Latest / Deep Values: Value Investing Deep Dives / How Big Tech Hides AI Infrastructure Debt?
Transcript
- 0:00Silicon Valley tech giants are on track to spend a staggering
- 0:05$700 billion on AI infrastructure this year alone.
- 0:09Yeah, it's a massive number. Just completely unprecedented,
- 0:13right? But to pay for that build out,
- 0:15they aren't just dipping into their famous cash reserves.
- 0:18They're actually actively dusting off the exact same off
- 0:21balance sheet financial engineering that, you know,
- 0:24brought down Enron. And fueled the 2008 financial
- 0:27crisis, Yeah. Exactly.
- 0:29So welcome to today's deep dive. We are looking directly at the
- 0:34hidden financial plumbing holding the AI boom together for
- 0:37you today. And it is a fascinating topic
- 0:39because when you look past the shiny front end of what AI can
- 0:43do and focus strictly on how it's being financed, well, the
- 0:46whole ecosystem starts to look intensely fragile.
- 0:49It really does. And today we're unpacking A
- 0:52incredibly sobering analysis from Steve Eisman.
- 0:55For anyone following the markets, you know Eisman is the
- 0:57legendary investor who notoriously shorted the housing
- 1:00market before the great financial Crisis.
- 1:02Right, he's the guy who saw the cracks before anyone else did.
- 1:05Exactly. And he isn't spending his time
- 1:07evaluating like large language models or neural networks.
- 1:11He is looking squarely at the balance sheets.
- 1:13Because that's where the real story is.
- 1:15He's identifying structural parallels to past financial
- 1:18disasters that are being deployed right now, today, by
- 1:21the biggest companies in the world.
- 1:23So the thesis here is essentially that Enron era
- 1:25accounting tricks are back with a vengeance.
- 1:28But before we get into the hidden debt, we kind of need to
- 1:31understand the macro pressure cooker that's forcing tech into
- 1:34these shadowy corners, right? Yeah, we have to understand why
- 1:37they're suddenly so desperate for cash.
- 1:39Because it wasn't that long ago that the narrative around Big
- 1:42Tech was that they were basically printing money.
- 1:45I mean, they had so much free cash flow, they were acquiring
- 1:47startups just to kill them. Right, but that era is
- 1:50definitively over. To understand why tech is
- 1:53strapped today, we have to ground ourselves in the current
- 1:55macro reality. Eisman distills this down to
- 1:58just two primary variables that matter right now.
- 2:00OK, what are they? Oil prices and the 10 year
- 2:04Treasury yield and the 10 year yield is the critical 1.
- 2:07He points to a 5% yield on the 10 year as the market's Rubicon.
- 2:12The Rubicon meaning if we cross it and stay there, there's no
- 2:16going back. Exactly.
- 2:17If it stays consistently above 5%, a massive market correction
- 2:21is imminent. And we know why that 5%
- 2:24threshold is so important, right?
- 2:25When the risk free rate sits that high, borrowing just
- 2:28becomes prohibitively expensive. It completely crushes the equity
- 2:32risk premium. Right, the weighted average cost
- 2:34of capital skyrockets. So for tech companies that are
- 2:37used to near 0 interest rates, they just can't sustain their
- 2:41historic growth rates. Precisely, And that high cost of
- 2:44capital is slamming head first into the reality of AI
- 2:47development. We are transitioning out of the
- 2:49capital light tech era. The capital light era meaning
- 2:54like software. Yeah, for 20 years, software was
- 2:56a margin machine. You write the code for a social
- 2:59platform once, and scaling it to a billion users is relatively
- 3:02cheap, right? AI breaks that model completely.
- 3:06AI is heavy industry. It requires massive physical
- 3:08hardware, acres of data centers, cooling infrastructure, and just
- 3:12gigawatts of electricity. Which brings us back to that
- 3:15$700 billion figure, the hyperscalers.
- 3:18So Amazon, Google, Microsoft, Meta, Oracle, they are
- 3:22projecting massive capital expenditures.
- 3:25They have to. They have no choice.
- 3:27And on top of that, NVIDIA is raising chip prices by like 15%
- 3:32because they have absolute pricing power right now.
- 3:34Right, They're the only game in town for the chips you need, so
- 3:37the tech giants just have to pay whatever NVIDIA asks.
- 3:41So the free cash flows they used to rely on are just being
- 3:44completely swallowed by physical infrastructure costs.
- 3:47Which forces them into the debt markets at exactly the wrong
- 3:50time. And we're already seeing the
- 3:52cracks in the foundation here. Take Oracle for example.
- 3:55Oh yeah, the downgrade. Yeah, they recently took on so
- 3:58much leverage to fund their AI data centers that Standard and
- 4:01Poor's actually downgraded their credit rating to BBB minus.
- 4:04Which is an incredibly precarious place to be.
- 4:07I mean, BBB minus is just one single step above junk status.
- 4:11Right, it's the bottom rung of investment grade.
- 4:14And if they get downgraded one more notch to junk, massive
- 4:18institutional investors and pension funds are legally
- 4:20mandated to dump their bonds. It creates A catastrophic
- 4:24liquidity crunch. Exactly.
- 4:25It's a huge warning sign, and if Oracle is a warning sign,
- 4:28SoftBank is a flashing red siren right now.
- 4:31Oh man, the SoftBank move is wild.
- 4:33Yeah, Masayoshi Son is reportedly raising $11 billion
- 4:37in junk bonds to fund a venture style equity investment in Open
- 4:40AI. Let's just pause on that because
- 4:42the mechanics are insane. He is raising highly speculative
- 4:45junk debt to take an equity position in a private AI
- 4:49company. Right?
- 4:50You have fixed income high yield liabilities directly funding an
- 4:54illiquid, high risk venture play.
- 4:56It is playing with fire on a tightrope, but it illustrates
- 5:00this extreme tolerance for risk. They are driven entirely by this
- 5:03desperate need to not fall behind in the AI arms race.
- 5:07They need the capital at all costs.
- 5:09Exactly. But if they issue straight debt,
- 5:11they ruin their credit ratings, just like Oracle almost did.
- 5:13So they have to find another way.
- 5:15Right. They can't afford to hit to the
- 5:16balance sheets. So if you're a tech giant and
- 5:19you need billions for data centers but can't scare the
- 5:22rating agencies, how do you? Get it?
- 5:23You hide it. You hide it, yeah.
- 5:25You use off balance sheet financing, specifically special
- 5:28purpose vehicles, SPV's or variable interest entities VI
- 5:33ES. And this is where Eisman's
- 5:35historical parallels come in, right?
- 5:37Absolutely. To understand how they're doing
- 5:39this today, you have to look at how it was abused in the past.
- 5:41Look at Enron in the late 90s. Right, Enron.
- 5:44They needed to mask their massive debt load to keep their
- 5:47investment grade rating, so their CFO Andrew Fastow
- 5:50engineered these SPV's, the famous LJM partnerships.
- 5:54And Enron basically just transferred its toxic debt into
- 5:57these external shell companies that Fastow controlled.
- 6:00Exactly. The ratings agencies who really
- 6:02only look at the official public facing balance sheet, saw Enron
- 6:05as this healthy, low leverage energy giant.
- 6:08Well, the auditors, Arthur Andersen, just looked the other
- 6:11way and signed off on the whole charade.
- 6:13Right. The debt didn't magically
- 6:15disappear, it was just legally cordoned off.
- 6:17Which brings us to the great financial crisis.
- 6:20Right. Fast forward to 2008 and the
- 6:22banking sector pulled a very similar maneuver with what they
- 6:25called Sivs, structured investment vehicles, right?
- 6:28These were basically shadow banks created by the major
- 6:31commercial banks, deliberately kept off the main balance sheets
- 6:34to bypass regulatory capital requirements.
- 6:37The mechanics of those 2008 Sivs are fascinating, especially that
- 6:41concept of what is it called surfing the curve.
- 6:44Let's break that down, because it perfectly illustrates the
- 6:46liquidity risk today. Yeah.
- 6:48So surfing the curve is fundamentally an arbitrage game.
- 6:51It's built on a duration mismatch.
- 6:54These Sivs would fund themselves by issuing asset backed
- 6:57commercial paper. So they're borrowing money on a
- 7:00very short term basis, right? Like overnight or 30 days?
- 7:0390 days Max. Exactly.
- 7:04Borrowing short term at extremely low interest rates
- 7:07from institutional investors and they took that short term cash
- 7:10and bought long term illiquid, higher yielding assets.
- 7:14The subprime mortgage-backed securities.
- 7:16Yep, and they profit on the spread between the two.
- 7:19But the fatal flaw is that the 30 day commercial paper matures
- 7:22constantly. The SIV has to reissue it.
- 7:25They have to roll the debt over every single month just to stay
- 7:28solvent. So the entire structure assumes
- 7:30infinite uninterrupted liquidity from the market.
- 7:33Exactly. It's essentially like it's like
- 7:36maxing out a series of short term credit cards to buy a 30
- 7:39year mortgage, but right before the bank runs your credit check,
- 7:42you legally transfer all that credit card debt into your dog's
- 7:45name. Yes, your credit profile looks
- 7:48pristine, the debt disappears. Poof, magic.
- 7:51It's a great analogy for the obfuscation part of it.
- 7:54But let's look at what happens when the mechanism breaks down.
- 7:57When the underlying asset, those subprime mortgages, started
- 8:00defaulting, the institutional investors panicked.
- 8:03They realize the Sivs were stuffed with toxic garbage.
- 8:06Right. So when the 30 day paper
- 8:09matured, the investors just flat out refused to buy the new
- 8:11paper. The rollover stopped.
- 8:14The liquidity completely evaporated.
- 8:17The dog can't pay the credit card bill.
- 8:18Exactly, and because of the major commercial banks had
- 8:21sponsored these Sivs, they were ultimately on the hook.
- 8:24They were forced to step in and finance those long term toxic
- 8:27assets themselves. So overnight, all of that hidden
- 8:30shadow bank debt boomeranged straight back onto the main
- 8:33balance sheets of the big banks. The poof magic reversed and it
- 8:38blew a hole in the global financial system.
- 8:41Which brings us to the present day.
- 8:43Eisman's warning is that these exact vehicles, these SPV's, are
- 8:47being actively used by Silicon Valley right now to build AI
- 8:50data centers. The playbook has been completely
- 8:53dusted off. OK, let's look at a concrete
- 8:55example. Meta is currently building a
- 8:57sprawling $27 billion data center project down in
- 9:00Louisiana. Right, a massive facility.
- 9:02And they didn't just issue corporate bonds to pay for it,
- 9:05they set up an SPV. They called it very cutely
- 9:08Beignet Investor LLC Meta technically only owns 20% of
- 9:13this SPV financial firms like Blue Owl and PIMCO own the
- 9:16remaining 80%. Right, and that 8020 ownership
- 9:19split is exactly how they justify keeping the debt off
- 9:22Meta's main balance sheet. But when you dig into the actual
- 9:25mechanics of the agreement, the risk profile tells a totally
- 9:28different story. That's the crux of it.
- 9:30Meta isn't just a passive 20% owner.
- 9:33They signed a 20 year lease agreement to rent the data
- 9:36center from the Beignet SPV. Yep, and furthermore, Meta
- 9:40explicitly bears the cost of any project delays or budget
- 9:43overruns. They are holding all the
- 9:45residual risk and they are guaranteeing the cash flow for
- 9:47two decades. But under current variable
- 9:50interest entity accounting rules, that massive liability
- 9:53doesn't sit on Meta's balance sheet.
- 9:55It's just wild. And it's not just meta.
- 9:57Oracle is executing a nearly identical structure with a
- 10:00project called Jupiter also partnering with Blue Owl, Right?
- 10:03And it's really worth noting here that the debt tied to
- 10:06Oracle's project, Jupiter SPV, is already trading at a stress
- 10:10level. It's trading at around $0.90 on
- 10:12the dollar due to force majeure construction delays.
- 10:15Wow, so the stress in these hidden vehicles is already
- 10:18materializing in the credit markets?
- 10:19Already happening, yeah. OK, but if the risk is so
- 10:22obvious and we've all lived through Enron in 2008, how are
- 10:26the auditors allowing this to happen again?
- 10:28Well, this is where it gets interesting.
- 10:30If you look at Meta's 10K financial filing, Ernst and
- 10:33Young, their auditor, actually includes a note labeling this
- 10:36exact arrangement a critical audit matter.
- 10:39Critical audit matter. Yes, and this requires us to
- 10:42really read between the lines of corporate accounting jargon.
- 10:45The rules post Enron basically revolve around who has the quote
- 10:48UN quote power to direct the activities of the entity.
- 10:51Right. Who's actually in control of the
- 10:52SPV? Exactly, and in their 10K, Ernst
- 10:55and Young writes that auditing this beignet SPV was especially
- 10:59challenging due to the significant judgment required
- 11:02and assessing which party has that power.
- 11:05So they're hedging. Eisman argues that if we
- 11:07translate that corporate speak into plain English, Ernst and
- 11:10Young is basically waving a massive red flag.
- 11:13They are saying, look, we are extremely uncomfortable with
- 11:16this transaction. We aren't entirely sure this
- 11:18debt shouldn't be on Meta's main balance sheet.
- 11:20But Meta is a massive client. They provided their own internal
- 11:24assessment and we are relying on their description to avoid
- 11:27rocking the boat. It it reads like an auditing
- 11:29firm desperately trying to paper its own files to avoid becoming
- 11:33the next Arthur Andersen. That's exactly how Eisman reads
- 11:36it. I got to push back a little here
- 11:38though. Let's play devil's advocate for
- 11:40a second. Meta is a trillion dollar
- 11:42company. They generate 10s of billions in
- 11:44free cash flow every single year.
- 11:46Sure, yes, $27 billion for the Beignet SPV is a ton of money,
- 11:51but relative to Meta's total enterprise value, is this really
- 11:55a systemic risk? We aren't exactly talking about
- 11:58a Lehman Brothers sized hole here, right?
- 12:00It's a vital question, and Eisman addresses it by pointing
- 12:03out that it isn't the absolute dollar amount of this single SPV
- 12:06that matters. It's the precedent.
- 12:08It's the behavioral pattern. OK, He references a specific
- 12:11scene from that 1961 film Judgment at Nuremberg.
- 12:15There's a moment where a defense attorney is brutally cross
- 12:18examining a witness using the exact same abusive tactics
- 12:21common in Nazi courtrooms years earlier.
- 12:23Right. And one of the defendants, a
- 12:25former judge, stands up in horror and shouts hair off.
- 12:28Are we going to do this all over again?
- 12:30Wow. So it's about the normalization
- 12:32of the behavior. Exactly.
- 12:34If Mehta is utilizing these SPV's and Oracle is utilizing
- 12:37them, and the major auditing firms are just signing off on
- 12:39them with a nervous shrug, then that 27 billion in Louisiana is
- 12:43just the very tip of a much, much larger industry wide
- 12:47iceberg of shadow leverage. Right, because everyone's going
- 12:49to start doing it. And if you hold an S&P 500 index
- 12:52fund, your portfolio is heavily weighted toward these tech
- 12:55giants, which means you are unknowingly exposed to this
- 12:58hidden off balance sheet debt. It's the return of a financial
- 13:01culture that prioritizes hiding leverage to maintain a stock
- 13:05multiple over transparent accounting that is the systemic
- 13:08risk. Exactly.
- 13:10Which forces us to ask, I think, the most critical question of
- 13:13all. We have this colossal physical
- 13:15infrastructure being funded by shady hidden debt that depends
- 13:19on uninterrupted corporate guarantees.
- 13:21But what happens when the actual AI products hit the market?
- 13:25Is the consumer application of this technology lucrative enough
- 13:28to generate the massive ROI needed to justify all this
- 13:32financial engineering? This is where the macro analysis
- 13:35collides with the consumer reality, and the irony here is
- 13:38just palpable. Let's look at agentic AI.
- 13:41Meta has an autonomous agent called Muse, right?
- 13:44Unlike a a standard large language model where you type a
- 13:47prompt and get a text summary and agentic, AI actually
- 13:50executes tasks autonomously on your behalf.
- 13:53It takes actions. Exactly.
- 13:56The example given is a user deploying Muse to buy theater
- 13:58tickets. A human would have given up
- 14:00after checking like 2 sites, but Muse just autonomously scoured
- 14:05the web, haunted for secure payment links, and eventually
- 14:07procured cheaper tickets that the human user never would have
- 14:10found. Which on an individual level, is
- 14:13fantastic utility for the consumer.
- 14:15But let's extrapolate the macroeconomic ripple effects of
- 14:18that behavior. OK, if AI agents begin
- 14:21autonomously scouring the web to book your hotel rooms or your
- 14:24international flights, completely bypassing traditional
- 14:27aggregator platforms, what happens to the digital moats of
- 14:30companies like Booking or Expedia?
- 14:33They get entirely disintermediated.
- 14:36A search engine monetizes attention and clicks, but an
- 14:39autonomous agent optimizes for utility without you ever even
- 14:43seeing the screen right. If the AI buys the ticket, who
- 14:46clicks the ad? Nobody clicks the ad, and that
- 14:50obliteration of the traditional Moat extends perfectly into
- 14:52brand loyalty to the Nike sneaker example in the source
- 14:55illustrates this beautifully. Right, so let's say you are a
- 14:59historically loyal Nike runner, but now you have an AI health
- 15:03agent. You tell it, hey, I need a
- 15:04running shoe perfectly optimized for my foot arch and my specific
- 15:08running gate to prevent knee pain.
- 15:10The AI does not care about the Nike swoosh.
- 15:12No, it analyzes your biometric data, cross references it with
- 15:16material science data and says actually this specific shoe from
- 15:19Brooks is 15% better for your bio mechanics and it happens to
- 15:23be 20 bucks cheaper. The brand premium completely
- 15:26evaporates if AI perfectly optimizes for consumer utility
- 15:30based on raw data. Brand loyalty dies.
- 15:34But here is the ultimate tragic irony for the tech giants.
- 15:37How does Meta generate the 10s of billions of dollars in free
- 15:41cash flow they need to pay off that $27 billion SPV data
- 15:44center? They sell advertising to brands
- 15:47like Nike. Exactly.
- 15:49If Nike pays Meta millions of dollars for targeted ad space,
- 15:52but Meta's own AI agent intercepts the consumer and
- 15:55directs them to a competitor based on pure utility, the AI is
- 15:59actively cannibalizing the exact advertising revenue that Meta
- 16:03needs to survive. They are essentially funding
- 16:06their own disruption. The tech giants are taking on
- 16:08massive hidden off balance sheet debt to build out an
- 16:11infrastructure that is designed to perfectly optimize consumer
- 16:14choice, which inherently destroys the attention based
- 16:17advertising business models that currently sustain their
- 16:20valuations. It is a literal race to the
- 16:22bottom, financed by shadow debt. It's like going into massive
- 16:26hidden debt to build a trillion dollar machine whose only real
- 16:29function is to eat your own wallet.
- 16:31That's one way to put it, yeah. OK, let's pull all of this
- 16:34together for you. We started with the tech
- 16:37industry that is desperately trying to maintain the illusion
- 16:39of infinite capital light cash flow while simultaneously
- 16:43spending hundreds of billions on physical infrastructure.
- 16:46And to protect their credit ratings, they are resurrecting
- 16:48Enron style SP VS to shift that massive debt off their balance
- 16:52sheets, relying on auditors who are frankly too afraid of losing
- 16:55their biggest clients to stop it.
- 16:57All to fund consumer AI agents that are highly likely to
- 17:01destroy their core revenue engines.
- 17:03It is, quite literally, a House of Cards built on top of a
- 17:06server rack. But before you wrap up today,
- 17:08there's one final critical piece of this puzzle that builds on
- 17:11everything we've discussed. Oh.
- 17:13Let's hear it. We know that these SP VS like
- 17:15beignet investor, they only survive because the tech giants
- 17:19are guaranteeing the 20 year leases and covering the cost
- 17:21overruns, right? That financial structure assumes
- 17:24the data centers will eventually get built, turn on, and generate
- 17:28processing power. But what happens if they hit a
- 17:32physical bottleneck that no amount of financial engineering
- 17:34can solve? You mean the power grid?
- 17:37Specifically, the global energy grid.
- 17:40We are already seeing reports that the grid simply cannot
- 17:42generate or transmit enough gigawatts to power all these
- 17:45projected AI data centers. That is the ultimate hard
- 17:48constraint. And raises A terrifying
- 17:50scenario, right? If the power isn't there, the
- 17:53data center stall, they can't operate.
- 17:55They can't operate. They generate 0 revenue.
- 17:58And if they generate no revenue, the SP VS collapse.
- 18:00Exactly. And if those SP VS collapse,
- 18:03those 20 year lease guarantees are called in, and all of that
- 18:06hidden massive shadow debt boomerangs right back onto the
- 18:10balance sheets of Meta, Oracle, and Microsoft.
- 18:13Exactly like the shadow banks in 2008.
- 18:15Yep. And if that happens, we might
- 18:17suddenly find out if the Too big to Fail doctrine that the
- 18:21government invented to save Wall Street now applies to Silicon
- 18:24Valley. Which is exactly why you always
- 18:27have to read the footnotes. Always read the footnotes
- 18:30because the real story is never in the press release.
- 18:33Thanks for joining us on this deep dive.
- 18:35We'll catch you next time. None.
- 19:38I'd like to know that.