Trillions in data centers, funded through the same opaque vehicles that broke the last cycle.
I wrote in the past that the AI rollout feels a lot like déjà vu of the 1999 telecom bubble. Today it is also starting to feel like the 1999 bubble is being supersized into something closer to what led to the 2008 financial crisis.
What is the difference between the two?
The 1999 bubble had two parallel and interrelated dynamics: overvaluation of certain segments of the market (dotcoms, beneficiaries of the new economy, and a slew of other stocks), and overinvestment by telecom companies in internet infrastructure. When the bubble burst, it deflated overvalued stocks and brought value stocks back from the dead. It also revealed overcapacity in telecom infrastructure while sending many of those companies to meet their maker. And it brought a mild recession.
The 2008 crisis started with a housing bubble, but housing is not what nearly took down the economy. It was the collapse of housing-linked financial instruments that infected the banking and financial system. That is what went into the history books as the Great Recession.
Today’s AI bubble has elements of both. Let me take them one at a time.
AI brings a transformational promise and, with it, incredible optimism, which has led to a data center buildout marching toward a trillion dollars a year. At the tip of the spear are OpenAI (creator of ChatGPT) and Anthropic (creator of Claude). The growth and sheer size we see here are unlike anything we have seen before: Anthropic’s revenue has grown nearly 10x in a year, three years running. But its losses accelerate along with its revenues.
Behind them stand the major tech companies: Google, Microsoft, Meta, Amazon, xAI (creator of Grok), and Oracle, a more recent player in AI infrastructure, with multi-hundred-billion-dollar commitments to OpenAI. The relationships among these companies are complex. They are often partners, competitors, and vendors: at times all three at once.
If AI were only a capital-light business, then whatever happens in AI land would stay in AI land. Instead, these businesses make airlines look capital light. Growth requires data centers to support it, and that is where things get dangerous very fast.
Here is Dario Amodei (Anthropic’s CEO) in February: “If my revenue is not 1 trillion dollars, if it’s even $800 billion, there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt if I buy that much compute… If I’m just off by a year in that rate of growth, or if the growth rate is 5x a year instead of 10x a year, then you go bankrupt.”
What Dario is telling us is that he has to commit hundreds of billions of dollars without knowing what demand will be. If his revenue forecast is off by 20% or a single year, he is done.
And Dario is on the conservative end of this race. He is the one saying the number out loud and planning against being wrong. His competitors see AI as an existential threat and have put the pedal to the metal building data centers.
This creates incredible inflation in everything the buildout touches, starting with GPUs and memory chips. Nvidia and Micron, responding to insatiable demand, have raised prices and now earn margins as if they were software companies.
Their customers, many of whom had the capital-light profile of software companies, have gone from generating enormous free cash flow to being cash flow negative, issuing debt and even equity for the first time in decades. And since they are all competing for the same chips, the same generators, and the same construction labor, they are paying multiples of what these goods and services would cost in a more rational environment.
This is the 1999 element of the story. These data centers cost a lot of money and incur significant fixed costs. But the profitability of AI is elusive. Companies that went all in on AI are struggling with the bill, and many are starting to ration their usage. Open-source and Chinese models are delivering results comparable to frontier models (developed by OpenAI, Anthropic, Google, and xAI) at a fraction of the cost. Good enough is a powerful argument when it costs a tiny fraction of the alternative.
All of this means that what Dario was worried about may come true. It starts to look like a race to the bottom, as trillions of dollars worth of data centers compete for the same customers. At some point, the prices the big guys charge will have to decline to meet the alternatives. This is exactly what happened in fiber optics after 1999. The same thing will happen in the sexier but still highly capital-intensive business of AI.
Then the bankruptcies come, and we find out who was swimming naked.
And this is where 2008 comes in: it is about who financed the swimmers.
These assets need to be financed. Some of the capital comes from the cash flows, equity, and debt of the companies themselves, but a large chunk is going into special purpose vehicles (SPVs): opaque investments packaged by Wall Street that regulators struggle to value and are sold to pension funds, private investors, and insurance companies.
There is also a lot of incestuous vendor financing going on. For example: Nvidia is in talks to guarantee roughly $250 billion (yes, with a B) in lease and debt financing for an OpenAI data center campus in Ohio, and separately to help finance up to $350 billion of OpenAI’s purchases of Nvidia chips. The customer cannot afford the product, so the vendor cosigns the loan. Nvidia’s guarantee exists precisely to reassure the lenders backing these financing vehicles, as the borrower does not have an investment-grade rating.
Speaking of Nvidia: I lived through 1999 and vividly remember how companies like Nortel and JDS Uniphase were respected and how their CEOs were celebrities on CNBC. Then one day the script changed and they collapsed. It is hard to imagine something like this happening to Nvidia, but the semiconductor business has been and will remain a cyclical business. This time is no different. I would be very cautious trying to find a bottom in Nvidia when it goes down.
A recent paper by Pranjal Drall and Andrew Granato, “Private Credit’s State Backstop,” describes how private equity firms have bought up life insurers and stuffed their balance sheets with opaque private credit. When one of these insurers fails, it does not go through bankruptcy. Other insurers are assessed to cover the shortfall, and in most states they recover the cost through tax credits. Roughly 86% of the bill quietly lands on taxpayers. No vote, no headline. The paper is about private credit broadly, but AI data center debt is exactly the kind of asset this machine is built to absorb. The consequences of the AI bubble will likely spill into the financial system, and the pipes are already laid.
How do we position portfolios for this scenario? I have talked about this before, but I will repeat myself. It begins with humility. I really have no idea how things will play out, and nobody does. So we have been reducing our position sizes: our portfolio went from 20-25 stocks to about 30. We bought AI losers: stocks left for dead because they are not AI beneficiaries. We bought a basket of software companies. We diversified internationally into Europe and Latin America. We are limiting our exposure to the financial sector. And when we cannot find enough stocks to buy, we hold more cash.
Skeptics always sound smarter than optimists. If I had to choose, I would rather be an optimist. They are the ones who end up changing the world. As an investor, I do not have to choose either extreme, because there is a third option available to me: being a realist.
None of this is investment advice.
Key takeaways
- The data center buildout is marching toward a trillion dollars a year, and nobody knows what demand will be. Dario Amodei said it out loud: if his revenue forecast is off by 20%, or if he’s off by a single year, there is no hedge on earth that stops him from going bankrupt.
- These are not capital-light businesses — they make airlines look capital light. Companies that used to generate enormous free cash flow are now cash flow negative, issuing debt and even equity for the first time in decades.
- Everyone is bidding for the same chips, generators, and construction labor, so everyone is overpaying. Nvidia and Micron now earn margins as if they were software companies. The semiconductor business has been and will remain cyclical — this time is no different.
- Good enough is a powerful argument when it costs a tiny fraction of the alternative. Open-source and Chinese models are delivering comparable results, which sets up a race to the bottom as trillions of dollars of data centers compete for the same customers. This is exactly what happened in fiber optics after 1999.
- The 2008 part of the story is about who financed the swimmers. A large chunk of this is going into special purpose vehicles — opaque investments that regulators struggle to value, sold to pension funds and insurers — alongside incestuous vendor financing. The pipes for the damage to spill into the financial system are already laid.
yss