NeoClouds: The Junk-on-Junk Structure That Is Cracking AI Finance

Something cracked in AI finance this month. It was not the technology. It was not the demand. It was not NVIDIA, ASML, TSMC, or any of the companies that actually build the infrastructure of the next decade. What cracked was a specific financing model — opaque, highly leveraged, and structurally identical to the subprime mortgage structure that imploded in 2007.

To understand what happened, you need to understand the difference between two groups that the media has been calling "AI companies" as if they were the same thing. They are not.

Two Worlds Inside "AI"

The first group is the hyperscalers: Microsoft, Alphabet, Amazon, Meta. These companies own the cloud. They own the applications. They own the user base. Together, they generate approximately $1.2 trillion in annual operating cash flow. When they issue debt, it carries AAA or AA credit ratings. They are not in trouble. Their Q3 results this month confirmed it: ASML, which supplies the lithography machines that make advanced chips possible at all, reported revenue up 55% and earnings per share up 136%. Its CEO said projected supply for 2027 and 2028 may not cover all the demand already placed. TSMC confirmed the same direction — the market is evolving from chatbots to autonomous AI agents, and the infrastructure buildout is accelerating.

The second group has a different name, though few people use it: NeoClouds. These are former Bitcoin mining companies that repositioned themselves as AI compute providers. CoreWeave, Iris Energy, Nebius, Galaxy Digital, Applied Digital, Cypher. Their business model: borrow money, buy NVIDIA GPUs, rent computing power by the hour to whoever will pay. They do not own the software. They do not own the customer relationships. They do not generate free cash flow.

That distinction matters enormously right now, because it is the NeoClouds — not the hyperscalers — that are breaking.

The Structure That Should Not Have Existed

JPMorgan's infrastructure analysts put a number on the problem this week. For a NeoCloud to generate positive free cash flow, it needs approximately four to five times more operational computing capacity than capacity currently under construction. The current ratio across the sector is 0.9x. That means the industry, in aggregate, has slightly less operational capacity than it has in construction — the opposite of what profitability requires. JPMorgan's conclusion: positive free cash flow for NeoClouds arrives, at the earliest, in 2030.

The market for NeoCloud debt already reflected this conclusion before the analysts published it. CoreWeave's bonds maturing in 2032 are trading below 80 cents on the dollar, implying a yield above 15%. Its five-year credit default swaps have exceeded 1,100 basis points, pricing in a probability of default above 60% over five years. For context: Lehman Brothers' CDS were trading at approximately 700 basis points the week before the firm filed for bankruptcy.

But the structure goes one level deeper, and this is where the 2007 parallel becomes uncomfortable.

There are approximately $14 billion in junk bonds financing data center construction projects whose sole tenant — the only company paying rent — is CoreWeave. CoreWeave itself is financed entirely by junk bonds. A building financed by junk debt, occupied by a company financed by junk debt. The bond market has a phrase for this kind of structure. It is not a compliment.

In 2007, the equivalent structure was mortgage-backed securities containing tranches of subprime loans that were themselves backed by other subprime assets. The rating agencies called them investment grade. The market eventually called them what they were. The process of that recognition caused the largest financial crisis since 1929.

The NeoClouds will not cause a crisis at that scale. The dollar amounts are different. The systemic connections are different. But the structural logic — layered junk, opacity, and a financing model that only works if the underlying asset appreciates indefinitely — is the same.

The Amplifier Hidden in Plain Sight

There is a second mechanism driving the severity of the current correction that has nothing to do with AI specifically.

Margin debt — money borrowed from brokers to buy securities — has reached $1.45 trillion in the United States, up 37% from a year ago. Leveraged ETFs that multiply the daily return of an index by two or three times have accumulated $218 billion, up 60% since March. Leveraged semiconductor ETFs alone have grown 175% in the same period.

When an asset falls 5%, the holders of a 3x leveraged ETF experience a 15% loss. At a certain loss threshold, brokers issue margin calls: pay now or we liquidate. Liquidation at scale pushes prices lower. Lower prices trigger more margin calls. The spiral is mechanical, not emotional.

This is the technical explanation for why the VIX — the measure of expected volatility in the S&P 500 — rose 31% in a week while the S&P 500 itself fell only 2%. The underlying index barely moved. The leveraged positions around it moved violently. The volatility index was measuring the leverage, not the fear.

What This Means for the Companies That Actually Matter

Michael Burry moved his timeline for a severe correction from 2028 to the next twelve months this week. Ray Dalio, speaking in Singapore, said the AI boom has all the characteristics of a classic bubble. Both are right about the NeoClouds. Both are worth listening to carefully. But neither observation, if correct, implies that ASML's order book deteriorates, that TSMC's customers stop buying chips, or that Meta's 3.2 billion daily active users stop using its platforms.

The S&P 500 is trading at approximately 19 times forward earnings. The top ten stocks in the index are at 20.9 times. These are not the valuations of irrational exuberance. The dot-com peak was 40 times. The current NeoClouds trade at valuations that cannot be computed, because they have no earnings.

Bank of America's analysis of hyperscaler capital expenditure projects a valley — a period of reduced free cash flow — between now and 2027, followed by acceleration. The hyperscalers will absorb this valley with their existing cash generation. The NeoClouds with negative cash flow, junk-rated debt, and 2030 profitability targets will not.

The correction happening now is real. The rotation is real. The pain in semiconductor stocks — the SOXX index recorded nine consecutive down sessions, something not seen since 2006 — reflects genuine repricing of risk that was mispriced at the margin. But the correction is happening inside a sector, not across an economy. The companies that build the actual infrastructure of AI, the ones that supply the machines, the chips, and the platforms, are reporting the strongest fundamentals in their histories.

The Question Worth Sitting With

Firmus, an Australian NeoCloud backed by NVIDIA and Blackstone, cancelled its planned $30 billion IPO this month. The cancellation happened because it proposed to issue $30 billion in debt against projected 2028 earnings that were six times smaller. Investment funds rejected the offering. The Broadcom CEO said publicly that a significant mass of investors is directing capital toward vehicles they do not understand at all.

AirTrunk, a data center REIT, cancelled its Singapore listing for related reasons.

These cancellations are not the beginning of the end of artificial intelligence. They are the market doing its job: withdrawing financing from structures that should not have received it. The technology is real. The demand is real. The applications are real. The question was always whether the financing was real.

For the NeoClouds, it was not. That answer is now arriving. The hyperscalers, the chip designers, and the semiconductor equipment manufacturers are a different question — and the data from their Q3 earnings this month provided a different answer.

---

CongressFlows tracks unusual trading patterns in congressional stock disclosures and connects them to macro developments. Nothing in this post constitutes financial advice.