A Bigger Short?

The Political Economy of the AI Bubble

Minsky and the Bear

The renowned Big Short investor Michael Burry has taken a bearish position on Caterpillar, a major beneficiary of AI hyperscaler’s expansive drive to build AI data centres. He also briefly expanded his short position on Palantir before reducing it again after a rally in the stock price following their deal with Nvidia to help develop AI models for US Government customers. However, there is little doubt about an overarching truth: Burry is bearish. So bearish in fact that he has suggested on his Substack Cassandra Unchained that plans in the tech heavy South Korean market to significantly increase AI infrastructure capex marks the “beginning of the end”.

The fear of course is the AI bubble that has grown at an astronomical level over the last few years, with the five biggest hyperscalers in the US on track to spend more than $1 trillion on AI between 2025 and 2026.

He is not the only one, Jeremy Grantham, who also predicted both the 2008 housing bubble and the dot com bubble of the 1990s and early 2000s has urged investors to sell their holdings of US equities. Ann Pettifor, another economist who predicted the global financial crisis in her 2006 book The Coming First World Debt Crisis suggested we may well be entering a Minsky moment for Ponzi finance after the NASDAQ declined 2.2% from June 20^th to 22^nd.

The trouble with Minsky moments is that for many investors (at least those who can take a step back in the midst of debt-fueled speculative growth) they are pretty much bound to happen, but the timing is always fairly uncertain. It is highly likely that this AI capex cycle is still in an expansionary phase, meaning even the most risk averse investors won’t want to get out too early and miss the growth.

The Minsky moment, named after economist Hyman Minsky, follows a stage in financial markets known as “Ponzi finance”. This is when financial markets, incapsulated by prolonged bull markets, have taken on so much debt they have to borrow to meet their existing liability payments, making them dependent on rising asset prices for collateral.

Understanding the Minskyian boom bust cycle is likely to prove as essential to understanding the AI bubble as it was to understanding the 2007/08 crash. In brief terms, the cycle begins with a “historic economy” in which a crisis has taken place and banks and firms are more risk averse, meaning they only lend for conservative projects and avoid excessive leverage. However, as the economy recovers, those projects begin to do well and markets look at their returns and think “if only we were more leveraged”.

They then take on further debt and begin to revalue assets, such that we enter the “euphoric economy” where the decline in risk aversion increases the money supply, increases investment and increases asset prices, making speculation more profitable. As debt levels rise, Ponzi financiers emerge whose cash flow from investments are less than their debt servicing costs, meaning they make profits by selling assets in a bull market and have a never ending demand for debt to cover up what is essentially insolvency. Charles Ponzi and Bernie Madoff are the classic examples.

When the increase in the supply of credit drives up market interest rates, previously conservative projects become speculative and, inevitably, previously bullish investments fail to meet expectations, causing a sell off of assets as investors struggle to cover spiraling loan payments. This means the era of asset price inflation that Ponzi finance depended on is no longer in operation, which rapidly amplifies losses and causes a Minsky moment, or market collapse.

It is easy to see how this applied to 2008, with the dot com bubble representing the historic economy, the mortgage backed security (MBS) fueled growth of the 2000s representing the euphoric economy, and the crash representing the Minsky moment. It seems increasingly likely amongst economists and traders that we may be in the midst of a new euphoric economy, with the global financial crisis now representing the historic economy. This cycle is intrinsic to capitalism, particularly financial capitalism, which, as Minsky suggested, has a fundamental instability that flows upwards, turning recovery into yet another speculative boom.

The Bubble

Investor and finance whiz Paul Kedrosky highlights that this particular bubble has managed to combine four components only previously seen in separate bubbles.

First, it has a speculative real estate element given that AI data centres are first and foremost pieces of commercial real estate rented out to both large hyperscalers and risky startups (often in the same rental unit to drive yields up in what essentially amounts to the securitization of tenants).

Second, there is a revolutionary technology at its centre with the potential to drastically change society and the economy, as railroads did in the late 19^th century. Importantly, the bubble bursting is not at all dependent on whether the technology is revolutionary, indeed most revolutionary technologies coincided with speculative bubbles as they were financed, including railroads.

Third, it is fueled by loose credit from banks, corporate bonds, and increasingly the shadow banking sector. The Bank for International Settlements has published a report highlighting specifically the dangers of AI infrastructure capex coming from private credit markets.

Fourth, and perhaps the most important for thinking about how we come out of the likely crash, the US state is actively expanding the market. In the run up to the global financial crisis Fannie May and Freddie Mac were heavily involved in securitizing mortgage bonds into Asset Backed Securities (ABS) and guaranteeing them.

In this crisis, the AI race has become a national security issue, with the US Government accelerating the pace of permits, offering tax incentives and subsidies to reduce the cost of building AI infrastructure, and handing out public procurement contracts to companies like Palantir and Anthropic. This will have serious implications on how a potential crash might be framed, but more on that later.

When thinking about the AI bubble, we therefore need to understand that it is not just a story about inflated stock market valuations leading to a market correction in the way that the dot com bubble was. It is helpful to think about three distinct contributing figures: infrastructure capex, private capital markets, and public market valuations.

Infrastructure capex

As OpenAI launched Chat GPT in 2023 and the AI race was triggered, aggregate capex from the five biggest hyperscalers (Amazon, Microsoft, Alphabet, Meta and Oracle) has grown from $224 billion in 2024 to between $660 billion and $690 billion this calendar year. Total worldwide spending on data centre infrastructure has exceeded $1 trillion and, according to McKinsey, will reach $7 trillion by 2030. This is largely because AI data centres are highly capital intensive, requiring enormous amounts of energy and constant processing power to run General Processing Units (GPUs).

At first, this was largely funded by large technology companies with enormous cashflows. However, as these public companies have been spending increasing amounts on data centres, shareholder value has diminished, not least because there has been less cash available for share buybacks as capex began exceeding cashflow. Increasingly, external financing, even for public companies with lots of free cashflow, has been sustaining this huge expenditure as a result.

Corporate bond issuance thus began to grow exponentially, with Amazon and Alphabet issuing $60 billion in corporate bonds in multiple currencies in the last 12 months. To keep much of the debt financing off companies balance sheet however, Special Purpose Vehicles (SPVs) have been set up which allow partners to invest capital into a separate legal structure which retains the rights to the project. These will include pension funds, infrastructure funds, banks and private credit funds.

These are of course securitized investments, with the GPUs acting as the collateral. Problematically, tech companies extended the depreciation schedules for the data centre assets around the same time as the drive in GPU data centres, despite the fact that GPUs are used 24/7 for the intensive data processing needed to run a Large Language Model (LLM) like ChatGPT. This means they experience thermal degradation and have a life span of around eighteen months when used this intensely, whilst on company’s balance sheets they have a depreciation schedule of around five years.

This not only means there is a mismatch between the cashflow and depreciation schedule but also that the underlying collateral financing the data centres has to be constantly replaced. And how are they replaced? Via more debt with the GPUs serving as the collateral. This creates a Ponzi finance-like dependency on rising asset values and rising demand for GPUs even just within the data centre. The minute asset prices fall, companies struggle to pay the loans on their existing liabilities and struggle to replace the collateral on those loans.

This seems like a Minsky moment waiting to happen, but this time the collateral is constantly losing value at a rapid rate meaning the dependency on loose credit and rising asset prices is even greater as there is a constant need to turn over the base (replace the collateral). It’s not exactly surprising therefore that the fastest growing market now lending to hyperscalers is the opaque private credit market.

Private capital markets

Private credit funds are the fastest growing market financing data centre capex. Morgan Stanley estimates they could supply over half of the $1.5 trillion needed for the data centre buildout until 2028.

Private credit, or shadow banking, has grown since the global financial crisis as regulatory changes constrained lending in the traditional banking sector. The industry has faced serious scrutiny, particularly recently, due to its opaque nature, complex debt structures, loose regulation and poor rating standards for borrowers. There are also significant concerns that the industry’s exposure to standard software companies at risk of being upended by AI could cause a correction in private credit as soon as 2027, when those software loans need to be refinanced.

One of the key means of investing in these data centres, particularly for private lenders, is through the ABS. Private credit, for example, is increasingly investing into collateralised loan obligations (CLO). If you think that sounds familiar, it should… it is a type of collateralised debt obligation (CDO). The kind that Ryan Gosling explains with Jenga blocks in The Big Short. These are fairly standard means of raising capital and aren’t necessarily crisis prone as an instrument in isolation, but their complexity allows them to disguise risk as diversification. This was particularly true in 2008 when subprime mortgage loans were packaged together and given investment grade ratings simply because some were in California and some were in New York, for example.

Traditional banks are exposed to these funds in a number of ways which has raised concerns about a repeat of 2008 in the private credit market. Importantly, private credit CLOs are far less leveraged than mortgage-backed CDOs before the global financial crisis. CDOs were often leveraged 10x–15x and embedded within increasingly complex structures such as the CDO-squared and synthetic CDOs, which used credit default swap exposure to existing CDOs as the collateral. CLOs generally have lower leverage and simpler cash flow structures.

However, the scale at which private credit seems to be financing AI data centres is perhaps the most troubling piece of this complex bubble, creating much greater demand for more complex debt structures and increased leverage. Of course, as Minsky shows, the moment the underlying collateral - which in this case is the future cash revenue from the data centres - falls in value and those overly leveraged companies can’t repay their loans, the system collapses. And if private credit collapses because of its exposure to an AI bubble, we may well enter unseen territory, particularly given private credit may well be a bubble in itself.

The traditional banking sector’s exposure to the private credit industry is supposedly low, and certainly isn’t at 2008 levels according to the US Office for Financial Research, but even by their own admission it is almost impossible to accurately measure. The issue primarily is upstream funding where banks lend to institutions who invest the money into a fund as a Limited Partner. This kind of leverage exposure is nearly impossible to measure.

Meanwhile the measurable exposure, where banks lend directly to the fund through subscription lines, Net Asset Value loans and repos is supposedly being offset by significant risk transfers. These allow banks to offload their balance sheet exposure to private credit funds by buying default protection on their loans to the sector. Banks have been doing this to a large scale in order to free up further lending capacity to private capital markets, their most profitable clients. However, private credit funds have been a large part of the market selling protection on these loans, which they often use banking leverage to finance. So how exactly is the risk is leaving the banking sector, you ask? I don’t really know.

If this all sounds too confusing to understand, you are not alone. If the fund managers and investment bankers properly understood it they’d likely be a bit more cautious. But what should be clear is that this is not just a stock market bubble caused by overvaluations. It is an infrastructure bubble, a debt bubble, a real estate bubble and a valuation bubble all at once. But of course the catalyst that could spark the doomsday scenario economists have been envisioning would be a fall in the stock price of public companies as their cashflows become incapable of paying their existing liabilities, never mind producing a return.

Public market valuations

The big issue is not AI’s ability to change the way we operate or genuinely revolutionise large parts of our economy. It is whether it can make enough money while doing so to make the large scale investment worthy and to meet the demands of this sustained period of debt fueled growth. Nvidia’s stock valuation is of course historic, but it is fueled by expectations on future revenue streams. This is not risky, it is uncertain. The difference is that we can calculate risk, we cannot calculate uncertainty. We can only attempt at an estimate.

When you consider that the greatest stock market valuations of all time are based on these “vibes”, it becomes clear why overvaluation could be an issue. This is particularly true when we consider that AI has negative unit economics, meaning that rather than making a profit on each additional customer, it makes a loss. This is because the operating costs of LLMs are constant and grow alongside usage.

Standard software is different in its unit economics as new customers bear no (or very little) additional cost, so once you have sold enough individual units to make a profit you are in the green. Nvidia of course makes its money largely from selling GPUs, but that still means its customers have negative unit economics.

LLMs require constant energy power as each prompt is processed at a scale so large that the costs of each additional user are generally higher than the profits. Most LLMs will get their profits from Application Programming Interface (API) bills. Think of these as being like electricity bills, only for AI usage. One company is essentially renting the use another company’s AI and receives a bill at the end of the month based on the amount of processing done.

These can be enormous bills and do make up a large portion of OpenAI and Anthropic’s cash flow. However, the concentration of API customers contributing to this pool of revenue is low, meaning one loss can have a significant impact, and these customers are generally large enough that they will likely develop their own AI systems eventually.

OpenAI has diversified its revenue pool more than Anthropic, and is even looking into advertising. But if advertising becomes the monetising tool used by this revolutionary technology then evidently it was always far more revolutionary in terms of its application than its ability to generate profit. If a market correction does take place and the value of these stocks fall, and with it the cost of GPUs, then the one thing sustaining the debt in the Ponzi finance era, rising asset prices, cannot be relied upon. And then we truly are in a Minsky moment.

The Politics of Sustained Bullshit

This is all very apocalyptic, I know. Believe me, I’m having more than just one pint after writing this one. Whether it gets this bad is not something I can predict with certainty. But evidently the question is moving away from “is this a bubble?” and more towards “when will it burst?”

However, what progressive economists should really be asking is, how does this all get framed in the aftermath? This is where the politics of Minsky’s cycle becomes important, and it is what I call, in professional terms, the politics of sustained bullshit. The economic argument presented by Minsky is clear. What is less clear is how the politics plays out. Why do we accept the never ending cycle? Why doesn’t the state intervene in boom bust cycles rather than expand them?

The state absolutely can and sometimes does step in and provide counter cyclical macroeconomic measures to smooth business cycles. But evidently the neoliberal orthodoxy at the central bank and Treasury are not willing to do what’s necessary to reverse the era of financial speculation that has marked the economy since Thatcher and Raegan. This is because neoliberalism has a seemingly never ending ability to fail; when it fails, it “fails forwards.”

At the centre of this is an inherent paradox within the neoliberal state. The state, as I have argued before, is not small under neoliberalism, it is both marketized and marketizing. This means it plays this crucial role of market expansion that sets neoliberal economics apart from classical liberalism, but also acts as the inevitable scapegoat when the market fails.

After 2008, neoliberals blamed the state’s expansion (now framed as interference) in the housing market (as well as immigrants and poor people “biting off more than they could chew”) and were able to suggest that the problem was that we weren’t neoliberal enough… We didn’t let the market sort it out. Of course this is bullshit, not least because the state ultimately had to bail the market out. But it does explain how the politics of a Minskyian boom bust cycle is framed, shifting the blame away from financial speculation and excessive leverage and towards the state, poor people and immigrants.

As a result, if the AI bubble is real (which seems beyond reasonable doubt at this point), it may well be the fourth component of this bubble, namely that the state is expanding it, that allows neoliberalism to get up and go again. This will mean state bail outs, another round of quantitative easing, and maybe even austerity again. But it seems undeniable that this would be political suicide. Maybe then, just maybe, the gig is up and the leftist Keynesian Jedi can have their Luke Skywalker moment and destroy an AI bubble shaped Death Star, defeating the neoclassical and neoliberal Sith Lords that currently rule the galaxy. Or will we just end up going even further right than Emperor Donald Trump?