원문 보기
Michael Burry says the AI infrastructure boom is beginning to resemble the financial engineering that preceded 2008. JPMorgan sees something different. Unprecedented borrowing certainly, but not necessarily a credit crisis. They are looking at the same machine and reaching very different conclusions.
Burry's warning is not that artificial intelligence is fake. His concern is what sits underneath the build out. Circular financing, special purpose vehicles, off-balance sheet commitments, and increasingly complicated structures that allow enormous investment to continue without all the leverage appearing in one obvious place.
That is why Burry keeps returning to 2008. The housing crisis was not created simply because people bought expensive homes. Risk became dangerous because leverage was repackaged, redistributed, and obscured until investors could no longer see clearly who ultimately carried it. Burry argues that AI now contains its own version of that opacity.
Chip makers invest in customers. Cloud providers finance AI developers. Infrastructure vehicles borrow against future compute demand. Money can travel around the ecosystem several times before genuine end-user revenue enters the system. Oracle illustrates how complicated the funding mix has become. After generating record operating cash flow, it still produced deeply negative free cash flow as infrastructure spending surged.
It raised tens of billions in financing while customers also prepaid or supplied hardware for major AI contracts. This is the structural change that matters most. The first phase of the AI boom could largely be financed from extraordinary big tech cash flows. JPMorgan now says those cash flows are no longer sufficient to fund the full build out.
Credit markets are taking over. JPMorgan estimates the full data the build out could cost roughly $5 trillion through 2030. Around $2 trillion of that may eventually be financed through investment grade credit markets. AI is becoming as much a financing story as a technology story. The acceleration is extraordinary.
J.P. Morgan puts hyperscaler bond issuance at just $17 billion in 2024, 109 billion in 2025, and almost 200 billion in only the first half of 2026. And the publicly visible bonds are only part of the system. Data centers are also financed through bank construction loans, private placements, and special purpose structures that may sit outside traditional corporate bond indices.
That hidden scale is exactly what makes Burry uncomfortable. But this is where the 2008 analogy can become misleading. Moving financing into a separate vehicle does not make the debt imaginary, and it does not automatically imply fraud. Project finance has funded power plants, aircraft, and infrastructure for decades.
J.P. Morgan's credit team therefore reaches a very different conclusion. Yes, issuance is enormous. Yes, spreads are responding. But the biggest borrowers still have some of the strongest balance sheets in the investment grade market. Their argument is simple. Heavy debt supply is not the same thing as insolvency.
And there is evidence supporting that view. J.P. Morgan says US hyperscaler bond offerings were typically oversubscribed roughly three and a half to four and a half times during the first half of 2026. There has been no buyer strike. But the market is beginning to push back. Technology spreads now trade wider than the broader investment grade benchmark.
And one surprise hyperscaler deal attracted only around 1.4 times demand. Investors will still lend, but increasingly at a price. And here is the twist. J.P. Morgan is not uniformly bullish on AI. Its technical strategists have separately warned that parts of the AI stock rally resemble the divergence seen before the dot-com collapse.
Credit can remain solvent even while equity valuations fall dramatically. So, Burry and J.P. Morgan are not simply arguing about whether AI is expensive. They are arguing about transmission. If expectations collapse, does the damage stop with shareholders? Or has leverage become distributed widely enough that a repricing starts damaging lenders, too?
The 2008 comparison works in one important sense. Leverage can amplify mistakes when risk becomes difficult to trace. But today's system is different. AI borrowing is not centered on millions of fragile household mortgages or banks with the same pre-crisis capital structures. Burry's case becomes much stronger if several things happen together.
Compute utilization drops. Customers cannot meet long-term commitments. Special purpose vehicles need support. Sponsors honor guarantees and refinancing becomes difficult precisely as asset values are falling. That is how a credit problem could emerge. Not because one AI model disappoints, but because falling expected returns reduce collateral values.
Wider spreads make new projects uneconomic. And lenders discover that apparently separate exposures all depended on the same assumption. Permanently rising demand. J.P. Morgan wins the argument if demand keeps absorbing the supply. High utilization produces cash flow. Enterprise AI broadens the customer base.
Refinancing remains available. And investors continue buying the debt. Under that scenario, financial engineering is not hiding a crisis. It is simply financing infrastructure. Michael Burry sees the early architecture of another credit accident. JPMorgan sees an enormous capital cycle that markets can still finance.
We will only know who is right when the infrastructure starts producing mature cash flows and we discover who was really carrying the risk all along.
댓글 0