and control over the system This is the third part of a series on how modern US wealth is structured. In the first part, "The Geometry of Capital," we examined who owns the assets. In the second, "Money Is Made of Debt," we looked at how the credit system inflates the money supply and sustains the growth of asset values. Now the question is different: what might happen to this construction next? By the end of the second quarter of 2026, the total wealth of American households reached $185.58 trillion. At the start of 2020 it was $101.78 trillion. Over six and a half years, $83.8 trillion was added, that is, almost $84 trillion. This is a huge sum. But behind it lies not only the growth of production, wages and savings. A significant part of the increase is connected to the fact that the assets themselves rose in price: stocks, real estate and other claims on future income. This is exactly where the main question of the article begins. If the value of wealth has grown faster than real incomes, how sustainable are the promises on which it rests? What will happen if profits fail to meet expectations, the credit chain breaks down, or the digital infrastructure comes under attack? > Thesis. > > $84 trillion is mostly a revaluation of claims on future income, so the question is not whether the sum is large, but what backs these claims. There is also a contradiction deeper than financial calculations. The owners of large fortunes and influential institutions can set the rules by which everyone else lives. But this does not yet mean that they fully control the system on which the value of their own wealth depends. Banks, markets, cloud platforms, power grids, software and supply chains are connected so tightly that no single participant controls them entirely. Another dependency is emerging as well. Management requires reliable data and the ability to verify conclusions. AI can sharply increase the volume of analytical documents, but by itself it does not guarantee that people will become better at assessing what is happening. If ready-made explanations displace verification of the source data, the problem already affects not only the quality of management but also trust in those who make decisions. I am not going to guess the date of the next crisis. Instead, we will examine three possible paths for the American economy. Then we will look at how cyber threats can affect asset values, why the concentration of ownership does not mean full control over the system, and what will happen if society and its leaders begin to lose the ability to assess reality independently. The data in the article are given as of October 30, 2026, unless another date is indicated next to a number. ## Wealth has grown. But what exactly has grown? On September 15, 2026, Forbes published its next list of the 400 richest Americans. According to the publication's estimate, the combined fortune of this group reached $8 trillion. A year earlier it was $6.6 trillion. This figure is impressive in itself, but it is more important to understand what exactly we call wealth. GDP shows the flow of goods and services produced by a country in a year. The fortune of an owner of stocks or real estate is structured differently. It is an estimate of claims on future income: dividends, rental payments, company profits and interest on debt securities. Such claims have a price today, although the money on them may arrive over years. Therefore wealth cannot be reduced to the amount of cash in accounts. Behind it stands a large network of obligations and expectations. A company must earn a profit in order to pay dividends. A tenant pays for housing or office space, a borrower repays a debt with interest, a bank processes payments, and a registry confirms who owns the asset. As long as the whole chain works, claims on future income retain their value. This is the difference between economic growth and the growth of asset valuations. Production can increase gradually, whereas company stocks can rise sharply. If investors begin to expect that future profits will be much higher than current ones, they are willing to pay for them today. Market value grows even though the additional money has not yet been earned. From the first quarter of 2020 to the second quarter of 2026, the wealth of American households grew from $101.78 trillion to $185.58 trillion. The increase amounted to $83.8 trillion. Over the same period, accumulated household savings, if the monthly PMSAVE series is summed, amounted to about $10 trillion. The difference between the indicators is about $74 trillion. The Federal Reserve's financial accounts (Z.1) allow us to check what stands behind this difference. According to the table of changes in net assets of the "households and nonprofit organizations" sector, the increase over the same period equals $86.7 trillion. Of this, $71.9 trillion is the revaluation of existing assets (holding gains), and $15.1 trillion is accumulation through transactions: savings, purchases of assets, contributions to pension funds. The identity balances almost without remainder. In other words, most of the increase is not deferred money but a change in the price of claims on future income. Revaluation works in both directions: in 2022 it took $8 trillion away from the sector, and by mid-2026 it had returned it with a surplus. It is worth clarifying the perimeters. Z.1 gives levels of $109.2 trillion at the end of the first quarter of 2020 and $195.9 trillion at the end of the second quarter of 2026: it counts the entire sector, including nonprofit organizations. DFA (Distributional Financial Accounts), on whose data this article is built, excludes these organizations and gives levels of $101.78 and $185.58 trillion for the same dates. These are two perimeters of one accounting system; the article uses DFA throughout, because only it distributes wealth across groups. The decomposition of the increase is based on the Z.1 release of September 11, 2026. ![Fig. 1 — Two perimeters of one accounting system: DFA levels and the decomposition of the increase according to Z.1 — asset revaluation versus new savings.](/upload/upload-chto-budet-s-84-trln-fig1_wealth_en.svg) *Fig. 1 — Two perimeters of one accounting system: DFA levels and the decomposition of the increase according to Z.1 — asset revaluation versus new savings.* There is a second sign of a change in valuation. The trailing P/E of the US market, according to multpl, rose from 17.7 in 2019 to 25.86 on October 30, 2026. In other words, investors began paying noticeably more for each dollar of profit already earned. The Shiller CAPE is at approximately 41, which is substantially above the historical median of about 17.8. This does not mean that the market must fall immediately. But the price of expectations has become high, which means future profits must match these expectations. ## A pyramid with different foundations Another dimension of this story is connected to the distribution of wealth. According to DFA data for the second quarter of 2026, the top 10% of American households own 68.9% of capital. The middle 40% account for 28.8%. The bottom half accounts for 2.3%. Moreover, within the top 10%, the top 1% alone accounts for 32.5% of all capital. ![Fig. 2 — The ownership pyramid: 68.9% of capital held by the top 10%, 2.3% by the bottom half; within the top, 32.5% falls to the top 1%.](/upload/upload-chto-budet-s-84-trln-fig2_ownership_en.svg) *Fig. 2 — The ownership pyramid: 68.9% of capital held by the top 10%, 2.3% by the bottom half; within the top, 32.5% falls to the top 1%.* The main financial assets are concentrated at the top. The middle layer relies on housing, pension savings, wages and access to credit. The bottom half has almost no capital stock and depends far more on current earnings, rent, transfers and everyday payments. But this is a pyramid of ownership, not a map of full control over the economy. A large share of capital gives influence, access to resources and the ability to shape the rules. However, even the largest owner cannot single-handedly keep the entire chain on which his property depends running. Stocks exist thanks to companies and markets, settlements pass through banks and payment systems, data centers depend on electricity, networks, equipment, software and suppliers. > Clarification. > > The share of ownership measures the right to income, not the ability to keep the system running: in a failure, the first thing to lose value is what cannot be restarted alone. The top can influence the living conditions of the lower layers, but it itself depends on a system it does not fully control. The lower layers hardly own this system at all, yet they use its services every day and submit to its rules. As long as everything works, this contradiction is invisible. Failures can show that influence over people and control over the entire system are not the same thing. The same crisis will therefore hit different groups differently. What matters is not only who has how many assets, but also who depends on specific mechanisms and who has a reserve of time to survive their stoppage. There is also a less visible difference. Some can influence the rules, while others mostly adapt to them. If decisions turn out to be wrong, far from everyone has the opportunity to demand explanations or secure a review. ## Three roads for the American economy The future does not switch on according to a schedule. It is made up of several variables: the cost of money, the profitability of AI investments, the availability of credit, political decisions and energy prices. They can change simultaneously, so the scenarios below should not be taken as three mutually exclusive forecasts. They are three different mechanisms that can intersect. > Paradox. > > The longer the market waits for earnings to catch up with valuation, the larger the sum of expectations that will have to be justified — that is, the waiting itself increases the price of the subsequent revaluation. ### First road: earnings catch up with price The most favorable variant for current valuations looks like this: companies really do start earning as much as the market expects from them. Then high multiples gradually acquire a real economic foundation. What today looks like an expensive bet on the future over time turns into the value of a working business. The main test is tied to investment in artificial intelligence. Capital expenditures of the largest technology companies in 2026 are estimated at roughly $720–745 billion. The money goes to data centers, accelerators, networks, power supply and land. These investments are justified only if the new infrastructure creates sufficient cash flow: through cloud services, corporate subscriptions, advertising, software or new products. The problem is that a data center cannot be paid back with a single promise of future demand. Electricity, cooling, maintenance and financing require money right now. And computing accelerators become obsolete far faster than buildings and engineering networks. GPUs may need replacement within a few years, whereas the service life of the building itself is measured in decades. This scenario has three conditions. First, the earnings of AI-related companies must grow faster than capital expenditures for at least several quarters in a row. Second, long-term rates must stop weighing on the cost of financing. Third, energy and logistics constraints must not sharply increase the cost of building and operating data centers. Why can't this option be dismissed? Because construction has already become part of the economic and political agenda. It creates jobs, supports demand for equipment and infrastructure, and a whole chain of suppliers has grown up around it. Bringing it to a halt abruptly is not easy. But there is also a flip side. The longer the market waits for earnings to catch up with valuations any moment now, the larger the sum of expectations that will have to be justified becomes. If the catch-up does not happen, the subsequent revaluation may prove more painful. ### Second road: the market cools without a big crash In the calmer variant, the market gradually returns to moderate valuations. This does not require the economy to collapse. It is enough for investors to stop paying for future earnings as generously as they do now. As of October 30, the trailing P/E of the US market was 25.86. Imagine it falls to 22, 18 or 15.1. If company earnings do not change, a decline in the multiple alone will reduce market value by roughly 15%, 30% or 42% respectively. If earnings simultaneously fall by 10–20%, the drop will be deeper. In the calculation examples given, it ranges from 23% to 53%. ![Fig. 3 — The arithmetic of normalization: a decline in P/E from 25.86 to 22, 18 and 15.1 separately and together with a 10–20% fall in earnings.](/upload/upload-chto-budet-s-84-trln-fig3_pe_en.svg) *Fig. 3 — The arithmetic of normalization: a decline in P/E from 25.86 to 22, 18 and 15.1 separately and together with a 10–20% fall in earnings.* This is not a forecast and not an indication of a specific date. It is simple arithmetic: the price of an asset depends both on earnings and on how much the market is willing to pay for them. When both indicators go down, the losses add up. This process can be called normalization. The market does not necessarily stop believing in technology or in the American economy. It simply begins to demand a more realistic price for future income. Market capitalization declines, some projects are postponed, companies cut costs, and investors again pay attention to cash flow, not just to promises of growth. In this scenario, the first to lose are the owners of expensive stocks and other assets whose valuation depends most heavily on expectations. But the consequences will not stay confined to the stock market. If companies start economizing, this will affect contractors, employment and wages. For the middle layer, the cost of housing and mortgages matters. For the bottom half, a stable job, the ability to pay for housing and access to credit are especially important. That is precisely why a "soft landing" for the market does not mean that everything will be easy for everyone. It only means that the revaluation does not turn into a rupture of the credit chain. ### Third road: the credit loop breaks The most unpleasant economic scenario begins not with a stock decline as such, but with the fact that weak participants in the chain can no longer borrow and refinance debts. Here one needs to look especially carefully at the riskiest corporate bonds. The CCC OAS spread is understood as the additional yield that investors demand on bonds of companies with the lowest credit rating compared with Treasury securities. The higher this spread, the more the risk of default costs the market. As of October 30, the indicator reached 13.25%, and the gap between CCC and the broader high-yield segment was 9.99 percentage points. This is a level that speaks of strong distrust toward the weakest borrowers. Another risk zone is tied to private credit. This is the name for loans to companies outside the ordinary public bond market. According to Fitch, the default rate in this segment over the last 12 months to July 2026 was 6.1%. If it exceeds 8%, this will become an additional signal that the problems are ceasing to be isolated. For AI infrastructure, the danger lies in the fact that its financing increasingly resembles a closed loop. Equipment suppliers sell accelerators, financial structures help pay for construction, companies build data centers, and future revenue from leasing computing capacity must cover costs and new purchases. As long as money is available, the loop works. If creditors begin to refuse risk, construction may stop not because executives suddenly changed their minds, but because financing no longer arrives. In such a situation, the fall in share prices becomes not the cause but the consequence of a more serious problem. If projects cannot raise money, demand for equipment contracts. If data centers are not filled with tenants, the expected revenue does not arrive. If suppliers are forced to wait for payment, they themselves begin to restrict financing for buyers. The chain that was accelerating investment starts working in reverse. At the same time, another loop linking energy prices, interest rates and government debt may intensify. Expensive energy sustains inflation and hinders rate cuts. A high rate increases the cost of servicing debt. The more the budget spends on interest, the less remains for other tasks. And attempts to compensate for expenses with new borrowing again increase the debt burden. Therefore, the main signal will not be a single drop in quotations, but the coincidence of several signs. Spreads on the riskiest borrowers are rising, defaults are increasing, projects are losing access to financing, and the market is increasingly reluctant to buy debt instruments tied to infrastructure construction. ## Cyber threats: a risk that runs through the entire pyramid Three economic scenarios describe what may happen to incomes, asset prices and credit. But there is another risk that runs through all levels of the economy at once. It is the vulnerability of digital infrastructure. Modern wealth does not exist only as ownership records. Behind every such record stand banks, registries, trading venues, cloud systems and communication networks. A payment must be sent, received and confirmed. A transaction must be executed, and the right of ownership of an asset verified. If these mechanisms stop working or become doubtful, the owner may still be listed as the proprietor, but exercising that right becomes harder. Cyber threat therefore should not be treated as a fourth economic scenario. It is a separate layer of risk on top of the entire pyramid. It can manifest during growth, during an ordinary correction or in the midst of a credit crisis. And its consequences are not limited to a temporary outage: they are capable of changing investors' expectations and the value of the assets themselves. ![Fig. 4 — Cyber risk as a separate layer: it does not replace the economic scenarios but runs on top of each of them.](/upload/upload-chto-budet-s-84-trln-fig4_cyber_en.svg) *Fig. 4 — Cyber risk as a separate layer: it does not replace the economic scenarios but runs on top of each of them.* ### The base of the pyramid: old devices that no one updates The most convenient entry point is often not where the money is concentrated. It may be home and office routers, cameras, network-attached storage, VPN gateways and other devices that sit on the network for years and are barely maintained. The manufacturer has stopped issuing fixes for them, the owner does not notice the problem, and the device keeps forwarding traffic and executing commands. Such devices are valuable not necessarily in themselves. They can become an intermediate link that hides the source of an attack, opens access to an internal network or passes someone else's traffic through itself. That is why a large botnet of routers and IoT devices is not just a collection of infected gadgets. It is a distributed staging ground that can be used against someone else's infrastructure. The history of such operations is already long. Mirai in 2016 turned household devices into an instrument of large-scale attacks. VPNFilter in 2018 infected, by Cisco Talos's estimate, about half a million routers and network-attached storage devices in 54 countries. In the case of Volt Typhoon, compromised routers were used as a proxy layer that concealed the group's traffic. In February 2024, CISA and partners issued a joint advisory about this infrastructure. In April 2026, the US Department of Justice and the FBI reported on Operation Masquerade: according to the agencies, US specialists, acting under a court order, reset the settings of thousands of compromised TP-Link SOHO routers. The authorities linked the campaign to a GRU unit. It is important to separate the fact of the operation from the attribution. The link to a specific state structure reflects the position of the US agencies. A single paragraph of an article is not enough to independently prove such a link. The M-Trends 2026 report describes campaigns against network devices and the use of the BRICKSTORM backdoor. The report speaks of intruders establishing a foothold on network perimeter devices and maintaining a long presence in the infrastructure. This is a particularly unpleasant class of threat: security tools installed on ordinary computers may simply not see what is happening inside a router or VPN gateway. For the lower part of the economic pyramid, the consequences may look like an internet outage, a banking app failure or a delivery service failure. A business may lose access to the network, data and corporate systems. The financial system risks facing compromise of the channels through which data and commands pass. An attack may start with a cheap device, but its consequences will affect systems far beyond the home or organization that owns it. ### The next layer: AI accelerates the search for weak spots With the advent of language models, finding vulnerabilities is becoming more accessible and faster. This does not mean that AI already independently breaks into any system. But it helps to parse code, look for errors, write auxiliary tools and turn technical knowledge into a more accessible service. Here a race of two speeds emerges. Defense must find the problem, verify it, prepare a fix and deliver it to the user. In practice, months may pass between the discovery of a vulnerability and the update of a device. Especially if the device is old, the manufacturer releases firmware slowly, and the telecom operator or corporate administrator must separately approve the update. According to GrapheneOS, some Android fixes reach public bulletins with a delay of several months after being handed to manufacturers. This is a statement by the project's developers, and it must be read precisely as the position of a participant in the ecosystem, not as universal statistics for every Android device. But the problem itself has been familiar for a long time: the existence of a fix at the developer does not yet mean that it has reached every device. There is also the opposite example. In Chrome versions 149–151, about 1,500 vulnerabilities were closed, and, according to Google, 94% of the findings were made by internal AI agents. When a single vendor controls the browser, the release of updates and the delivery channel, the defensive model works much better. This is an important clarification: AI accelerates not only the work of attackers but also that of defenders. The result depends on who turns a finding into a fix faster and whether they can deliver it to all users. The most vulnerable position is held by devices that have already been taken out of support. For them there may be no new fix at all. A study published in IEEE Transactions on Dependable and Secure Computing in 2024 estimated that millions of active devices remain on the network beyond support; a significant share of known vulnerabilities in them are discovered after the end of updates. These devices do not disappear from the internet when support ends. They keep working and over time turn into long-lived points of risk. ### AI agents: a new door into other people's systems There is another problem not directly related to router infection. An AI agent may gain access to email, documents, the browser, corporate databases, and in some cases also to payments and trading operations. It reads external information and makes decisions based on it. This means an intruder does not always need to break into the operating system. In some cases it is enough to influence the content that the agent will read. This class of attacks is known as indirect prompt injection. An instruction may be hidden in a web page, an email, a document or other content that the agent treats as an ordinary source of information. If the system does not separate data from commands well enough, external text can change the agent's behavior. The mechanics have been studied since 2023. In July 2025, a Nikkei investigation described cases of hidden instructions in scientific preprints on arXiv: the text was designed for AI systems that could read and evaluate documents. This is not yet a story about stolen money. But it shows that instructions for machines are already being placed inside ordinary content, where a person may not even notice them. Financial consequences will appear where agents are given real authority. While an agent only retells an email, the consequences of a single error are limited. If it can send payments, change trading orders, sign documents or communicate with counterparties on behalf of the company, the price of an error becomes entirely different. So the question is not whether AI will become "evil" or acquire a will of its own. The risk arises in a far more down-to-earth situation. An agent has been given useful work and access to the necessary systems, but no reliable boundary has been established between the text it reads and the actions it is actually permitted to take. An external instruction can turn into a command, and then into an operation with real consequences. ### AI malware: from individual techniques to a closed loop Malicious programs have already turned to language models during real operations. In July 2025, Ukraine's CERT-UA described LAMEHUG, also known as PromptSteal. The malware contacted a language model through an open API and used the commands it received to collect information and copy documents from infected computers. Google Threat Intelligence later also classified this case as one observed in real operations. The M-Trends 2026 report describes QUIETVAULT: the malware checked infected machines for the presence of local AI command-line tools and ran prepared prompts, including to search for configuration files. This shows that attackers can use AI tools already installed on the computer. But it is important not to jump past the evidence. PromptFlux, which used a model to create new disguised variants of its own code, was an experiment. PromptLock was a research prototype of ransomware. It had no confirmed victims. Academic papers have also shown that instructions can self-propagate through email or interactions between agents. However, a laboratory demonstration does not yet mean that such infection has become a mass phenomenon on the internet. For now we see individual parts of a future construction, not a fully autonomous AI swarm. For such a system to become truly dangerous, it must simultaneously find vulnerabilities, exploit them, adapt to defenses and spread further without constant human involvement. A full cycle of this kind had not been documented by October 2026. This is precisely where the boundary lies between what has already happened and what so far remains a scenario. The use of a model to generate commands has already been recorded in a real attack. The existence of a self-modifying worm that independently finds a target, penetrates it, changes its own code and spreads further has not yet been confirmed. ## How cyber threats reduce the value of wealth To understand the financial consequences, it is not necessary to imagine a complete shutdown of the banking system. The value of assets can change long before such an event. It is enough for investors, creditors and insurers to begin to consider future losses more likely. > Turn. > > Losses are counted not by stolen sums but by the price of risk: repricing is triggered before the attack itself occurs. This process has two main paths. The first is related to income. A hack, a data leak or a halt of digital systems requires money for recovery, verification of operations, compensation and protection. If a company cannot operate for several days, it loses revenue. If the problem affects its suppliers or customers, the damage spreads further. Expected profit becomes lower, and along with it the company's valuation may decline. The second path runs through the price of risk. Even if a major attack has not yet occurred, investors may demand a higher return for holding an asset they consider vulnerable. When future income looks less reliable, they are willing to pay less for it. Thus market value can decline without an immediate drop in current profit. This is not an automatic reaction to every piece of news about a hack. What matters for repricing is the scale of the threat, the recurrence of incidents, the quality of protection and whether the problem affects shared infrastructure. Artificial intelligence can amplify this mechanism. It helps to parse code faster, look for errors and adapt tools to a specific target. This does not mean that AI can already independently break into any system. But if preparing an attack becomes cheaper while fixes still take a long time to reach devices, companies and investors are forced to account for a wider range of possible losses. For owners of large fortunes, this means the risk of a decline in the value of stocks, bonds and stakes in companies. Part of the losses may arise from real damage, and part from the market valuing future income differently and demanding an additional premium for risk. The more the financial system depends on a few shared digital platforms and vendors, the broader the repricing may be. The middle 40% will feel the consequences through employers, pension savings, mortgages and access to credit. If a company spends more on protection and recovery or loses the ability to operate, this may affect hiring and wages. If financial assets fall in value, the value of pension savings invested in the market also decreases. The lower half owns only a small share of capital but depends more heavily on the continuous operation of payments, transfers and basic services. Here the damage may be not so much in the repricing of property as in the loss of access to money and services precisely when they are needed. For a person with no savings, even a short delay in wages carries a different weight than the temporary unavailability of part of an investment portfolio. Therefore the consequences cannot be measured only by the amount of stolen money. A cyberattack can reduce the value of wealth through lost profit, additional costs and a rising risk premium. At the same time, it can disrupt the daily life of people who have almost no financial cushion. These are different types of losses, but they arise within the same pyramid. ## Duality: power over people and dependence on the system Now we can return to the contradiction that runs through the entire article. The upper layers of society hold most of the capital and can influence the rules of access to money, work, housing and services. But this does not make them masters of the entire infrastructure. They depend on banks, exchanges, cloud platforms, energy, communications, software, suppliers and state institutions. These links are connected to one another, and no single participant can guarantee the outcome of the whole system. At the same time, the lower layers depend on this system in almost everything, but hardly own it. They have to use banking apps, employer platforms, credit services, identification systems and government portals. The terms are often set without their participation. A refusal by an algorithm or a service failure can deprive a person of access to money or a service, and there is often no one to turn to for an explanation. This is how the duality arises. The top can influence people, but cannot fully control the system on which it itself depends. The lower layers have almost no influence on how the system is structured, yet they experience its decisions in everyday life. Power and responsibility diverge: some have more opportunities to set the rules, others have fewer opportunities to refuse those rules, but no one has full control over the entire chain. It is important not to turn this thought into the claim that the system has a will of its own. The system is not a separate subject that has decided to subjugate the owners of capital. It is a combination of institutions, algorithms, infrastructures and obligations. Together they create constraints that no single executive or owner can override by decree. A company can control its own platform, but not the entire network, power supply, software and financial settlements on which it depends. As long as the infrastructure works, this contradiction stays in the shadows. But if failures become noticeable and repeat, people may see that those who demand compliance with the rules are themselves unable to ensure the reliable operation of the system. It is especially dangerous for trust if the losses are distributed unevenly: large institutions are helped to recover, while workers, customers and taxpayers bear the consequences through lost income, higher prices or new budget expenditures. This is not an inevitable political crisis. Between a technical failure and a loss of public trust there are many intermediate conditions. What matters is the duration of the problem, the transparency of decisions, the availability of compensation and who ultimately pays for the recovery. But if people increasingly see a gap between the promise of control and the real ability to manage, the legitimacy of institutions may weaken. AI can accelerate this process. It increases the speed of digital systems, helps automate decisions and at the same time lowers the cost of some types of attacks. The more important actions are delegated to programs and agents, the harder it is for a person to understand why the system made a particular decision and who is responsible for the error. So the question is not that machines have suddenly gained power. The question is that society may become increasingly dependent on mechanisms it can neither fully control nor clearly contest. Cyber risk can reinforce each of the three economic roads. In the favorable scenario, security spending and downtime can eat into part of the profit that AI investments are supposed to bring. Under ordinary market normalization, rising perceived risk can further lower asset valuations, even if no major attack has occurred. In a rupture of the credit loop, a cyber incident can accelerate problems if it affects a supplier, a bank or a digital platform on which financing depends. Therefore cyber threats do not form a separate fourth scenario. They can amplify the consequences of any of the three. Financial revaluation and a crisis of trust are connected. When investors doubt future earnings, the value of assets falls. When people doubt the ability of institutions to keep the system running, what is at stake is no longer only valuations but also public consent to live by the established rules. ## When conclusions outnumber verifications Modern governance grew up around measurements. States collect statistics, companies track production and sales, financial institutions assess risks, researchers test hypotheses. An indicator by itself explains nothing. It must be compared with other data, its methodology checked, what remains outside the measurement understood, and only then can conclusions be drawn. AI changes the economics of this work. It can quickly summarize documents, cross-reference large data sets, prepare forecasts and write analytical memos. The same set of materials can be turned into hundreds of versions of a report for different agencies, executives and audiences. Producing text becomes cheap. Verifying how well conclusions match the source data does not automatically become cheap. That is where the risk lies. The number of analytical documents can grow faster than the number of independent checks. Several reports may look like different studies, even though they all rely on one source or repeat one erroneous premise. A persuasive text begins to be perceived as evidence that a full analysis has already been carried out. But the detail of a document, the number of tables and a confident tone say nothing about the quality of the evidence. > Turn. > > Cheap text does not make knowledge cheap: a thousand reports from one source yield one independent confirmation, no matter how many tables they contain. The problem is not that AI necessarily produces falsehood. It can speed up genuine research work and help detect errors. The risk arises when an organization stops maintaining its own ability to verify conclusions and simply starts consuming them. An executive is handed a ready-made picture, but does not always see which data went into it, which explanations were discarded, and where measured facts end and assumptions begin. This can produce a strange situation: there is more information, reports come out faster, but understanding of reality does not improve. A person literally receives more and more ready-made conclusions, but less and less often goes through the path from observation to independent assessment. ### The double loss of the ability to assess reality At the bottom of the pyramid this can happen almost unnoticed. Most people are not obliged to verify state statistics daily, understand monetary policy or study how banking systems work. In ordinary life a person delegates some of these tasks to institutions, specialists and the media. As long as services work, salaries arrive, payments go through and the usual order holds, there is little reason to spend time checking every decision. This is not a lack of intelligence and not the fault of the ordinary person. That is how the division of labor works. The problem begins when the habit of trusting turns into an inability to verify, and a ready-made explanation is accepted only because it matches previous beliefs or comes from "one's own." At the top a similar but not identical dependence arises. Political and business leaders receive more and more condensed briefings, forecasts and recommendations. First, complex work is done by advisers and analytical units, then more and more intermediate operations are handed over to algorithms and models. A leader may retain the right to the final word but gradually lose the habit of independently checking the source data and considering alternative explanations. This is how a double vulnerability appears. Society gets used to receiving ready-made interpretations from institutions, and the institutions themselves become increasingly dependent on ready-made interpretations prepared by analytical systems. At the bottom a person may not see the grounds for a decision. At the top a leader may see only the result of processing, without understanding what was left off-screen. Formal powers are retained, but the ability to assess the reality on which the decision is based may weaken. This does not mean that any use of AI leads to a loss of independence. A model can highlight an inconsistency, check a calculation and suggest an alternative. The question is whether independent verification is preserved. Who is responsible for the quality of the source data? Can the course of the analysis be reconstructed? Does the leader see the arguments against the proposed decision? Does a person have the opportunity to contest an automated conclusion? ### When words lose weight There is an old parable about a boy who several times shouted that a wolf had attacked the flock. When the danger actually appeared, the villagers no longer believed him. He had abused their trust so often that he deprived his own warnings of force. In modern politics the mechanism may look similar. A leader announces that the situation is improving, then encounters facts that contradict this, and offers a new explanation. Changing one's position is not shameful in itself: circumstances really do change, and new data require a revision of previous conclusions. The problem arises when the earlier statement is not acknowledged as mistaken, the contradiction is not explained, and instead of verification another version is offered that is supposed to preserve trust in the speaker. Over time the audience stops evaluating the arguments themselves. Some accept the explanation because they trust their political camp. Others reject it for the same reason, regardless of the quality of the evidence. The shared space of verification narrows. As a result, what is lost is not only trust in a particular politician. It becomes harder for people to understand when they are facing another manipulation and when a truthful warning has actually been sounded. AI can accelerate the production of such explanations. For a single event it can quickly prepare many versions, equip them with tables, quotations and links, and adapt them to different audience groups. But a thousand texts do not become a thousand independent confirmations if they repeat one original version. The persuasiveness of the presentation does not replace verification of the data's origin. That is why the parable of the shepherd matters here not as a comparison with an individual politician but as a description of the depletion of trust. When words often diverge from observable events, society risks ceasing to believe even those messages that correspond to reality. For the state this is dangerous: a warning about a real threat may get lost among yet more explanations and justifications. ### When trust is replaced by regulations Informal ties are also part of the structure of society. Reputation, mutual obligations, the habit of honoring agreements and the ability to resolve disputes without constantly turning to the state allow people to interact without a separate instruction for every case. Such relations are not always beneficial: they can conceal cronyism, closed-door arrangements and exceptions for "one's own." But they show that public life does not rest on laws alone. When trust weakens, people begin to demand formal guarantees. What was once settled by agreement is formalized in documents. New checks, approvals and reporting requirements appear. In some cases this is necessary: the law protects rights, limits arbitrariness and gives a person the opportunity to contest a decision. But formalization by itself does not create trust. If the rules become excessively complex and their application depends on broad discretion of an official or intermediary, new points for pressure and corrupt rents appear. A person has to look not only for a way to meet the requirement but also for someone who can speed up the procedure, explain an exception or secure the needed decision. Informal ties do not disappear. They may return around the bureaucratic apparatus, but now access to them determines who can get through the system faster than others. Therefore not every new law strengthens corruption, just as not every informal arrangement is useful. What matters is the simplicity and clarity of the rules, equal application of requirements, transparency of decisions, the possibility of appeal and independent oversight. Automation does not resolve the question by itself either. AI can process applications faster and find inconsistencies, but it can also scale up opaque criteria if no one checks how they are set and what consequences they lead to. This is how a vicious circle forms. Distrust generates demand for additional rules. Complex rules create new administrative barriers. If the barriers distribute opportunities unevenly, trust falls even further. The state gets more procedures and reports, but not necessarily more ability to solve the underlying problem. ## Who sets the picture of reality Out of these processes grows the question of power. Previously it was discussed primarily through property, positions and the right to make decisions. Now that is not enough. What also matters is who defines the indicators, chooses the data sources, sets the criteria of analysis and decides which conclusions the leader will see. One can govern not only by direct order. One can influence which options even reach the table, which risks the system flags as significant and which alternatives appear too costly or unrealistic. This does not necessarily mean that the developer of the model secretly governs the state. Influence can arise from ordinary dependence on an infrastructure supplier, a closed methodology, a limited dataset or a system that the client is unable to verify independently. In twenty years formal institutions may still be called parliament, government, court and corporation. But the real question will be who controls the conditions under which decisions are prepared and who is able to verify their foundations. Who determines what counts as success? Who chooses the metrics? Can the primary data be seen? Who has the right to challenge the algorithm's conclusion? Who is accountable if the decision turns out to be wrong? AI will not necessarily become an independent political ruler. The more down-to-earth and therefore more important problem is different: people may retain the right to make decisions but lose the ability to independently assess the reality on which those decisions are based. Then power will remain formally human, yet an ever larger part of its practical content will depend on systems, methods and data to which access is distributed unevenly. ## What can be done so that the pyramid does not depend on a single mistake Protection here does not come down to buying a more powerful antivirus. What is needed are measures that retain their meaning even when the attacker uses an unknown vulnerability. First, networks need to be segmented. If one router, server or workstation is compromised, this should not automatically open access for the intruder to all other systems. Second, the authority of AI agents must be limited. An agent should not be given more rights than are needed for a specific task. A large payment, a change of bank details, the signing of a document or a trading operation should not go through without separate confirmation. Third, action logs must be stored in such a way that they cannot be quietly rewritten after the fact. What matters is not only the recording of an event but also independent verification of what exactly happened. And finally, organizations will have to account for the entire fleet of devices, not only employees' computers. Routers, cameras, network storage and other peripheral devices are also part of the infrastructure. If the manufacturer no longer issues fixes for them, one needs to decide in advance what to replace them with. Updates, antiviruses and signatures are still necessary. But they depend on knowledge of a specific threat: first the problem has to be detected, then a fix prepared and delivered. Segmentation, limiting authority, confirmation of critical operations and independent logging work differently. They are not obliged to know in advance which particular method the attacker will use. There is a similar principle for analytical systems. An important decision cannot be justified by a single generated conclusion. References to the source data, the calculation methodology and the course of verification must be preserved. In the report, an observed fact should be separated from interpretation and recommendation. The leader should be shown not only the arguments in favor of the chosen decision but also the strongest objections against it. If a decision affects human rights, a clear review procedure must remain, and responsibility cannot be shifted onto the model or the software supplier. Such requirements do not eliminate errors and do not restore trust automatically. They at least preserve the possibility of establishing on what grounds a decision was made, who checked it and where exactly the failure occurred. Without this feedback loop, automation accelerates not only useful work but also the spread of errors. ## What to watch To understand which road the economy is taking, a few groups of indicators are enough, not dozens of scattered news items. For trust and the quality of public analysis there is still no single series comparable to stock indices or credit spreads. Therefore one should not invent an exact numerical threshold for them. Here it is more important to observe whether agencies publish source data and methodologies, whether an automated decision can be challenged, whether publicly refuted statements are corrected and whether independent verification of analytical systems exists. | What to watch | Why it is needed | Signal worth paying attention to | | ------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- | | Long-term rates, especially the yield on 30-year Treasury bonds | Show the price of long-term financing | A sustained move above 5.5% | | Profits of the largest technology companies and their capital expenditures | Show whether the construction of AI infrastructure pays off | AI-chain profits grow more slowly than capex | | CCC spreads and private credit defaults | Show how quickly the riskiest borrowers are weakening | CCC−HY above 10.18 pp; private credit defaults above 8% | | Brent, TTF gas, tanker freight and traffic through Hormuz | Show whether physical inflation is returning through energy and logistics | A prolonged rise in energy prices and shipping costs | | Linux kernel releases, Android bulletins and the CISA KEV catalog | Help assess the speed of vulnerability fixes and the scale of problems on the network perimeter | Vulnerabilities are actively exploited while fixes take long to reach devices | | Documented cases of prompt injections and agent actions | Show how close AI agents have come to real financial operations | A confirmed case of a payment or trade on an external instruction | | Company disclosures about downtime, losses and recovery costs, cyber insurance terms | Show whether cyber risk is turning into a permanent expense line and whether it affects business valuation | Recurring incidents, rising costs or worsening insurance terms; there is no single numerical threshold yet | These indicators do not give a ready answer to the question of what will happen tomorrow. They help to understand which part of the structure is becoming weaker. A high rate in itself does not mean a crisis. A fall in stocks does not prove that investment in AI has failed. The discovery of a vulnerability does not mean it has already been exploited. And a rise in cybersecurity spending does not yet prove that investors have already revised the value of the company. For such an assessment one needs to compare incidents, financial reports, insurance terms and market reaction. What matters are combinations of signals and how they change over time. ## Who pays | Scenario | Top 10% | Middle 40% | Bottom 50% | | --------------------------------- | -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------- | -------------------------------------------------------------------------------- | | Earnings catch up with price | Receive confirmation of high asset valuations | May gain from employment and income growth | May get new jobs, but not necessarily capital growth | | Market normalization | First to lose part of portfolio value | Feel pressure through housing, pension savings and employment | Depend more heavily on work, rent and credit availability | | Credit loop breaks | Lose asset value and income from them | Face reduced employment and lending | Feel job loss and worsening credit conditions before others | | Cyber risk on top of any scenario | Possible repricing of stocks and debt securities, higher spending on protection and recovery | Pension savings, jobs, mortgages and access to credit may be hit | The fastest consequences relate to wages, transfers and access to basic services | Capital shares help show the difference in scale and nature of losses. The top 10% own 68.9% of wealth, so even a small percentage decline in the value of their portfolios can mean enormous sums. The bottom 50% own 2.3%, but for them the loss of part of current income or a delayed payment can be far more painful in everyday life. The same failure therefore looks different. For the owner of a large portfolio it may be a drop in asset valuations. For a family with a stable salary and a mortgage it is the risk of job loss or more expensive credit. For a person without savings, even a short halt in transfers can disrupt the entire monthly budget. The pyramid distributes not only wealth but also the ability to survive a failure. There is also a social consequence. If people see that large owners and institutions can influence the rules but are unable to keep the system working, and that those with the fewest reserves pay for failures first, distrust can spread from a specific company to institutions as a whole. This is not an automatic result of an attack, but one of the paths by which a technical problem becomes political. ## Seven bets with the future So that a conversation about the future does not turn into a set of elegant assumptions, let us fix seven statements that can be tested. These are not point market forecasts but control questions for our picture of the world. > Bottom line. > > The point of a bet is that refutation must be possible in advance: a statement that cannot fail says nothing about the future. | No. | What we test | Series or source | Check date | What refutation will mean | | --- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------- | ----------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------ | | 1 | Trailing P/E of the US market will leave the 20–30 corridor | multpl, monthly values | 31.12.2027 | The indicator will stay inside the corridor | | 2 | Private credit default rate will exceed 8% over the trailing 12 months | Fitch PCDR | Q4 2027 report | The indicator will be no higher than 8% | | 3 | Hyperscaler capital expenditure in 2027 will grow by less than 10% year on year | Company reports | January–February 2028 | Growth will be 10% or more | | 4 | A US regulator or law enforcement agency will publish a case in which an AI agent is a substantial participant in the scheme and performs real actions, for example placing orders, making payments, signing documents or sending messages to third parties | SEC, CFTC, FINRA, US prosecutors | 31.12.2027 | No such case will exist by that date | | 5 | The ratio of household wealth to annual nominal GDP will leave the 5.0–6.0 range | Z.1 and DFA | Q4 2028 data, publication approximately in March 2029 | The indicator will stay inside the range | | 6 | The average time from vulnerability disclosure to mass exploitation in Mandiant reports for 2026 and 2027 will be below −7 days | M-Trends, annual reports | M-Trends 2028 publication | At least one year will be no lower than −7 days, or the methodology will change without a comparable recalculation | | 7 | A CERT, regulator or security vendor will publicly link the compromise of a mass fleet of routers or IoT to vulnerabilities found with the help of AI | CISA, CERT, vendor reports | 31.12.2027 | There will be no confirmed case by that date | Each bet has a weak spot. For example, the time from vulnerability disclosure to mass exploitation depends on the counting methodology and the sample composition. Therefore a change in methodology cannot be passed off as confirmation or refutation of the thesis itself. The same applies in the financial part: one quarter does not always show a sustained reversal. ## Finale: who is capable of assessing reality and managing the system? Over six-plus years, US wealth grew by almost $84 trillion. Behind this sum are houses, stocks, pension savings, bonds and other assets. But the value of each of them depends not only on future income. It also depends on whether companies can operate, banks can settle payments, registries can confirm rights, and people and organizations can trust the information on which they base decisions. > Conclusion. > > The point is not what will happen to the market, but whether society will retain the ability to verify what is actually happening. The economic part of this story has three possible continuations. Earnings may catch up with high valuations. The market may cool without a major crisis. Or the credit loop may break and drag down investment, employment and financing. Cyber threats run on top of all three scenarios. They can reduce earnings through downtime and recovery costs, and even before a major attack force investors to price future risks differently. But the resilience of wealth depends not only on digital infrastructure. It also depends on society's ability to verify what it is told about what is happening. Modern governance relies on indicators, statistics and analysis. AI can make the preparation of reports almost instantaneous, but it does not remove the need to verify sources and methodology. If, instead of independent analysis, leaders receive a stream of ready-made conclusions, the number of documents grows while the connection between decisions and measurable reality may weaken. This may affect both sides of the pyramid. Most people delegate the assessment of complex processes to institutions because they cannot independently verify everything happening around them. Leaders, in turn, depend ever more heavily on analysts, algorithms and models that prepare a picture of the world for them. At the bottom, a person may not see the grounds for a decision. At the top, a leader may not notice what assumptions and limitations are hidden behind the final recommendation. Formal authority remains, but the ability to independently assess the consequences of decisions is not guaranteed. When public explanations constantly change without acknowledgment of errors and without checking contradictions, trust gradually erodes. When trust disappears, society demands new guarantees and rules. Laws are necessary, but excessively complex procedures and broad discretion in their application create additional administrative barriers and opportunities for corrupt rents. AI can speed up the work of such an apparatus, but the speed of processing documents is not the same as fairness and quality of decisions. This is how a broader circle of mutual dependencies arises. Owners of capital influence society but depend on infrastructure and markets they cannot fully control. Society depends on the decisions of institutions but often lacks the means to verify their grounds. The state tries to maintain order with ever more complex rules, yet the procedures themselves can become a source of new conflicts and distrust. None of these links means that the system has acquired a will of its own. The point is that participants depend on many mechanisms that no one fully controls. Hence the question about the next twenty years. Who will govern society if political leaders depend ever more heavily on AI, citizens receive ready-made interpretations instead of verifiable explanations, and bureaucratic systems automate rules that people themselves are not always able to contest? What matters is not only who holds an office or signs a decree. What matters is who defines the indicators, chooses the data, sets the criteria of analysis, verifies the conclusions and bears responsibility for the consequences. AI will not necessarily become an independent ruler. Far more significant is the risk that people will retain the formal right to make decisions but gradually lose the ability to independently assess the reality on which those decisions are based. This is not a predetermined outcome. It can be contained if access to primary data, independent verification, transparent methodology, the ability to contest an automated decision and the responsibility of specific people for its application are preserved. American wealth rests not only on property and debt. It rests on working infrastructure, trust and the ability to distinguish measured facts from convincing explanations. Therefore the question of $84 trillion is ultimately broader than the question of the stock price. It concerns who is capable of keeping the system working, who can verify its decisions, and whether society will retain the ability to independently judge what is actually happening. [← Part one: “The Geometry of Capital”](/en/article/middle-class-gone-en) [← Part two: “Money Is Made of Debt”](/en/article/money-made-of-debt-en) ## Sources 1. Financial Accounts of the United States (Z.1), release of September 11, 2026, including the decomposition of the change in net assets of the household and nonprofit organizations sector, and the Distributional Financial Accounts — Federal Reserve: federalreserve.gov/releases/z1; federalreserve.gov/releases/z1/dataviz/dfa/ 2. FRED series (Federal Reserve Bank of St. Louis): DFA shares (WFRBST01134 and related), PMSAVE, TNWBSHNO, CMDEBT, DGS2, DGS10, DGS30, CPIAUCSL, CPILFESL, BAMLH0A3HYC, BAMLH0A0HYM2: fred.stlouisfed.org 3. Trailing P/E and Shiller CAPE: multpl.com (S&P 500 PE Ratio; Shiller PE Ratio) 4. Forbes 400, publication of September 15, 2026 (data as of September 4): forbes.com/forbes-400/ 5. FOMC meeting of October 27–28, 2026: statement and voting results: federalreserve.gov 6. Private credit defaults over 12 months to July 2026: Fitch Ratings 7. Incidents involving network devices: Mirai (2016); VPNFilter — Cisco Talos (2018); Volt Typhoon — joint advisory by CISA and partners (February 2024); Operation Masquerade — US Department of Justice and FBI (April 2026; attribution given as the agencies' position) 8. M-Trends 2026: campaigns against network devices, BRICKSTORM backdoor, QUIETVAULT malware, time from vulnerability disclosure to mass exploitation — Mandiant / Google Cloud 9. LAMEHUG (PromptSteal), July 2025: CERT-UA; attribution to real-world operations: Google Threat Intelligence 10. Hidden instructions for AI systems in arXiv preprints: Nikkei, July 2025 11. Delays in delivery of Android security bulletins: GrapheneOS (October 2026); Chrome vulnerabilities in versions 149–151: Google 12. Study of devices beyond support: IEEE Transactions on Dependable and Secure Computing (2024) 13. Known Exploited Vulnerabilities Catalog: CISA; Linux kernel releases: kernel.org