The AI Bubble, the Debt Beneath It and the Next Federal Reserve Rescue

By Matthew Parish, Associate Editor

Sunday 16 August 2026

There are two quite different propositions that are frequently confused in discussions of artificial intelligence. The first is that AI is an extraordinary technology, likely to alter substantially the way human beings work, communicate, research, write software and organise businesses. The second is that the companies presently building AI infrastructure are therefore worth the enormous sums attributed to them by financial markets. The first proposition may be true without the second being true at all.

That distinction lies at the heart of a striking paper published by Myrmikan Research on 14 August 2026, entitled AI Debt Failure Will Prompt Another Wave of Fed Bailouts. Its argument is considerably more sophisticated than the familiar claim that AI shares are in a bubble. Indeed its central concern is not principally with shares. Myrmikan argues that beneath the spectacular valuations of AI laboratories, semiconductor manufacturers, cloud-computing companies and data centres there has developed an increasingly elaborate structure of debt. If anticipated productivity gains from AI arrive too slowly to service that debt, the consequences may escape the technology sector altogether and enter insurance companies, pension funds and ultimately the balance sheet of the American state.

The paper is written from an unmistakably Austrian-school perspective. Its starting point is monetary history. Before fiat currencies, it argues, gold deposits and withdrawals communicated information about the appropriate price of credit. Central banking replaced this dispersed price-discovery mechanism with administrative judgement. From that moment central bankers were required to decide whether money was too expensive or too cheap — and political and institutional incentives repeatedly encouraged them to err on the side of cheapness. The Great Depression, Arthur Burns’s inflationary 1970s, the rescue of Continental Illinois and Alan Greenspan’s response to the 1987 crash are presented as iterations of essentially the same phenomenon. Once sufficiently large financial structures become dependent upon cheap credit, monetary authorities cannot permit them to liquidate without endangering the financial system.

This historical argument supplies the architecture for everything that follows. Myrmikan’s contention is that the Federal Reserve has repeatedly created a ratchet. Credit expansion generates investment and rising asset prices; rising asset prices make the policy appear successful; leverage accumulates; and eventually raising rates to suppress inflation threatens the institutions that have become dependent upon cheap money. The central bank then retreats. The famous “Greenspan put” was merely the most explicit version of the arrangement.

The productivity problem

The paper’s most interesting argument concerns productivity. It draws an analogy between contemporary enthusiasm for AI and Alan Greenspan’s enthusiasm for computers around the turn of the millennium. Greenspan believed technological progress was increasing productive capacity sufficiently rapidly that monetary expansion could be accommodated without consumer-price inflation. Myrmikan invokes Robert Solow’s famous observation that the computer age could be seen everywhere except in the productivity statistics, and argues that the eventual productivity improvements associated with computing proved slower and less spectacular than enthusiasts expected.

This is important because there is an enormous difference between technological capability and measured economic productivity. A machine can perform astonishing tasks without making the economy proportionately richer. Implementation costs matter. Organisational disruption matters. Training matters. Errors matter. Regulation matters. Above all, the price of the technology matters.

Myrmikan contends that precisely this difficulty is appearing with generative AI. It points to reports of unexpectedly high corporate AI expenditure, failures of implementation and surveys suggesting that many enterprises have yet to obtain measurable returns from generative-AI investment. Agentic AI makes the problem more acute because an autonomous system undertaking a task may consume vastly greater computational resources than a system merely answering a question — while mistakes by an agent can have financial consequences rather than merely producing an incorrect paragraph.

There is a genuine insight here. Much public discussion assumes that because models are improving rapidly, productivity must improve at approximately the same rate. There is no economic law requiring this. The transition from capability to productivity involves redesigning institutions around the capability. Electricity transformed manufacturing, but factories had to be rebuilt around electric motors before the gains became overwhelming. Computers eventually transformed offices, but businesses spent decades purchasing software that was expensive, badly integrated and sometimes counterproductive. AI may ultimately prove still more important — but “ultimately” is an uncomfortable word when one has borrowed money repayable next Tuesday.

Nevertheless this part of the Myrmikan argument should not be accepted too readily. Solow’s paradox was a paradox precisely because computing eventually did generate enormous economic changes whose value conventional productivity statistics sometimes struggled to capture. AI may similarly create consumer surplus and improvements in quality that national accounting measures imperfectly. A lawyer who conducts three hours of research in twenty minutes has experienced a real productivity improvement even if billing practices, organisational inertia or GDP statistics disguise part of it.

The correct conclusion is therefore narrower than Myrmikan’s rhetoric occasionally suggests. There is not yet compelling evidence that AI will not generate the productivity revolution promised for it. There is considerable reason to doubt whether it will generate that revolution quickly enough to validate every investment presently being made in anticipation of it.

That is a much more dangerous proposition for investors.

Utility is not value

Myrmikan makes this distinction particularly effectively towards its conclusion. AI can be revolutionary and AI investments can nevertheless be disastrous. The paper compares the present boom with canals, railways and the internet. Railways remained useful after railway shareholders were ruined. Fibre-optic cable remained useful after telecommunications companies collapsed. Web browsers became indispensable precisely while becoming virtually impossible to monetise directly. The destruction of scarcity can increase social utility while destroying the economic rents upon which investors had relied.

This is perhaps the strongest conceptual point in the paper.

Technological investors frequently ask the wrong question: will people use this technology? The economically relevant question is: who will capture the surplus generated by its use?

AI unquestionably has utility. But competition between models is intense. Chinese developers and open-source models create downward pressure upon inference prices. Algorithms become more efficient. Hardware improves. Knowledge diffuses between laboratories. A model capability that costs hundreds of millions of dollars to develop may become a commodity surprisingly quickly.

That creates an uncomfortable possibility. AI might become ubiquitous precisely because AI itself becomes cheap.

If so, the winners might ultimately be businesses and consumers using abundant intelligence rather than the companies that borrowed extraordinary sums to manufacture it.

The circular AI economy

The paper then turns from technology to corporate finance and becomes considerably more alarming. It portrays the contemporary AI industry not as a conventional supply chain but as a network in which suppliers, customers, financiers and investors increasingly finance one another.

AI laboratories need computing power. Hyperscalers need AI laboratories as customers. Specialist “neocloud” data-centre businesses need hyperscalers and laboratories to sign long-term purchasing commitments. Nvidia needs all of them to keep buying GPUs. Capital therefore moves around the system in circles.

The paper gives the example of Nvidia investing in AI laboratories and neoclouds while those businesses use financing to purchase Nvidia hardware. Nvidia has also undertaken commitments relating to unused computing capacity. Myrmikan’s objection is not that these arrangements are fictitious — the chips and data centres certainly exist — but that apparent demand at one point in the chain may partly depend upon financing supplied elsewhere in the same chain. Revenue can therefore be economically less independent than accounting statements make it appear.

That is a serious issue. There is nothing inherently improper about vendor financing or strategic investment. Aircraft manufacturers, property developers and telecommunications companies have employed comparable arrangements for generations. But circular financing becomes dangerous when investors mistake financed demand for final demand.

The distinction resembles building an enormous number of hotels because banks are willing to finance hotel construction. For a while everybody involved prospers: builders build, suppliers sell concrete, employees receive wages and banks collect fees. Yet eventually somebody must sleep in the rooms at a price sufficient to service the debt.

For AI, that somebody is ultimately the ordinary business customer.

The hyperscalers

Myrmikan therefore examines Microsoft, Amazon, Google, Oracle and the other hyperscalers. These companies are not speculative start-ups. They possess immensely profitable existing businesses. That fact makes the AI boom simultaneously safer and more dangerous.

It makes it safer because Microsoft or Google can absorb investment mistakes that would destroy a conventional start-up. But it makes it more dangerous because their creditworthiness permits them to mobilise capital on a scale no start-up could contemplate.

The paper emphasises the divergence between operating cash flow and AI-related capital expenditure. Its charts depict hyperscaler operating cash flow continuing upwards while adjusted free cash flow falls sharply below zero, accompanied by rising net debt issuance.

The intended message is that some of the richest corporations in history are increasingly consuming more cash through AI investment than their businesses generate after relevant investment commitments and shareholder distributions are taken into account.

Again, the interpretation deserves qualification. Capital expenditure is not intrinsically evidence of financial weakness. A company should borrow or spend heavily when it has unusually attractive investment opportunities. Amazon spent extraordinary sums building infrastructure long before AWS became one of the world’s most valuable businesses.

The question is therefore not whether the hyperscalers are spending too much relative to current cash flow. It is whether the assets they are purchasing will earn returns above their cost of capital before technological obsolescence destroys their economic value.

GPUs are particularly awkward assets in this regard. They are expensive but relatively short-lived. Today’s scarce accelerator can become tomorrow’s second-rate chip remarkably quickly. A railway constructed in 1870 might still carry trains fifty years later. A data centre remains useful for decades, but its most expensive computational contents may not.

The hidden balance sheet

The next stage of the argument concerns neoclouds such as CoreWeave and Nebius. Instead of hyperscalers constructing every data centre themselves, they can contract with specialist providers for future computing capacity. The neocloud raises debt against those contracts and constructs the infrastructure.

Economically, Myrmikan argues, this can resemble capital expenditure by the hyperscaler while appearing differently on its accounts. The paper cites estimates of enormous lease and purchase commitments that are not reflected straightforwardly as conventional balance-sheet debt.

The expression used in the paper — “laundering capital expense” — is deliberately provocative and probably too strong. Off-balance-sheet contractual obligations are not necessarily concealed and accounting rules distinguish leases, purchase agreements, guarantees and funded debt for legitimate reasons.

Yet beneath the polemic lies a valid analytical question: who ultimately bears the risk if projected demand for computing power fails to materialise?

A hyperscaler may not own a data centre. But if somebody borrowed billions to construct it because the hyperscaler signed a long-term take-or-pay contract, the economic connection remains real.

The paper reports an estimate of $1.2 trillion of debt tied to artificial intelligence and describes proposed financing structures using computing capacity as collateral.  The danger is therefore increasingly not merely that shareholders have paid absurd prices for AI companies. Shareholders can lose everything without necessarily causing a financial crisis. Credit is different.

Debt connects bubbles to institutions that cannot afford losses.

From AI to life insurance

Here Myrmikan makes its most original argument.

Private-equity and alternative-asset managers have increasingly acquired or partnered with life insurers. The attraction is obvious. Conventional private-equity funds periodically return capital and must raise new funds. Life insurers receive a continuing stream of premiums against liabilities extending decades into the future. They therefore provide something approaching permanent investment capital.

Private-credit managers also originate precisely the sorts of loans required by data centres and other capital-intensive AI infrastructure projects.

The potential conflict is apparent. An asset manager can originate private credit, collect fees for arranging it and place some of the resulting assets within insurance businesses over which the same financial group exercises substantial influence. Myrmikan argues that the insurer receives the limited upside associated with debt while ultimately carrying much of the downside risk.

The scale is substantial. The paper reports estimates placing private-equity-controlled life-insurance assets between hundreds of billions and approximately $1.5 trillion, depending upon definition. It also points to the disproportionately large appetite of PE-owned insurers for privately placed instruments.

Reinsurance adds another layer. Liabilities may be transferred to affiliated offshore reinsurers operating under different capital, accounting and disclosure regimes. The paper notes that US life insurers had ceded $2.1 trillion of reserves by 2023, with offshore reinsurers accounting for a rapidly growing proportion. Life insurers meanwhile owned $849 billion of the roughly $2 trillion private-credit universe.

This is where the analogy with 2008 becomes genuinely interesting.

The next financial crisis need not resemble the previous one institutionally. Before 2008, systemic leverage accumulated conspicuously in banks, mortgage securities and shadow banking. Regulators subsequently made banks substantially safer. But capital does not disappear when one channel is regulated; it searches for another. Private credit, insurance balance sheets, captive reinsurance and long-duration infrastructure financing may represent parts of the contemporary shadow financial system.

Myrmikan may overstate how much AI debt has actually migrated into life insurers — indeed the paper acknowledges that the opacity of private markets makes the amount impossible to establish precisely. That is an important evidential gap. Demonstrating that insurers own a large share of private credit and that AI infrastructure increasingly relies upon private credit does not establish exactly how much AI risk insurers possess.

But the absence of transparent information is not entirely reassuring. Opacity is itself one of the characteristics that makes financial contagion difficult to manage.

Can an insurance company suffer a bank run?

The paper argues that it can.

Life-insurance liabilities are less liquid than bank deposits because policyholders may face surrender charges and tax consequences. But they are not completely locked away. As policies mature, surrender penalties diminish. If policyholders become concerned about an insurer’s solvency, rational individuals may prefer an immediate tax cost to a potentially larger future loss.

The insurer must then find cash. Liquid assets are sold first, leaving the remaining policyholders increasingly exposed to illiquid private assets whose market prices may be uncertain. The result could resemble a slow-motion bank run.

The paper points to the rescue of AIG in 2008 as precedent. Contemporary insurers may in some respects be more difficult to stabilise because private-credit assets cannot necessarily be sold quickly at observable prices.

This is the crucial bridge between an AI investment bust and the Federal Reserve.

If an AI laboratory fails, Washington need not rescue it. If a data centre defaults, Washington need not rescue that either. If an insurance company holding the debt fails, matters become politically different. If several insurers face simultaneous redemptions, pension funds hold impaired securities and financial markets cannot establish which institutions are solvent, the government confronts the familiar problem of systemic contagion.

At that point the provenance of the crisis becomes almost irrelevant.

The Federal Reserve trap

Myrmikan predicts that the Federal Reserve will ultimately respond exactly as it has in previous crises: it will create liquidity.

The paper’s contention is that monetary policy confronts an insoluble political asymmetry. High interest rates may preserve currency stability but expose malinvestment. Low rates preserve leveraged institutions but encourage inflation and further leverage. When the choice becomes immediate financial collapse or future monetary debasement, governments almost invariably choose the latter.

Hence the paper’s memorable final formulation: “print or die”.

This conclusion is rhetorically effective but economically too deterministic. The Federal Reserve possesses instruments other than simply printing money indefinitely. Emergency liquidity facilities can be collateralised and temporary. Insolvent shareholders can be wiped out while systemically important liabilities are protected. Fiscal authorities can recapitalise institutions. Regulators can force mergers or restructurings. Quantitative easing need not translate mechanically into consumer-price inflation.

Nevertheless, Myrmikan identifies the political economy correctly. Governments facing systemic financial panic have repeatedly preferred intervention to liquidation. The interesting question is not whether they could allow a disorderly collapse. It is whether any elected government actually would.

History suggests considerable caution before betting upon heroic monetary restraint in the middle of a panic.

Where the argument is strongest — and weakest

The paper is strongest when treated not as a precise forecast but as a map of vulnerabilities.

Its most persuasive propositions are these: AI’s technological importance does not prove current valuations; productivity gains may arrive much more slowly than model capabilities improve; AI infrastructure is extraordinarily capital-intensive; hardware depreciates rapidly; parts of the industry’s demand are financially interdependent; off-balance-sheet commitments complicate assessments of leverage; private credit has become an important source of AI infrastructure finance; and insurance companies have become increasingly intertwined with private-equity and private-credit markets.

Those propositions together describe a financial system deserving considerably more scrutiny than the cheerful phrase “AI investment boom” conveys.

The paper is weaker when possibility becomes certainty. Neoclouds may collapse; it does not follow that they inevitably will. AI laboratories may require restructuring; equally, inference costs might fall rapidly enough to produce enormous new demand. Hyperscaler investment may prove excessive; it might instead build the infrastructure for a genuinely transformed economy. Life insurers may contain significant AI-related private credit; the paper cannot establish precisely how much. And a crisis in private credit would not automatically require monetary intervention on the scale imagined.

There is also an ideological tendency to force disparate episodes into one Austrian monetary narrative. The claim that virtually every major crisis results from administratively suppressed interest rates is too simple. Financial crises involve leverage, maturity transformation, regulation, expectations, technological change, fraud, international capital flows and ordinary human error. Monetary policy matters enormously, but it is not the only variable.

Some of the paper’s political rhetoric likewise detracts from its financial analysis. One does not need to accept its characterisation of Keynesianism, government or contemporary American political parties to recognise the balance-sheet risks it identifies. Indeed, those rhetorical excursions occasionally make an interesting financial argument easier for sceptical readers to dismiss.

The deeper issue

The deepest question raised by the paper is therefore not whether AI is a bubble.

Almost every revolutionary technology produces a bubble because human beings have difficulty distinguishing the eventual importance of an invention from the present value of the companies selling it.

Railways were transformative. Railway investors nevertheless went bankrupt.

The internet changed civilisation. Pets.com nevertheless disappeared.

Telecommunications traffic exploded after the dot-com crash. That did not rescue shareholders who had financed redundant fibre at bubble prices.

AI could therefore fulfil almost every extravagant technological prediction made about it while simultaneously producing one of the largest investment busts in modern history.

The paper captures the broader danger particularly well. It places rapidly increasing life-insurance policy loans beside broker margin lending under the heading “Leverage on Leverage”. The point is not merely that there is debt. Modern capitalism could scarcely operate without it. The problem arises when one leveraged claim becomes collateral for another leveraged claim, whose value ultimately depends upon assumptions about assets whose profitability has never been tested through a complete economic cycle.

And this is why the Myrmikan paper deserves attention even from readers who reject its apocalyptic conclusion.

The central risk may not be that artificial intelligence fails. It may be that AI succeeds spectacularly while becoming much cheaper than investors expect. If computational intelligence becomes abundant, model prices collapse, hardware efficiency rises and competition eliminates excess margins, society could receive an extraordinary technological dividend while owners of the infrastructure receive disappointing financial returns.

That would be excellent news for civilisation and terrible news for anyone who borrowed a trillion dollars on the assumption that scarcity would persist.

Myrmikan ultimately predicts a sequence of falling technology valuations, collapsing AI-backed bonds, pressure upon pensions and insurers, government intervention and renewed monetary expansion. That precise sequence is by no means inevitable. Financial history rarely repeats itself with such theatrical neatness.

Yet the underlying warning is harder to dismiss. The dangerous stage of a bubble begins when exuberance migrates from equity into credit. Equity investors can lose fortunes. Debt failures can destroy institutions. Once the liabilities of those institutions are implicitly guaranteed by the state, private speculation acquires a public balance sheet.

At that point the debate is no longer really about artificial intelligence.

It is about who pays when everybody discovers that something immensely useful was not nearly as valuable as they had imagined.