Crazy Train

The Credit Strategist - August 2025

Crazy Train

“Well-nigh two thousand years and not a single new god!”

Friedrich Nietzsche, The Antichrist (1888)

Or maybe not. The world believes it discovered a new god in artificial intelligence (AI) and machine learning. Many of the gains in today’s market bubble – which looks like the Internet Bubble on steroids – are driven by the AI boom. Each financial bubble seems to exceed the last in size and this one is no exception. But it’s important to understand why this bubble is so enormous. There are several reasons:

· First, the financial system and quantum of global debt are much larger than twenty-five years ago. Massive fiscal deficits and low interest rates inflated the amount of money in the system. According to the International Monetary Fund (IMF), total global debt (including government, corporate and household debt) was around $64 trillion in 2000. As of mid-2024, the IMF estimated total global debt was $314 trillion or 500% higher than twenty-five years ago. Global stock market capitalization peaked at $44 trillion in March 2000 before the bubble burst; as of mid-2024, global stock market capitalization was $110 trillion and is closer to $120 trillion today, or about 300% of its size during the Internet Bubble. And there are only half as many U.S. public companies today as twenty five years ago in which this capital can be invested.

· Second, market structure evolved since 2000 to make it much easier for investors to buy stocks through passive vehicles like ETFs while relieving them of the need to perform any research. Explosive growth of options markets with new products like zero-day options also expanded market participation without knowledge, turning markets into a new form of gambling. Today, unlike in 2000, investors can trade stocks on their phones as easily as playing computer games (with even less thought), which attracts more participants and increases volatility.

· Third, the massive fiscal and monetary stimulus of the last three decades inflated the value of all financial assets, not just stocks. Real estate, art, collectibles, commodities – every asset denominated in fiat currencies rose sharply in value. The S&P 500 is trading at nearly 30x earnings compared to a historical average of 16x and the Nasdaq is trading at 40x earnings compared to a historical average of 22x. Other valuation measures such as the Buffett Indicator, Schiller Cyclically-Adjusted P/E, Price/Sales Ratio, and Dividend Yield are at or near record levels, meaning the levels of the Internet Bubble. These valuations are far above the average earnings growth rate of even the most successful companies (with rare exceptions like NVDA whose growth rate is unsustainable) which many (myself included) consider a reliable long-term valuation benchmark. Such stock prices are unsustainable because the growth rates underlying them can’t support them.

There are different ways to define a bubble. I define a stock market bubble as a trading level far above companies’ ability to generate sufficient income to support their stock prices. By that definition (and others), we are squarely in a bubble now (both in equity and credit). Bubbles can continue much longer than seems possible but ultimately collapse under their own weight (i.e., stock prices move lower toward their companies’ growth rates). This one will end the same way, especially because it is inflating as economic growth is tepid.[1] Real (inflation-adjusted) growth is going to struggle under the weight of tariffs (whose effects have yet to set in) and mounting debt and AI will take years to counter those headwinds (if it ever does). The market is inflated by AI hype, massive fiscal deficits, negative real interest rates, and understandable relief from the end of four years of the Biden Administration’s war on growth (which looks worse the farther in the rear view mirror it appears). Stock prices are also aided-and-abetted by the imbecilic game of managed earnings whereby analysts massage down their earnings estimates (with the complicity of companies) to levels that can be easily beat, confirming that life remains like high school long after we survive our senior prom.

Just as companies around the turn of the millenium spent huge amounts of money building out infrastructure to accommodate future internet growth (which eventually materialized though many companies failed), companies are today spending even more gargantuan sums on LLMs (large language models). The top four hyperscalers (GOOG, META, AMZN and MSFT) are spending more than $300 billion this year on capex – roughly 55% of their annual cash flows. Additional hundreds of billions are being spent by ORCL, SFTBF, IBM, OpenAI, Chinese competitors, Middle Eastern sovereigns and others. This spending orgy is based on almost religious belief that AI will change virtually every aspect of human life. Maybe it will, but maybe it won’t, and maybe every aspect of human life doesn’t need to be changed (or maybe the parts that need to be changed can’t be changed by computers). It seems to me that AI’s greatest promise lies not in LLMs (large language models) but in projects requiring the analysis of mass data like medicine, space travel, military applications and the like that may take time to generate significant financial returns. But it appears Silicon Valley is again focusing on consumer and social media applications rather than more productive applications as Palantir’s Alex Karp warned in his book The Technological Republic. Certainly Mark Zuckerberg’s uninspiring July 30^th statement describing “Personal Intelligence” suggests a need for a more compelling vision for AI than the one that fueled the explosion in the solipsistic social media and entertainment applications that waste so much human capital today.

Some of the financial commitments made today are going to be difficult to accomplish. For example, OpenAI recently committed to pay more than $30 billion annually to ORCL starting within three years for 4.5 gigawatts of datacenter capacity according to The Wall Street Journal.[2] That is equivalent to the power generated by two Hoover Dams or enough to power four million homes. OpenAI recently projected $10 billion in annual revenue so it will need explosive growth to fulfill this contract. The $500 billion Stargate partnership among OpenAI, Softbank, and ORCL has yet to complete a single deal and scaled back its near-term plans as the partners hash out issues such as where to build sites. They now hope to build one small data center by the end of the year according to the Journal. And just like during the Internet bubble, investors are using not only equity but debt to fund these projects. CRWV raised billions of dollars of debt to purchase GPUs from NVDA and build huge AI datacenters and then went public and now trades like a meme stock while waiting to become profitable.[3] META (which still derives most of its revenues from advertising) is also joining the debt bandwagon by raising $26 billion in debt (and $3 billion in equity) from large private credit shops led by APO to fund its AI investments (which include throwing exorbitant pay packages at AI scientists).[4]

Not to be undone, Elon Musk is desperately trying to keep pace by raising billions of dollars at xAI. Weeks after raising $10 billion of equity and debt in June, he is now working with his pal Antonio Gracias’s Valor Equity Partners to raise another $12 billion to buy a massive supply of GPU chips from NVDA that would be leased to xAI for a new data center to train and power the AI chatbox Grok (he is targeting one million chips). Musk recently directed SpaceX to invest $2 billion in xAI and collateralized the $5 billion debt component of the June financing with Grok’s intellectual property among other assets. xAI has very little revenue and is spending money hand-over-fist with 2025 cash burn expected to be about $13 billion according to the Journal. And this spending is just the beginning. Musk says that xAI plans to have 50 million H100-equivalent AI computers in five years. To place these numbers in context, OpenAI’s GPT-4 (state-of-the-art when released two years ago) was trained on just 25,000 GPUs. OpenAI will have one million or forty times that number by year end and xAI plans to have 200x that number in five years. Leaving aside the very material question of how these companies will pay for these chips, they must also obtain the power required to run them since that power doesn’t exist and has to be built. Valor is going to have its work cut out for it and will likely have to focus on Middle Eastern sovereign funds and similar sources to meet xAI’s financing needs [5]

Anyone who questions AI faces the challenge of answering technology experts (genuine and self-proclaimed) who claim superior knowledge of the topic. But the projections on which future AI revenues and profits are based remain highly speculative because they are based on events that have yet to happen; put bluntly, they are based on predictions about the future that is always unknowable. But one characteristic of the future that is knowable is that it is reflexive, meaning human beings and organizations react and adjust to changes rather than remain static. As such, arguments that AI will eliminate jobs without creating new ones or arbitrage away margins without creating new margin opportunities are questionable. Further, it is unlikely that multiple LLM models addressing the same market will all prove successful; more likely, ruthless competition will create a small group of winners and many losers with a great deal of capital consumed in the process.

The quantum of AI spending dwarfs anything previously applied to a specific product or sector in such a concentrated period of time. All of this spending has yet to produce a commensurate amount of revenue or profit but we are still in early days and investors are convinced that it will. Whether the world needs numerous LLMs that perform similar tasks remains to be seen; it’s not clear that these models can