The Smart People Have Started Looking for the Exit — Transcript
Full transcript
- 0:00Last month, the Bank for International
- 0:01Settlements, which is basically the
- 0:03central bank for central banks,
- 0:05published a study of more than a
- 0:06thousand AI companies. The study focuses
- 0:09on one question. Using new data on AI
- 0:11firms, investment deals, and supply
- 0:13chain linkages, the study examined
- 0:15investment patterns among the largest AI
- 0:17firms globally, focusing on circular
- 0:19investment relationships. What they
- 0:21found is that between 2021 and 2025, 55%
- 0:26of the investment money flowing into AI
- 0:28companies came from other AI companies.
- 0:30So more than half the money that's
- 0:32funding the AI bubble is coming from
- 0:34inside the bubble, and it only gets
- 0:35messier as the checks get bigger. Some
- 0:38of these investors are also doing
- 0:39business with the companies they invest
- 0:41in, like a chip maker putting money into
- 0:43an AI company and that same AI company
- 0:45turning around and buying its chips.
- 0:47Deals like that make up only about one
- 0:49in six of the deals between AI
- 0:51companies, but when you look at the size
- 0:52of those deals, they're nearly half of
- 0:54all the money, which means the bigger
- 0:56the check, the more likely the investor
- 0:58is also the supplier or the customer.
- 1:00The study calls these relationships
- 1:02circular. The Bank for International
- 1:04Settlements flagged the problem in one
- 1:05very polite sentence. These circular
- 1:07arrangements tend to be opaque, making
- 1:09monitoring and supervision difficult,
- 1:12which is basically the same as the
- 1:13referees telling you they can't see the
- 1:14field. And until recently, that's where
- 1:17this story stopped. Everybody knew about
- 1:18the circular financing problem in AI,
- 1:21but it stayed mostly a so what type of
- 1:24hypothetical, the kind of thing that
- 1:25gets a thoughtful nod at a conference
- 1:28right before everyone heads to lunch.
- 1:30But recently, that's changed. Now we're
- 1:32seeing people starting to react to the
- 1:34risks in the AI bubble, and you can
- 1:35actually watch them react because
- 1:37there's one corner of Wall Street where
- 1:39fear has a price tag. That corner trades
- 1:41a product called a credit default swap,
- 1:44which is insurance on a company's debt.
- 1:46And if that sounds familiar, it's
- 1:47because this is the instrument Michael
- 1:49Burry used in the big short to bet
- 1:51against the housing market. It works
- 1:53like any other insurance. You pay a
- 1:54premium, and if the company stops paying
- 1:56its lenders, the insurance pays you,
- 1:58meaning the more worried the gets, the
- 2:00more that premium cost. Most of the
- 2:02time, nobody talks about this market.
- 2:04It's a quiet corner of finance that only
- 2:06makes noise when something's about to
- 2:07break, which means when it does start
- 2:09making noise, it's worth paying
- 2:11attention to. And as you can see from
- 2:13this chart, the CDS market is waking up.
- 2:15This shows the average amount of default
- 2:17insurance on AI companies traded each
- 2:19month over the last five years. Since
- 2:22the summer, the people who lent the AI
- 2:23bubble its money have been buying
- 2:25insurance against it at a record pace.
- 2:27In July, about $2 billion a week of
- 2:30default insurance was trading on the
- 2:31seven big AI linked names. By September,
- 2:34it was over 9 billion a week, more than
- 2:37four times the volume in just two
- 2:38months, and the prices followed the
- 2:40volume. A few weeks ago, insuring Meta's
- 2:43debt hit a record of about 119 basis
- 2:46points. That means investors are paying
- 2:48more for protection against a potential
- 2:50debt default, and investors don't pay up
- 2:52for protection unless they're starting
- 2:53to get worried. Then Nvidia, Broadcom,
- 2:56Alphabet, Oracle, and others all hit
- 2:58record highs as well. And at the riskier
- 3:00end of AI, the numbers stop looking like
- 3:02insurance and start looking like a
- 3:04prediction. But before we get into that,
- 3:06let's take a quick pause to talk about
- 3:08something else. According to Microsoft,
- 3:10the average worker gets 117 emails a
- 3:12day. It starts before breakfast. Of the
- 3:14people online at 6:00 a.m., 40% are
- 3:17already in their inbox. Then the workday
- 3:19gets in the way and the replies pile up.
- 3:21By 10 p.m., 29% of active workers are
- 3:24back in their boxes. That's hours of
- 3:26your week disappearing into a keyboard,
- 3:28which brings us to the sponsor of
- 3:30today's video, Whisper Flow. Whisper
- 3:32Flow is a voice-to-text tool that
- 3:34actually works. Here's an email that
- 3:36would usually take me about two and a
- 3:37half minutes to type. With Whisper Flow,
- 3:40the same email took me 30 seconds and
- 3:42there's nothing to fix. Whisper Flow
- 3:44isn't just for email either. Here's the
- 3:45same text spoken into my phone twice.
- 3:48Whisper Flow on the left, built-in voice
- 3:50on the right. One of these I can send,
- 3:52the other one I have to fix first.
- 3:54Whisper Flow turns your speech into
- 3:55clear, polished writing across any app
- 3:58and any device. If you're interested,
- 4:00Whisper Flow is free to use and you can
- 4:02get one month of Whisperflow Pro with
- 4:04unlimited words for free using the link
- 4:06below. Thank you Whisperflow for
- 4:07sponsoring this video. Now back to what
- 4:09the market's showing us with the AI
- 4:11bubble. The five-year CDS spreads for
- 4:13CoreWeave climbed past 1,100 basis
- 4:16points, while the yield on the company's
- 4:172031 bonds rose to 16 %. At those
- 4:21levels, the market is implying about a
- 4:2364% risk of default for CoreWeave over
- 4:26the next five years. That's more than a
- 4:28coin flip's chance. The market is
- 4:29pricing in a serious probability that
- 4:31CoreWeave won't be able to pay back its
- 4:33debt. And this is a company the AI
- 4:35bubble treats as critical
- 4:37infrastructure. What's weird is that
- 4:38nothing big in the AI bubble has
- 4:40actually defaulted yet. There isn't a
- 4:42single massive bankruptcy to point to,
- 4:44but the price of insuring these
- 4:46companies against one is hitting new
- 4:48highs anyway. And every insurance
- 4:50contract has two sides, someone
- 4:52collecting the premium and someone
- 4:53paying it. So who's actually moving this
- 4:55market? Well, you can't buy a credit
- 4:57default swap through your brokerage and
- 4:59there's no buy button for it on Robin
- 5:01Hood. In the US, this market is
- 5:03effectively closed to everyday investors
- 5:05by law. To trade over-the-counter swaps,
- 5:07you need more than $10 million in
- 5:09investments just to qualify. So the
- 5:11market is almost entirely made up of
- 5:13banks, hedge funds, and institutional
- 5:15money managers trading hundreds of
- 5:17billions of dollars a day. Part of the
- 5:19reason why comes down to one simple
- 5:20rule. Banks cap how much they'll lend to
- 5:23any one industry. And last year, the
- 5:25Chicago Fed found that the average large
- 5:27American bank had promised loans to
- 5:29AI-adjacent industries worth about a
- 5:31quarter of their Tier 1 capital. That's
- 5:33the bank's core capital, which is
- 5:35supposed to act as a safety net. It's
- 5:37there to make sure a bank can absorb
- 5:39large losses without threatening the
- 5:40stability of the institution itself.
- 5:43Loan promises are now worth a quarter of
- 5:45that safety net, all tied to a single
- 5:47theme. And when these banks started
- 5:48pressing against their own limits, they
- 5:50started looking for somebody else to
- 5:52hold the risk. In May, JPMorgan Morgan
- 5:55Stanley and one of the largest banks in
- 5:57Japan started looking for ways to
- 5:59offload their rapidly growing pile of
- 6:01data center debt. That same month, a
- 6:03survey of credit insurers who absorb
- 6:05risk from banks found several nearing
- 6:07their sector limits, or quote, are
- 6:09already full on data center exposure. By
- 6:12September, HSBC was working on a deal to
- 6:15hedge the default risk on more than a
- 6:16billion dollars in data center loans by
- 6:19transferring it to outside investors.
- 6:21Then, last week, the largest AI chip
- 6:23financing deal on record was signed. 63
- 6:26billion dollars raised to put chips in
- 6:28front of Anthropic, the AI lab behind
- 6:30the chatbot Claude. And the structure
- 6:32behind that deal is the strangest part.
- 6:35The money doesn't even go to Anthropic.
- 6:37It goes to a shell company that buys the
- 6:38chips and Anthropic leases them, which
- 6:40is a key detail we'll come back to. This
- 6:43was also a deal that almost didn't even
- 6:45close. Some of the banks underwriting it
- 6:47felt they weren't being paid enough for
- 6:48the risk and considered walking away.
- 6:51Broadcom's CEO spent the weekend before
- 6:53the announcement personally talking them
- 6:55off the ledge, but only in exchange for
- 6:57higher interest rates and more
- 6:58protection around Broadcom's guarantees.
- 7:00And two days after that deal closed,
- 7:02Bloomberg reported that Barclays is
- 7:04looking for default protection on more
- 7:06than 10 billion dollars of AI
- 7:08infrastructure debt. The stated reason
- 7:10is to create capacity for more lending.
- 7:12So the banks aren't pulling back from
- 7:14AI. They're still writing the loans and
- 7:16closing the deals. They just don't want
- 7:17to be the ones holding all the risk when
- 7:19they do. And we've seen a version of
- 7:21this from Wall Street before. It's the
- 7:22same playbook from the last great
- 7:24bubble, and we all know how that one
- 7:26ended. In 2005, Michael Burry started
- 7:29buying credit default swaps on mortgage
- 7:31bonds. He later said he was astounded at
- 7:33how willing the big investment houses
- 7:35were to sell him swaps on highly risky
- 7:37packages of mortgages. But a year and a
- 7:39half later, Wall Street caught on and
- 7:41started to agree with him. And we know
- 7:43exactly what that looked like from the
- 7:44inside because a Senate investigation
- 7:47later reconstructed everything using the
- 7:49bank's own emails. It started on
- 7:51December 14th, 2006. Goldman Sachs's
- 7:54mortgage desk had just lost money 10
- 7:56days in a row. That was enough for the
- 7:58chief financial officer, David Viniar,
- 8:00to pull the mortgage team into his
- 8:01conference room. His instruction was
- 8:03simple. Get closer to home. By closer to
- 8:06home, Viniar meant for the mortgage
- 8:08department to assume a more neutral risk
- 8:10position. The next day, he summed up the
- 8:12meeting in an email to Tom Montag, one
- 8:15of the firm's most senior executives. On
- 8:17Goldman's big mortgage bet, he wrote,
- 8:19the position is reasonably sensible, but
- 8:21is just too big. And then he followed up
- 8:23with, let's be aggressive distributing
- 8:25things because there will be very good
- 8:27opportunities as the markets go into
- 8:29what is likely to be even greater
- 8:30distress. The translation for this, sell
- 8:33it now even at a loss because the real
- 8:36money will be made when everyone else is
- 8:37forced to sell. And whatever they
- 8:39couldn't sell, they insured. But Goldman
- 8:41didn't stop at just insuring its
- 8:43holdings. By the end of the first
- 8:45quarter of 2007, Goldman's mortgage
- 8:47department had swung from a $6 billion
- 8:49net long position in December 2006 to a
- 8:52$10 billion net short position in late
- 8:55February 2007, a $16 billion reversal. A
- 8:59senior Goldman executive later described
- 9:01a $3 billion net short position in
- 9:03subprime mortgage-backed securities as
- 9:06huge and outsized. Goldman's net short
- 9:08position in February 2007 was $10
- 9:11billion,
- 9:12more than triple that size. That summer,
- 9:15as the market started collapsing, Viniar
- 9:17got an update on the mortgage desk
- 9:18numbers. His response tells you what
- 9:20might be happening to people who don't
- 9:22have the big short. By the end of 2007,
- 9:25those bets against the mortgage market
- 9:26had brought in about $3.7 billion.
- 9:29The positions were so large and risky
- 9:31that the mortgage department repeatedly
- 9:33breached its risk limits, and Goldman's
- 9:35senior management responded by
- 9:36repeatedly giving the department new,
- 9:38higher temporary risk limits to
- 9:40accommodate its trading. But here's the
- 9:42part that matters for where we are today
- 9:44with AI. While Goldman was betting
- 9:46against mortgages, they were still
- 9:48selling them. In 2006 and 7 alone,
- 9:51Goldman built and sold about $100
- 9:53billion of mortgage deals. One of them
- 9:56was a $1 billion deal called Timberwolf,
- 9:58a hybrid cash and synthetic CDO squared.
- 10:01It closed in March 2007, around the same
- 10:03time Goldman was flipping its bet to the
- 10:05downside, and within 5 months, the deal
- 10:08had lost 80% of its value. and Tom
- 10:10Montag, the same executive who received
- 10:13Viniar's email in December, later gave
- 10:15his opinion on the deal in an email that
- 10:17was read aloud at a Senate hearing. Boy,
- 10:19that timberwolf was one shitty deal. So
- 10:22that's what the inside looked like last
- 10:24time around. The firms closest to the
- 10:25risk saw it first. They didn't announce
- 10:27it. They didn't stop. They sold what
- 10:29they could, insured what they couldn't,
- 10:31and kept the deals moving. In July 2007,
- 10:34Citigroup's CEO, Chuck Prince, described
- 10:37this all perfectly when he was asked why
- 10:39his bank was still making risky loans.
- 10:41His answer became one of the most famous
- 10:43lines in finance. When the music stops
- 10:46in terms of liquidity, things will be
- 10:47complicated, but as long as the music is
- 10:49playing, you've got to get up and dance.
- 10:51We're still dancing. That music he's
- 10:53talking about is liquidity, and 19 years
- 10:55later, the same music is back. Only this
- 10:58time, it didn't come from the head of a
- 10:59bank. It came from the richest man on
- 11:01Earth. A few weeks ago, Elon Musk posted
- 11:04on X and said, Debt markets have almost
- 11:06entirely shut off access to money for
- 11:08any company that doesn't have real
- 11:09revenue outside of AI data centers. Then
- 11:12he explained why his own company was
- 11:13fine. He said, SpaceX just closed $40
- 11:16billion of debt for GPUs, but we're
- 11:18investment grade and have Starlink and
- 11:20launch revenue. Then came the
- 11:22prediction. The more than 200 neoclauds,
- 11:25which are cloud companies that rent out
- 11:26AI computing power, Musk says they're
- 11:28headed for, quote, guppy level death and
- 11:31consolidation. And anyone who's ever
- 11:33owned a fish tank knows how guppy deaths
- 11:35go. There's no funeral and no
- 11:36announcement. One morning, the tank just
- 11:38has fewer fish in it, and the big ones
- 11:40look well fed. Musk never used Chuck
- 11:42Prince's words specifically, but it's
- 11:44exactly what he's describing. A debt
- 11:46market slowing its lending on the AI
- 11:48story and starting to lend on everything
- 11:50else a borrower has. We're three years
- 11:53into the most hyped technology bubble in
- 11:54history, and the survivors are being
- 11:56selected by their non-AI income. That's
- 11:58what it looks like when market liquidity
- 12:00starts to go because market liquidity
- 12:02isn't a number on anyone's balance
- 12:04sheet. It's an assumption. The
- 12:06assumption that when you need to borrow
- 12:08or you need to sell, somebody will be
- 12:10standing on the other side at close to
- 12:11yesterday's price. And market liquidity
- 12:14is there right up until everybody checks
- 12:16for it at once, which leaves one
- 12:18question. If the debt market really has
- 12:20shut its doors on companies that live on
- 12:21AI alone, how did the largest AI chip
- 12:24financing deal in history close last
- 12:26week? Anthropic is one of the most
- 12:28important AI companies on the planet,
- 12:30but from a lender's perspective, they're
- 12:32still exactly the kind of borrower Musk
- 12:34is talking about. They're only five
- 12:36years old. Nearly every dollar they earn
- 12:38comes from AI, the one thing lenders
- 12:40have gotten nervous about, and they have
- 12:41committed to more than $500 billion
- 12:44in future infrastructure spending, which
- 12:46means a company like Anthropic walking
- 12:48into the debt markets alone for $63
- 12:50billion should get an ugly rate attached
- 12:53to that debt, if they even get an answer
- 12:55from lenders at all. Except Anthropic
- 12:57didn't walk into the debt markets alone.
- 12:59They tagged in Broadcom with them. It's
- 13:01the same reason a landlord will rent a
- 13:0322-year-old their first apartment if a
- 13:05parent co-signs the lease. The landlord
- 13:08isn't approving the kid. He's approving
- 13:10the parent's credit score. Broadcom made
- 13:12the lenders a promise. They've agreed to
- 13:14provide residual value support equal to
- 13:16the first lean portion. Basically, if
- 13:18Anthropic can't pay whatever the senior
- 13:20lenders are owed, then Broadcom will
- 13:22cover their obligations up to about $43
- 13:25billion. And Broadcom made a nearly
- 13:27identical promise on another Anthropic
- 13:29deal back in June for up to $29 billion.
- 13:33When you add it all up, Broadcom is on
- 13:35the hook for as much as $72 billion of
- 13:37other people's debt, and they're also
- 13:39reportedly weighing more than $50
- 13:41billion of financing to help open AI by
- 13:43chips. Nvidia has already done the same
- 13:46thing for Open AI. They agreed to
- 13:48provide up to $105 billion in financing
- 13:51for a new artificial intelligence data
- 13:53center for Open AI in Ohio. And Google
- 13:55does the same thing, one level down.
- 13:57They guarantee the rent owed to the data
- 13:59center companies building the computers
- 14:01that Anthropic runs on. Alphabet's
- 14:03latest filing shows $48 billion of these
- 14:06backstops. At the end of last year, it
- 14:08was about 17. These three companies
- 14:10together are standing behind more than
- 14:12$200 billion of other companies'
- 14:14obligations, which raises the obvious
- 14:16question, why would anyone volunteer for
- 14:18that? It's because every borrower on
- 14:21that list is also a customer. Broadcom's
- 14:23backstop keeps Anthropic buying Broadcom
- 14:25chips. Nvidia's keeps OpenAI filling an
- 14:28Nvidia only campus, and the companies
- 14:30doing the backstopping even collect a
- 14:32fee for the signature. The parents
- 14:34aren't backstopping their debt out of
- 14:35love. They're financing their own sales,
- 14:38which brings us all the way back to that
- 14:40study from the Bank of International
- 14:41Settlements at the start of this video.
- 14:43They asked where the money that's
- 14:45fueling the AI bubble is actually coming
- 14:47from and found that more than half of it
- 14:49comes from inside the industry. This is
- 14:51how the suppliers keep guaranteeing the
- 14:53money so their customers can keep
- 14:54buying. And when you start to see
- 14:56default insurance rising to record highs
- 14:58across the industry, that's the market
- 15:00getting nervous because a parent can
- 15:02only co-sign so many leases before their
- 15:04own credit starts looking stretched.
- 15:06It's not just three companies holding up
- 15:08Anthropic and OpenAI anymore. They're
- 15:10holding up something much bigger, the
- 15:12American economy. Last year, the
- 15:14research firm Renaissance Macro found
- 15:16that for the first time ever, AI
- 15:18investment added more to US GDP growth
- 15:21than all of American consumer spending
- 15:23combined. And JPMorgan found that since
- 15:25ChatGPT launched in 2022, just 42
- 15:29AI-linked companies produced about 65 to
- 15:3175% of the market's returns, profits,
- 15:34and capital spending. Their conclusion
- 15:36was the generative AI investment theme
- 15:39has smothered the rest of the US equity
- 15:41market. If you take these 42 AI-linked
- 15:44companies out, the American stock market
- 15:46would have underperformed Europe, Japan,
- 15:48and China. So the music hasn't stopped
- 15:50yet. None of the big AI borrowers has
- 15:52defaulted. The deals keep closing, and
- 15:54by every official measure, the dancing
- 15:56continues. But you don't want to be the
- 15:58one still on the floor when it ends. 4
- 16:00months after Chuck Prince said Citigroup
- 16:02was still dancing in 2007, he was fired.
- 16:04His bank went on to take some of the
- 16:06biggest mortgage losses on Wall Street,
- 16:08and they needed one of the largest
- 16:09bailouts in American history to survive,
- 16:1145 billion dollars from taxpayers plus a
- 16:14government guarantee on more than 300
- 16:16billion dollars of their assets. So the
- 16:19institutional money isn't calling the
- 16:20top. It's doing something quieter. The
- 16:22banks are still writing the loans. The
- 16:24chip makers are still shipping the
- 16:26chips. The deals keep getting bigger,
- 16:27just like we saw in 2007, but the risk
- 16:30is being passed around from the banks to
- 16:32insurers and credit default swap sellers
- 16:35and from the borrowers back onto the
- 16:37chip makers and Big Tech who are now
- 16:39guaranteeing other companies' bills just
- 16:41to keep the deals closing. We even have
- 16:43the central bank for central banks
- 16:45saying these circular arrangements tend
- 16:47to be opaque, making monitoring and
- 16:49supervision difficult. More than half
- 16:51the money flowing into AI companies
- 16:53comes from other AI companies. America's
- 16:55economic growth has been leaning on that
- 16:57money and the people financing it are
- 16:59now quietly buying insurance against it.
- 17:02>> [music]
About this transcript
This page contains the full transcript of The Smart People Have Started Looking for the Exit by Casual Finance, generated from the public captions YouTube serves with the video. The transcript has 3,273 words across 495 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.