Latest / Deep Values: Value Investing Deep Dives / Follow the Money: Big Techs Two Trillion Dollar Missing AI Debt
Transcript
- 0:00Picture this $1.65 trillion. That is a lot of money.
- 0:06It really is. I mean, if you took that amount
- 0:08in crisp $100 bills and, you know, laid them end to end, you
- 0:13could wrap them around the Earth's equator 64 times.
- 0:17That is just a staggering visual to start off with.
- 0:19Right. And that figure, it's the
- 0:21combined debt issued by America's top five tech giants.
- 0:24So we're talking Amazon, Microsoft, Google, Meta and
- 0:28Oracle. But the the most compelling part
- 0:32of this isn't just the sheer volume of that debt.
- 0:34No, it's the fact that these companies, which are
- 0:37historically known as like the greatest cash generating
- 0:40machines in human history, they are actively and deliberately
- 0:43hiding it. Yeah, they are keeping it
- 0:45completely hidden from the public market.
- 0:46Exactly. So before we trace where all
- 0:48those hidden billions are flowing, a quick note for those
- 0:50of you tracking this kind of market data.
- 0:52Right. We have some resources for you.
- 0:53Yeah. To access more advanced
- 0:55researches, follow Guru's latest movement and leverage our latest
- 0:59Deep Values agents. Welcome to download Deep Values
- 1:02app from App Store. And to track stock market
- 1:06temperature, welcome to download iOS app Buffett Indicator from
- 1:09Apple App Store or Bubble Watch from Google Play.
- 1:12Links in the description and our standard disclaimer.
- 1:15You know, all analysis is based on publicly available sources
- 1:18and may be subject to revisions or differing interpretations.
- 1:21Right. All contents from this channel
- 1:23is for research and entertainment purposes only and
- 1:25does not constitute any investment or financial advice
- 1:28or service. Conduct your own due diligence
- 1:30before making investment or financial decisions.
- 1:33OK, with the housekeeping done, let's unpack this.
- 1:36Let's do it. Our mission for this deep dive
- 1:38is to explore the mechanics of shadow borrowing, figure out how
- 1:41hyperscalers are keeping a trillion dollar tab off their
- 1:44books, and look at the underlying collateral to ask
- 1:49whether we're brewing an AI subprime crisis.
- 1:51Which is a huge question. It is, but we have to start with
- 1:54the fundamental paradox here, right?
- 1:57Why do cash rich tech giants suddenly need to borrow hundreds
- 2:00of billions of dollars in secret?
- 2:02Well, the answer lies in the terrifying math of the AI arms
- 2:07race. OK, break that down for us.
- 2:09So since 2023, the operating cash flow of these five giants,
- 2:13the actual money they generate from their core software,
- 2:16advertising and cloud businesses, it's growing at a
- 2:19very healthy clip, about 23% per year.
- 2:22Which sounds great, right? 23 percent is solid.
- 2:24It is, but their capital expenditures or CapEx, the money
- 2:28they're pouring into land, power, cooling and GPU's for AI
- 2:32data centers that is growing at 70%.
- 2:35Oh wow, that divergent is basically a ticking clock.
- 2:39Exactly. If your spending grows 3 times
- 2:41faster than your income, the cash pile eventually vaporizes.
- 2:45So what happens if they just, you know, keep going at this
- 2:47rate? Projections show that if they
- 2:49maintain this current velocity, these juggernauts will
- 2:52collectively return to unprofitability by 2027.
- 2:55Unprofitable big tech That's that's insane to think about.
- 2:58By 2028, global data center CapEx is expected to hit $2.9
- 3:02trillion, and even with their massive margins, the
- 3:06hyperscalers own cash flow can cover maybe 1.4 trillion of
- 3:09that. Meaning they're staring down a
- 3:11$1.5 trillion hole that must be filled with outside capital.
- 3:15Right, and we are already seeing behavioral shifts that confirm
- 3:18this cash crunch is real. Like what?
- 3:21Well, for the last decade, tech companies didn't know what to do
- 3:24with their excess capital, so they engineered massive stock
- 3:27buybacks to reward shareholders. Right, which pumps up the stock
- 3:30price. Exactly.
- 3:32But those buybacks have essentially evaporated.
- 3:34In the fourth quarter of 2021, these five giants bought back
- 3:38$48 billion in a single quarter. OK.
- 3:42By the first quarter of 2024, that number plummeted almost 90%
- 3:47to just 4.6 billion. Wow.
- 3:49So if you're holding these stocks in your retirement
- 3:51portfolio, the whole capital return story is effectively
- 3:54paused. It is.
- 3:56Google stopped buybacks entirely in Q1.
- 3:58Meadow went three straight quarters with 0 buybacks, Oracle
- 4:02spent less than 100 million all of last year, and Amazon hasn't
- 4:05retired a single share since mid 2022.
- 4:08Because every spare dollar is being hoarded for compute
- 4:11infrastructure, right? But that brings up a glaring
- 4:14contradiction. I mean, it's like your rich
- 4:15friend who always picks up the tab suddenly asking everyone to
- 4:18split the bill while secretly applying for five new credit
- 4:21cards. That's a great way to put it.
- 4:23But wait, if Amazon or Google needs cash to build a data
- 4:26center, they are massive blue chip entities.
- 4:30They could just issue public bonds normally.
- 4:32Right, they could, but they are avoiding the standard public
- 4:35bond market for three very specific structural reasons.
- 4:39OK, what's the first one? The first is market exhaustion.
- 4:43The public debt markets can only absorb so much paper from the
- 4:46same entities before demanding higher yields.
- 4:49So people are just tired of buying Amazon debt.
- 4:52Basically, when Amazon recently issued public bonds, their
- 4:55oversubscription rate, which is the ratio of demand to supply,
- 4:59dropped from 5.3 times down to just 1.6 times in less than a
- 5:03year. Oh, that's a steep drop.
- 5:05Yeah, investors are saturated with tech debt, which forces
- 5:07Amazon to pay significantly higher interest premiums just to
- 5:11clear the issuance. OK, so borrowing publicly is
- 5:14getting more expensive. What's the second reason?
- 5:16The. Second reason feels more
- 5:17psychological, tied directly to the stock narrative.
- 5:19I mean Wall Street quality growth funds bought into tech
- 5:23stocks because they were asset light, cash rich, low debt, high
- 5:27margin businesses. Right, they sell software.
- 5:29Not, you know, heavy machinery. Exactly.
- 5:32If Meta or Microsoft suddenly dumps hundreds of billions of
- 5:36dollars of raw debt onto their balance sheet to pour concrete
- 5:39and lay fiber optics, they destroy that narrative.
- 5:42Oh I see. They transform from high margin
- 5:45software companies into capital intensive utility companies.
- 5:48And growth funds will dump the stock.
- 5:51The stock price is the ultimate currency and they will do
- 5:54anything to protect that multiple.
- 5:55Makes sense. And the third reason, the.
- 5:57Third reason is the most binding, and that's credit
- 6:00ratings. Public debt sits transparently
- 6:03on the balance sheet. Oracle has been borrowing
- 6:05aggressively defend its infrastructure, and S&P
- 6:08downgraded Oracle's debt rating to BB minus wait.
- 6:11BB minus isn't that like just one notch above junk status?
- 6:14It is exactly 1 notch above junk status which.
- 6:17Is a huge deal because that triggers strict internal
- 6:19mandates for institutional buyers.
- 6:21Oh. Absolutely.
- 6:23Pension funds and massive insurance companies have legally
- 6:26binding covenants. If a company's debt drops below
- 6:29investment grade, those funds are forced to liquidate their
- 6:32holdings. So Oracle pushed the absolute
- 6:35limit of what the public debt market will tolerate, Yes.
- 6:38So they need the cash to survive the AI race, but public debt
- 6:43ruins their stock narrative and threatens their credit ratings.
- 6:46That backs them into a corner, basically forcing them to turn
- 6:50to corporate accounting acrobatics to keep the debt.
- 6:52Hidden. Which brings us to the five
- 6:54primary strategies they used to hide a trillion dollars, the
- 6:57most prominent being the Special Purpose Vehicle, or SPV.
- 7:01Right, Since public debt is too painful, they turn to shadow
- 7:04borrowing. Let's breakdown Meta's Project
- 7:06Hyperion in Louisiana as a case study for this, because it's
- 7:09fascinating it. Really is.
- 7:10So Meta is building a massive data center facility there,
- 7:14requiring $27.3 billion in financing.
- 7:18But if you scrutinize Meta's balance sheet, you only find a
- 7:21$2.37 billion investment. Exactly.
- 7:25The other $25 billion effectively vanishes from public
- 7:28view. And they achieved this by
- 7:30isolating the liability. I mean, it's like a wealthy
- 7:33individual wanting to buy a sprawling estate, but instead of
- 7:37taking out a mortgage in their own name and ruining their
- 7:39credit score, they set up a shell company with a private
- 7:42equity firm. That's exactly how it works.
- 7:45They put the estate in the shell company's name, but secretly
- 7:48sign an. Precisely if Meta decides in
- 7:51five years that they overbuilt and they want to walk away from
- 7:54the Louisiana facility, they are legally obligated to make the
- 7:58SPV whole up to that $28 billion mark.
- 8:01So the collateral for the loan isn't just the physical data
- 8:04center, it's metas, corporate treasury.
- 8:06Yeah, and Meta pays a steep premium for this invisibility
- 8:09cloak. They're paying a 6.58% interest
- 8:13rate on this shadow debt. Instead of what, like 5.5% on
- 8:16the public market? Exactly.
- 8:18They are voluntarily absorbing hundreds of millions of dollars
- 8:22in excess interest expense every single year purely for the
- 8:25privilege of keeping the liability off their books.
- 8:27It's a massive friction cost just to manage optics.
- 8:30And, you know, Meta isn't the only one utilizing these
- 8:33structures. Let's quickly run through how
- 8:34the others play the game. Let's do it.
- 8:36So Microsoft. They've largely avoided issuing
- 8:38new public debt, but their capital lease obligations surged
- 8:42from 27 billion to nearly $63 billion, quietly hidden under
- 8:46other liabilities. Right and Google takes a
- 8:49slightly different route. They don't build or hold shares
- 8:52in these SP VS They often act as a payment guarantor.
- 8:56Third party developers guaranteeing up to $43.8 billion
- 9:01in data center loans, but they. Only book a fraction of that,
- 9:04right? Yeah.
- 9:04Google utilizes derivative accounting to mask that
- 9:07exposure. When you guarantee a loan, you
- 9:11aren't borrowing the principal. OK, so how does the math work?
- 9:14Under derivative accounting standards, Google only has to
- 9:17record the fair market value of the guarantees risk, essentially
- 9:21the statistical probability that they will actually have to pay
- 9:23out in the event of a default. Oh wow.
- 9:26So through their internal risk modeling, they've calculated
- 9:29that derivative value at just $815 million.
- 9:32Exactly. They shrink a $43 billion
- 9:35underlying liability into an $800 million footnote using
- 9:38probability math. That is wild.
- 9:41And meanwhile, Amazon and Oracle rely entirely on massive future
- 9:44leases. Amazon holds $106 billion in
- 9:47signed data center leases, and Oracle has 200 and $60 billion.
- 9:52And because the physical facilities haven't been
- 9:54completed and the leases haven't technically commenced yet, those
- 9:57massive obligations remain entirely off the active balance
- 10:00sheet. Right.
- 10:01So when you aggregate all these tactics, 445 billion in actual
- 10:05borrowed cash, 831 billion in signed but unbooked leases, plus
- 10:10all the future SPV commitments, you.
- 10:13Uncover a true underlying financial obligation of $2.13
- 10:18trillion across just five companies.
- 10:20Which forces us to ask the logical next question.
- 10:23If traditional banks aren't lending this $2 trillion, who
- 10:26is? Well.
- 10:27Private credit giants like Blackstone, KKR and Blue Owl
- 10:30eagerly stepped in. Traditional banking
- 10:32infrastructure simply cannot swallow that much highly
- 10:34customized, unstructured debt. We're not following the
- 10:37localized banking tremors we saw in 2023.
- 10:39Basil the Third Capital requirements heavily penalized
- 10:42traditional banks for holding this kind of paper.
- 10:45The capital reserves a bank would have to set aside make
- 10:47these long term data center loans wildly unprofitable.
- 10:50OK, though I know rules have recently relaxed a bit bringing
- 10:54banks back to the table, but private credit definitely filled
- 10:56that initial vacuum. They did.
- 10:59These alternative asset managers operate outside the strict
- 11:02leverage constraints of commercial banking, and what is
- 11:05fascinating is how rapidly private credit reallocated its
- 11:09resources. Right, because previously a
- 11:12massive chunk of this private lending went into leveraged
- 11:16buyouts for sauce companies. Software as a service.
- 11:18Exactly. Private credit loved sauce.
- 11:22The recurring revenue models provided incredibly stable cash
- 11:25flows to service debt. Right?
- 11:27But generative AI fundamentally threaten the sauce mode.
- 11:30Right, because why buy software if you can just use an AI agent
- 11:34to build it internally precisely?
- 11:37Anticipating this erosion, Private Credit diverted their
- 11:39capital firehose away from software buyouts and aimed it
- 11:43directly at the physical infrastructure powering the AI
- 11:45models. OK, but this leads to the most
- 11:47precarious part of this whole ecosystem.
- 11:49Usually when you deploy billions in private credit, you borrow
- 11:52against real estate, right? A building the last 30 years,
- 11:54right? But here they are borrowing
- 11:56against GPU's. How can a graphics card be
- 11:58treated like a house? This is the most fascinating
- 12:01shift in the market. Wall Street is now actively
- 12:04underwriting loans where the primary collateral is the
- 12:07microchips inside the building. Nvidia's Jensen Huang is
- 12:12aggressively lobbying to establish AI compute as a
- 12:15distinct investable asset class. Entirely separate from the real
- 12:19estate that houses it, right? They are partnering with
- 12:22alternative asset managers to construct specialized financing
- 12:25vehicles purely for GPU's but. Hardware is risky.
- 12:28How do they convince lenders to accept silicon as collateral?
- 12:31NVIDIA is stepping in to offer up to 25% residual value support
- 12:36to backstop the loans. Wait, they are effectively
- 12:39writing a put option on their own hardware?
- 12:41Yes, NVIDIA guarantees lenders that if a data center operator
- 12:44defaults and the secondary market value of those GPU's
- 12:48collapses, NVIDIA will step in and buy the hardware back,
- 12:51establishing a hard floor of $0.25 on the dollar.
- 12:54That floor drastically alters the risk calculus for private
- 12:57credit. It does.
- 12:58Because of that perceived safety net, the interest rates on loans
- 13:01backed purely by compute hardware have plummeted from 15%
- 13:05down to 5.9% in roughly a year. Wall Street is eagerly treating
- 13:09rapidly depreciating silicon with the same financial
- 13:12reverence as commercial real estate, Which?
- 13:14Requires aggressive accounting fiction to sustain.
- 13:17Because real estate appreciates, but computing hardware follows
- 13:21Moore's law, I mean. And NVIDIA H100 GPU loses
- 13:25approximately 55% of its resale value by year 3.
- 13:28By Year 5, it's commercially obsolete.
- 13:31O how do they hide this brutal reality?
- 13:33Tech companies are playing accounting games, stretching the
- 13:36math. Microsoft recently announced
- 13:39they are extending the depreciation schedule of their
- 13:41servers and data center equipment from 15 years to 25
- 13:45years. Extending depreciation to 25
- 13:48years on computing hardware is absurd.
- 13:50We're talking about servers deployed today remaining on the
- 13:53books until 2051. You're.
- 13:55Essentially taking out a 20 year mortgage on an iPhone.
- 13:57Seriously. And short sellers estimate that
- 14:00this industry wide accounting trick obscures hundreds of
- 14:03billions of dollars in true capital consumption.
- 14:06It dramatically reduces the annual depreciation expense they
- 14:08have to report on their income statement, artificially
- 14:11inflating their current operating margins.
- 14:13OK, so let's zoom out and look at the totality of the system.
- 14:17Trillions in hidden debt routed through special purpose
- 14:20vehicles, hyper depreciating assets treated as real estate
- 14:24collateral, and Wall Street funds aggressively trading this
- 14:28debt back and forth. It sounds familiar, doesn't it?
- 14:30This sounds uncomfortably like the 2008 housing crisis.
- 14:33Are we looking at an AI subprime bubble?
- 14:36The parallels in the financial engineering are undeniably
- 14:39stark. We see complex securitization,
- 14:42off balance sheet vehicles and yield chasing behavior and the
- 14:45market is actively pricing in elevated risk.
- 14:48Yeah. If you look at Oracle's credit
- 14:50default swaps, or CDs, essentially default insurance,
- 14:53the cost recently spiked to 203 basis points.
- 14:57That is the highest level of risk premium attached to Oracle
- 15:00since the depths of the 2008 financial crisis.
- 15:02It's. A huge warning sign and.
- 15:04You also see concerning levels of interconnection.
- 15:06The funds buying this debt like PIMCO are highly interconnected
- 15:09with the issuers like Blue Owl, creating a closed loop and the
- 15:13ultimate money funding. This comes from everyday
- 15:15people's pension funds. True, but financial
- 15:17professionals strongly push back against the 2008 systemic
- 15:21collapse narrative, bringing a more calm, analytical
- 15:24perspective. They point to three main reasons
- 15:26this isn't 2008. OK, what's the first one?
- 15:29The. 1st is sheer scale. Despite the massive numbers, the
- 15:32corporate data center debt market is still small compared
- 15:35to the 2008 US residential housing market.
- 15:38It's not the foundational bedrock of consumer wealth.
- 15:41So the blast radius is narrower. Exactly The second reason
- 15:45centers on who holds the bag. In 2008, heavily leveraged
- 15:49systemic commercial banks held the risk.
- 15:52When assets failed, it threatened the global clearing
- 15:54system. But today?
- 15:55Because of Basel, the third, commercial banks are largely
- 15:58insulated. The risk sits with private
- 16:00credit and pension funds. Right.
- 16:02Meaning a collapse is less likely to drag down the entire
- 16:04global banking system. It destroys wealth, but it
- 16:07doesn't break the plumbing. OK, but how do they explain away
- 16:10the spike in Oracle's default insurance?
- 16:12Credit analysts view that as a logical repricing, not a panic.
- 16:16The market is realizing Oracle is transforming from an asset
- 16:20light cash flow heavy software company into a capital intensive
- 16:24infrastructure builder. Like an oil company?
- 16:27Exactly. Industrial utilities carry
- 16:29structurally higher debt loads and demand higher risk premiums.
- 16:32The CDs spike is just the market pricing Oracle as utility.
- 16:36OK. Bringing it back to you, the
- 16:37listener, ultimately the market doesn't care about accounting
- 16:40tricks, cares about cash flow, right?
- 16:42Precisely. Microsoft stock recently jumped
- 16:459%, not because of their 25 year depreciation trick, but because
- 16:50their AI cloud revenue actually grew by 43%.
- 16:54The enterprise demand was tangible.
- 16:56So as long as the physical infrastructure generates
- 16:58sustainable real world revenue, the operating cash flow can
- 17:01service the debt regardless of how convoluted the SPV
- 17:04structures become. That's.
- 17:06The ultimate test. So just to recap this journey,
- 17:08we are witnessing Big Tech's transition from cash rich
- 17:11software companies into the most heavily indebted infrastructure
- 17:14builders in modern history. They're using clever SPV's and
- 17:17shadow leases to hide the tap. And Wall Street is willingly
- 17:21facilitating this by treating depreciating use like 30 year
- 17:26mortgages. It's a master class in financial
- 17:28engineering, but. I want to leave you with the
- 17:29final thought to Mull over as you watch these hyperscalers
- 17:33continue to spend all of this complex financial engineering.
- 17:37The SPV's, the shadow borrowing, the 25 year depreciation
- 17:41schedules. It relies on one single
- 17:44assumption that the. Demand for AI compute will grow
- 17:47exponentially forever. Exactly.
- 17:49But what happens if the timeline for artificial general
- 17:51intelligence takes a decade longer than Silicon Valley
- 17:54thinks if. The AI revenue slows down.
- 17:57These trillion dollar financial structures aren't backed by land
- 18:00or homes. They're.
- 18:01Backed by outdated microchips, if that cash flow falters, that
- 18:05$1.65 trillion in debt doesn't just wrap around the earth, it
- 18:09becomes a very heavy chain around the neck of the tech
- 18:11industry. Thanks for joining us on this
- 18:13deep dive. Keep questioning the numbers
- 18:15behind the headlines and we'll see you next time.