What if the AI Bubble Bursts (Day by Day) Transcript, AI Summary & Key Points

00:00 NVIDIA might be the most dangerous  “safe” stock in the market right now. You’ve seen the headlines. You’ve  watched the stock climb. It’s a no brainer, smart money invested in a safe,  cash-rich fortress at the perfect time. At least that’s what you thought. Right now, the graphics card titan has somewhere  between $27.8 and $33 billion sitting in a column called “Accounts Receivable.” That means  the chips are built, packaged, and shipped.

00:25 They just haven’t been paid for. And some of NVIDIA’s biggest customers are using those same chips as collateral  for loans they may never be able to repay. Now, one company’s next payment -  or missed payment - could decide whether NVIDIA keeps climbing... or crash  your retirement, and cost you your job. This is what happens if the AI bubble bursts.

00:45 T-12 Hours It’s Sunday, 9:47 p.m. in Livingston, New Jersey. The lights are still on  in the CoreWeave operations center. Which is strange. Sunday nights in the data infrastructure business tend to be quiet. When you are a company  that has made it by every public measure, there are automated systems and skeleton  crews that keep your machinery humming.

01:04 CoreWeave went public earlier that year. The IPO  was, depending on who you asked, a triumph or a warning sign. Investors smiled and clapped as the  stock opened on Wall Street. The company was now officially a publicly traded AI infrastructure  provider valued in the tens of billions. But the people still at their desks  on Sunday night aren’t celebrating.

01:23 They know that in 12 hours, the markets in  Tokyo, London, then New York will open.. And buried inside CoreWeave’s financial  statements is a number that should terrify every NVIDIA shareholder. It’s public.  It’s easy to find. And for months, the people who understand what it means have  been watching it like a ticking time bomb. CoreWeave’s debt-to-equity ratio sits  at 5.27 for the last quarter on record.

01:46 What that means, in plain English, is that for every single dollar  of actual value the company owns, it owes over $5 in debt. A debt-to-equity ratio  above 2 is considered high for many industries. And the AI sector is not known for  its conservative borrowing habits. You'd probably assume CoreWeave was founded by  veteran cloud engineers. Maybe former Amazon executives.

02:06 Maybe a group of MIT graduates  with a master plan to reinvent data centers. You'd be wrong. It was founded in 2017 by three commodities traders. Hedge fund  people. The kind of people who bought a single GPU in 2016 to experiment with Ethereum  mining out of a garage in New Jersey. Their original company was called Atlantic  Crypto, and they were good at what they did.

02:28 They kept buying GPUs, and the chips soon paid  for themselves at the height of the crypto boom. Then in 2019, Ethereum shifted away  from the mining model that had made them all handsomely rich. Almost overnight,  CoreWeave was stuck with warehouses full of graphics cards that were becoming worthless  for the one thing they'd bought them to do. So they pivoted.

02:47 They looked at what they had. And what they  had in abundance were physical machines, power infrastructure, cooling systems,  and technical expertise in running dense GPU clusters. They decided to  rent it to AI researchers instead. It was the right call. The timing was almost impossibly good.  Within 4 years, they went from a crypto mining operation nobody had heard of to the  backbone of the generative AI revolution.

03:05 Soon, they’d become the largest US tech IPO since 2021. The way CoreWeave built itself was not  unusual in the AI infrastructure space. That’s part of what makes it so alarming. The  model you're familiar with today is simple: borrow billions to buy NVIDIA GPUs, rent those  GPUs to AI companies, then use the very same GPUs as collateral to borrow even more money.

03:29 The GPUs  are both the product and the security behind the debt. That means the entire structure depends  on those GPUs continuing to hold their value. And the companies renting them  will keep paying…indefinitely. NVIDIA’s H100 GPU, the chip  at the center of this story, was selling on the secondary market  for around $50,000 at peak scarcity in 2024.

03:51 That number was embedded into all  kinds of loan agreements, leverage ratios, and spreadsheets from every bank that  financed the infrastructure buildout. But the people in that CoreWeave  office that night know something else, though. NVIDIA’s revenue isn’t spread  across thousands of customers in a comfortable, diversified way. Two anonymous buyers  referred to in regulatory filings only as Customer A and Customer B account for  39% of NVIDIA’s total quarterly revenue.

04:19 What does that mean? One of the most valuable companies  in the history of the stock market has built its quarterly income around 2  people who would rather remain nameless. Financially, it's like running a global  supermarket chain and discovering that nearly half your revenue comes  from just two families. And now, both have suddenly started buying  their groceries on borrowed money.

04:38 Nobody outside NVIDIA knows who they  are. The company won’t say. An NVIDIA spokesperson has declined to comment  every time the question has been asked. The best guesses among analysts and supply chain  reporters point toward Taiwanese original design manufacturers, or ODMs. Companies like Quanta  Computer, which assembles finished server systems and sells them to data centers,  or large system integrators like Dell.

05:01 It’s important to note that these  2 mysterious clients are NOT cloud companies or AI startups. They’re the  middlemen, hardware assemblers who buy chips in enormous quantities and sell the racks  to the end users who run the AI workloads. That distinction matters, because it means the concentration  risk runs even deeper than it appears. It means that NVIDIA doesn’t always  know, or can only estimate, who the ultimate end-use buyer of its chips actually  is.

05:25 The filing says so explicitly. The chips move through a chain of hands before they arrive  at their final destination. A destination that, in many cases, is a data center  owned by a company like CoreWeave. CoreWeave’s entire business  model is built on NVIDIA chips. Everything, from its debt structure to its revenue  and existence as a public company hinges on them.

05:47 The team in the office isn’t panicking. Panic would require surprise. What they  have instead is arguably something worse. Clarity. CoreWeave’s employees have watched the AI revenue  problem emerge over the past year like a slow tide. Sequoia Capital published a piece calling it  the $600 billion question: What is the gap between what the AI industry is spending on infrastructure  and what it actually generates in revenue?

06:10 The math was not encouraging. The industry is projected to pour over  $700 billion in capital expenditure into AI in 2026 alone. All the giants  - Microsoft, Google, Meta, Amazon, Oracle - are building, buying, and  expanding. It’s enough spending in a single calendar year to build several hundred  Burj Khalifas, the world’s tallest building. But building the data centers  was only half the challenge.

06:34 The real question was whether  they could ever pay for them. Because at the time, almost none of these  AI companies were making enough money to justify the billions they were spending. Companies like CoreWeave are the central pillars of that ecosystem. And  somewhere in the next 12 hours, the markets are going to open. Every analyst on  every trading desk in every timezone is going to look at the same sets of numbers that the people  in that office have been staring at all weekend.

06:57 The difference is that the people in the  office already know what those numbers mean. The analysts have not. Day 1 The markets open on Monday at 9:30 a.m., Eastern  Time. For about 45 minutes, nothing happens. This is pretty normal. The hour or  so of any trading day is often a kind of financial small talk. Prices get  adjusted and algorithms recalibrate.

07:12 It takes some time for the humans  to catch up with the machines. NVIDIA opens slightly down. CoreWeave is  flat. The AI sector, broadly speaking, is doing what it has done for most of the  past 2 years… quietly inflating. Then, at 10:17 a.m. Eastern, a single line item fails  to appear in a single bank’s clearing system. CoreWeave misses a scheduled debt payment.

07:34 The missed payment doesn't trigger alarms. It’s nothing catastrophic. This isn't  a company-ending default. There’s no need to issue a press release or strong-arm  their CEO into issuing a public statement. A payment that was supposed  to clear overnight didn’t. Now the payment is sitting somewhere  in the interbank settlement system, delayed just long enough for automated risk  models to start paying attention.

07:53 Within minutes, warnings begin firing through a dozen  institutions with exposure to CoreWeave's debt. Those institutions have  every reason to be nervous. By the end of its third quarter  of operations as a public company, CoreWeave was paying $311 million in  interest. That figure had tripled in less than twelve months. The company’s IPO, which  it originally targeted to raise $2.7 billion, was slashed to $1.5 billion after investors  looked at the balance sheet and got cold feet.

08:25 And that hurt. By 10:30 a.m., trading desks know what’s  going on. By 11:00, the first analyst notes are circulating. They’re draped in  cautious language. Hedge-everything kind of prose. It’s the financial equivalent  of someone quietly skirting to the door at a party where something has… or  may well soon go very, very wrong. CoreWeave drops 11% before lunch.

08:44 NVIDIA, whose largest cloud customers are also  CoreWeave’s largest lenders, drops 4% by noon. 4% of NVIDIA’s $5 trillion market  capitalization isn’t a small number. Depending on the exact moment you measure it,  it could be between $204 and $216 billion. All that, gone by lunchtime on a Monday. At this point, the market is still in denial. The story goes something like this: CoreWeave  is just a speed bump.

09:09 An isolated issue in an AI boom that remains fundamentally intact. After all,  the AI buildout is too large, too strategic, and backed by too much capital for one infrastructure  company to threaten the entire ecosystem. This is a contained event. These things happen.  Someone will step in and oder will be restored. The problem with telling yourself that  story over and over again is that everyone in the market is simultaneously  trying to ignore one other thing.

09:37 The GPUs. CoreWeave borrowed the money to buy the  chips. The chips were the collateral. The loans were written when H100  GPUs were worth $50,000 each. But the market had changed. For months, prices in the secondary GPU market had been drifting lower as new chips arrived  and demand began to slow. Suddenly, lenders were asking if they had to seize and liquidate the  collateral, what were these GPUs actually worth?

10:01 The estimates being discussed  weren't anywhere near $50,000. They were closer to $5,000 to $10,000. That’s a 90% drop on loans written at peak prices. A Bitcoin miner who bought his rigs at the  top of the 2021 cycle and tried to sell them in 2022 would have also found  themselves in a similar situation. The point is that anything at NVIDIA’s scale  is no longer a niche collector’s market.

10:24 The GPUs are the physical backbone of the  artificial intelligence revolution. And now, they’re worth cents on the dollar  compared to the prices that justified the loans that built the machine’s  infrastructure. The lenders know this, and have for a while. The question nobody  wants to answer out loud is who moves first. At 2:14 p.m. Eastern, a private credit  firm that holds a significant portion of CoreWeave’s debt instructs its legal  team to begin margin call proceedings.

10:50 The clock, already running,  now begins to accelerate. Day 2 It’s now Tuesday, 6:00 a.m. at a data center in New Jersey when the phones begin  to ring. It’s not even sunrise yet. It’s a moment not unlike the beginning  of many other financial crises. They’re rarely as dramatic as you’d expect. Nobody  is shouting or hurling insults. There is, however, a man holding a clipboard  in a security uniform.

11:09 On it are a set of instructions prepared by a  law firm that doesn’t make mistakes. The instructions say that the company’s  collateral, specifically the racks of NVIDIA H100 GPUs which have been running  AI workloads for the past 18 months, are now the property of the lending  institution. They’ll need to be catalogued, tagged, disconnected, and  prepared for liquidation.

11:32 The people at the data center don’t fight this. They can’t. They’ve seen the paperwork and  know what’s coming. Since 4 a.m., engineers have been scrambling to move active  workloads off the affected machines before access disappears. One veteran engineer  later describes the experience like having someone repossess your house while you're  still trying to move the furniture out.

11:51 Most people have never stopped to think  about what this process might actually look like. A GPU server isn’t a laptop. Good  luck sticking it in a bag and walking home. An H100 Server Weighs 36 Pounds (16 Kg) Per  Rack Unit.Build out a fully-loaded NVIDIA H100 AI server and we’re talking 400 pounds (181 kg)  of static weight. These machines aren't designed to be moved.

12:12 They require industrial-scale  cooling just to stay operational. They're also connected directly into the facility's power  and networking systems. The cabling, conduit, and supporting infrastructure were installed  when the building was originally fitted out. To disconnect one, you’d have to power it off,  log and tag every component, disconnect fiber and copper in the right sequence.

12:30 Then you need  to use specialized lifting equipment to move it without destroying either the hardware.  They’re made of reinforced steel frames, and are taller, more rigid, and deeper than  your average server rack. Each individual rack unit can cost $30,000 with a fully integrated  multi-GPU set up costing up to $8.8 million. This particular New Jersey facility has around  4,000 servers or so.

12:50 Which means removing them becomes a multi-day operation that resembles  something akin to a military logistics exercise. The lawyers have arranged for a third-party  asset management firm to handle the physical side. They have done this before, in smaller  venues, during the crypto mining collapse of 2022. They have the equipment, the  manifests, the insurance riders.

13:09 They are efficient and completely indifferent to  what’s happening in the building around them. Engineers who built these systems are now watching  them get boxed up and prepared for auction. By 9 a.m., thousands of servers  are offline in New Jersey alone. The significance of this goes beyond the  servers themselves. Every AI company running workloads on those machines is now getting a  very unhappy error message.

13:32 The compute they paid for and accounted for in internal revenue  projections and product roadmaps has vanished. Now the companies who built their own products  on top of CoreWeave’s infrastructure have to do something that is structurally  very difficult in the AI industry. They’ve got to stop. Customers start to notice. The secondary effects only take hours to  become apparent.

13:54 A startup that’s been using CoreWeaves compute to power  a legal document review tool sends an email to its customers explaining the  unexpected outage. Another company whose AI-generated marketing platform runs on the  same platform goes dark for 6 hours. Again, it’s not world ending stuff for either  company. But the signals are alarming. At NVIDIA, there are other alarming signs.

14:17 They’ve got $27.8 to $33 billion sitting  in accounts receivable right now. It's all the money they’re owed. Revenue  that’s already been recognized. It’s passed through the income statement,  been celebrated on earnings calls, and used to justify valuations to  otherwise hallucinatory levels. But everyone who knows anything  knows that accounts receivable is not cash.

14:38 It’s more like a promise.  One made in large part by companies that are now finding it extremely  difficult to keep that promise. Two anonymous customers account for  39% of NVIDIA’s quarterly earnings. Customer A and Customer B. And one  of those just missed a payment. That afternoon, NVIDIA’s CFO speaks with  investors on a call. In careful language, they deliver a calm message.

15:01 The underlying message,  stripped of its hedging, is not. Day 4 You’re sitting in your kitchen on Wednesday  morning. It’s 7:45 a.m. Maybe you don’t work in finance. Maybe you work in something  tech-adjacent. Maybe you’re a teacher, or run a small business, or work for  a global healthcare conglomerate. But somewhere in the past few years,  probably encouraged by every podcast and YouTube channel you’ve ever encountered,  you’ve been putting money into an index fund.

15:26 One that tracks the S&P 500. You were told to diversify your holdings. Spread your money across companies. Never put  your eggs in one basket. If one fails, who cares? The market will hold true in the  end, which means you can sleep well at night. Few people today, including you,  actually realize that the top ten stocks in the S&P 500 now account  for 40% of the entire index’s value.

15:47 Why should you care? Well, in almost 60 years of market history,  the S&P 500 has never been this concentrated. The last time a technology mania pushed stock  prices to seemingly implausible levels was during the dot-com bubble. At its peak in 2000, the ten  largest stocks accounted for 26% of the index. Then reality caught up. What followed was a very unpleasant few years for investors.

16:08 We’re well past that now. Most of those top ten stocks are, in one way or  another, in the AI trade. You pull out your phone that morning and open your brokerage app. You  have been watching the news with detached concern. CoreWeave’s missed payment, and  NVIDIA being slightly down only matters to rich people at large  institutions. It doesn’t apply to you and your sibling who occasionally  talk about meme stocks over coffee.

16:33 But you notice that the S&P 500 is down 6.4% this  morning. You rack your brain. Then it hits you. The last time you saw a drop  like that was in March 2020, the early days of the global pandemic when  nobody knew how bad things were going to get. The time the world briefly stopped and  the market fell off a cliff in real time. You remember that because you lived  through it.

16:52 You saw how it affected neighbors and friends and parents. You remember  the conversation with your sibling where they begged you to hold and not sell, since the  market would come back, stronger than ever. And it did come back. That time. You take a second to do the math.  If you’re like most of your mildly affluent peers, you’ve spent a decade or  two building up to $150,000 in your 401k.

17:15 $9,600 has been erased. The market has been open for just 11 minutes. There’s a reason why it's happening so  fast. And it’s always the same. When 40% of the S&P 500 is riding on just ten stocks,  all depending on the same story being true, the market stops being diversified. It  starts living and dying by a single idea. The AI revolution. The belief, again, is that artificial  intelligence will soon generate enough revenue to justify the debt that  built it.

17:39 The CoreWeave default is the first real piece of evidence  that the timeline might be wrong. Not slightly wrong. Fundamentally wrong. By noon on Wednesday, the S&P 500 is down  8.1%. You call your financial advisor. The call goes straight to voicemail. Their advice,  when it comes, will be the same as it always is. Stay the course! Don’t panic-sell!

18:00 The markets always recover. It’s true, historically. But it’s way easier to give advice when you are not the  one watching the numbers plummet. Also, that advice rarely acknowledges the  specific structural problem underneath this particular drop. The fact that the  fund you were told to invest in was not actually diversified like you thought it was.

18:20 NVIDIA, Microsoft, Apple, Amazon, Alphabet,  Meta. These are the names bankrolling the AI boom. They make up a disproportionate  share of what you “own” in the market. And they are all exposed to the question  of whether the $700 billion being spent on AI infrastructure this year will  produce revenue that justifies the spend. Day 5 On Thursday morning, a mid-sized AI  software company at a nondescript tech campus in Austin calls an all hands meeting.

18:47 They’re no different than many companies  like them. It felt almost criminal how easy it had been to raise the $200 million  in venture capital 18 months ago. All it took was a compelling demo and a revenue  growth chart that pointed mostly upward. The meeting is virtual. The company  has 400 employees across 3 time zones. The CEO, who is 37 years-old and has  built and sold one company before this one.

19:07 He has the look of someone that hasn’t slept. He announces that the company is  immediately reducing headcount by 35%. Soon, 140 people will receive  calendar invites from HR. It’s all he can say right now. The legal  team has advised, and the meeting ends. In conference rooms and home offices and coffee  shops and parked cars, 140 people open their calendar invites like a death sentence.

19:28 Some  of them have been at the company for 3 years; others for 3 months, having moved cross-country  securing their dream AI tech job. Many of these are still paying off student loans they took out  to get degrees that got them hired. Degrees that, increasingly, are looking less  and less relevant or necessary. The AI industry’s version of “doing  fine” has always required a certain willingness to not look too  closely at the faces behind the revenue numbers.

19:52 To say nothing  of the revenue numbers themselves. The compute costs were high, but demand  felt insatiable. The burn rate was high, but there was always another 18  to 24 months of runway. Investors were patient and the models kept  improving. And with each breakthrough, the payoff felt closer. For the people who  got in early, it felt less like investing in a technology company and more like showing up  in California at the start of the gold rush.

20:17 For any setbacks that came,  the market would figure it out. And the market, it turns out, has figured it out. The Austin firings repeat  themselves all over the U.S. It's happening in San Francisco, where AI startups  that were expanding just months ago are quietly giving back office space and cutting costs. It's happening in Seattle, where hundreds of companies were pulled into orbit  by Microsoft's AI ambitions.

20:35 Now, they’re discovering that when the center  contracts, the shockwaves travel outward. And it's happening in New York, where AI  infrastructure investors built entire businesses on top of other companies' AI products. They’re  are learning how fragile that dependency can be. Each of these geographic centers are  running their own math this Thursday morning.

20:57 They’re not loving what they  find. The numbers are far from reassuring. In the first quarter of 2026 alone, before any  of this week’s events have been fully processed, somewhere north of 50,000 tech workers  in the AI industry had already lost their jobs. It was the highest Q1 job losses  since 2023. Halfway through the year, that number had already climbed past 120,000.

21:19 The numbers are startling and worrying. IBM has eliminated hundreds of HR  positions and replaced them with AI chatbots. Salesforce cut 4,000 people and said  AI had reduced the need for staff. Microsoft, which laid off 15,000 people in 2025, had its CEO  tell employees this was a “new era shaped by AI”. What all the job cut announcements had in common  was a particular kind of corporate doublespeak.

21:43 In the short term, the same technology  that was supposed to create abundance was being used to justify the elimination  of the people who had helped build it. Everyone in the tech industry knows it  has always been cyclic. There have been plenty of downturns before. 2001.  2008. 2022. People got laid off, found other jobs eventually. The industry  contracts and expands, eventually.

22:05 This time, the question is whether  eventually has a reliable timeline. Or whether this particular correction is different  in character from the ones that came before. That’s because this time, the correction  centers on the nature of the AI revolution’s physical infrastructure. Everything that  got built in the boom based on demand that had not yet materialized.

22:22 Physical things  like chips, data centers, power contracts and fiber optic cables are much harder to  unwind than overpriced SaaS subscriptions. You can’t just turn off a data center. You  can’t return 17,000 GPUs to the store. And you can’t immediately redirect the careers  of hundreds of thousands of people who spent the last few years specializing in skills that  were, until recently, in extremely high demand.

22:48 The 140 people in Austin are, right  now, beginning to find that out. Day 5 It’s 9:am on Friday, in a coffee shop on the  corner of Austin’s South Congress Avenue. It’s been open for 11 years, open before the  tech companies came en masse. The owner has watched the neighborhood change over that  time. On the one hand, she is grateful the tech workers came.

23:05 They spend money. Buy $12  lattes. They frequently have meetings and run tabs on their corporate cards, tipping  the help staff well when they expensed it. On the other hand, she’s uneasy. She’s noticed  that over the past 6 months, some of her regulars have stopped coming. Gradually, fewer faces  file in through the sun-drenched door each morning. She chalked it up to remote work or  new coffee shops in the area.

23:24 The tech crowd always felt transient and temporary, the way  college undergraduates turn over every 4 years. One thing she hadn’t thought to  connect their absence to was an obscure debt-to-equity ratio she’d never heard of. The AI campus across the street from her shop  which has been her single most important revenue source is now 35% smaller than it was the  day before.

23:44 The Friday coffee rush she’s always relied on, planned for, staffed  for, is about half of what she budgeted. She doesn’t know it yet, either, but  it will be this way next Friday too. And the Friday after that. And the next. The corporate catering contract she signed in January is about to become non-existent. Not  because of anything she did. But because the company on the other side of the agreement  relied on AI infrastructure that is now being repossessed in New Jersey.

24:11 She’s thinking about next Saturday, and whether she should call her  part-time staff or adjust her milk order. In San Francisco, Peet’s Coffee, a Bay  Area institution that’s been serving the city since 1966, long before the  city’s eventual tech dominance, announced the closure of roughly 30  Bay Area locations. They cited what a spokesperson called “long-term growth  priorities and current market conditions.” Starbucks was in a similar boat  in San Francisco.

24:34 They, too, closed multiple locations simultaneously.  It included stores in corridors that served the daily ritual of thousands of office  workers whose offices were now half-empty. The foot traffic analytics firm Placer.ai found  that San Francisco had the lowest office visit rate of any major American city. It was more than  50% from pre-pandemic levels and still falling.

24:58 The coffee shops didn’t  cause that; they absorbed it. Economists call this a “secondary shock.”  It’s the moment when a financial crisis moves from the institutions of origin  to the people and businesses who had no direct exposure to the bad bets. People  connected by proximity, dependency, or the sad fact that the money has to flow  somewhere or it stops flowing everywhere.

25:18 In previous tech contractions, the secondary  shock was real but bounded. The dot-com crash cost about $5 trillion in market value, but the  internet itself survived and the companies that rebuilt on its foundation eventually  generated far more than what was lost. Will the same thing be true when the collateral  isn’t internet domain registrations and vaporware business plans?

25:37 Will it be true about the  physical computing hardware that is not in the earliest stages of being repossessed by teams with  clipboards dispatched to data centers before dawn? Day 7-14 By the following Monday, the S&P  500 has given back a number that will be cited in business school  case studies for a generation. Retirement accounts that took decades to  build are down 15-22% depending on their concentration in tech.

26:00 Workers in their 50s,  now 5, maybe 10 years from the finish line, are doing the grim arithmetic. They’re  coming to terms with the reality that they’ll have to work longer, perhaps  indefinitely, to make up the difference. These are the people who can least afford  to absorb the loss. These are the ones who, true to most financial crises, are going  to be forced to absorb the brunt of it.

26:21 NVIDIA’s stock is down 28% from its peak.  But it’s still a wildly profitable company, make no mistake. It bears reemphasizing  that NVIDIA, for all its foibles, continues to make real chips and GPUs that do  real things that real industries depend on. Their fate isn’t inextricably linked to the  dot-com crash. NVIDIA will survive this. Still, nobody at the company knows this early  whether the $33 billion they’re owed in accounts receivable will ever get collected.

26:45 Whether  Customer A and B and the rest of the highly leveraged cloud infrastructure companies can  actually pay the bills they’ve already racked up. This is a question that will take months or  years to fully answer. And in the meantime, the stock price reflects the uncertainty  of the moment, not the long-term thesis. The banks that funded the GPU buildout  are running their own stress tests.

27:06 Private credit firms who issued debt to AI  infrastructure now look at their portfolios, once fat with perceived  profits, much differently now. The venture funds who bankrolled all this are  now facing calls with institutional investors and endowments and pension funds who want to know  what their portfolio is really worth. Some of the pension funds are public employee retirement  systems.

27:24 People like teachers, firefighters, and sanitation workers who have nothing to do  with the AI revolution. They’ve never heard of CoreWeave or made any investment decisions of  consequence except the decision to invest at all. They’re coming to terms still with the fact that  their future security is, quite literally, tied to whether a New Jersey data center can be liquidated  efficiently enough to cover a margin call.

27:48 The California Public Employee’s Retirement  System, or CalPERS, is the largest public pension fund in the United States. It manages  almost half a trillion dollars in assets on behalf of roughly 2 million teachers, police  officers, and state workers. Its equity portfolio, like most institutional portfolios at  this scale, is awfully index heavy. Which means that whether anyone wanted it  this way or not, it is severely exposed.

28:11 When these stocks drop 20%, CalPERS loses  roughly the GDP of a mid-sized American state. The pension board will have to issue  a statement to the effect that the “long-term investment horizon” has suddenly  shifted. It will be accurate and irrelevant to the 63-year-old retired schoolteacher in  Fresno who was planning to stop working this year. They’re now being told by a benefits  administrator that the picture has changed.

28:35 Modern financial contagions don’t travel in  straight lines. This is the architecture. It travels and echoes through every system that  is connected to every other financial system just as the AI bubble of the early 2020s is now  connected to almost everything we see and do. And the human cost is real. How to Save the Future The most disorienting thing  about a situation like this, is that there is nobody really to be angry at.

28:58 Financial crises are supposed to have a  face. There was a CEO, a trader, a bank, somebody asleep at the wheel. You could feel  confident attaching their name to the damage. 2008 had its villains, real and constructed.  The mortgage brokers. The loan officers. The ratings agencies in bed with the banks. Whoever.  The knowledge that someone, somewhere had looked at the machine and decided to run with it  anyway is a form of comfort.

29:18 Even if it’s just a single executive that gets prosecuted  and the public gets to pay for the damages. That’s exactly what happened in 2008. But an AI crash like this is different. There's nobody obvious to blame. No villain. No  fraud. No single bad decision. What happened was the combined result of thousands of choices that,  taken individually, were completely rational: The pension fund manager overweighted  tech because that's what the benchmark did.

29:46 The benchmark overweighted tech because  that's where the market's money was flowing. The money flowed there because analysts  believed the growth justified it. The analysts believed it because the revenue was real. The problem wasn't that the revenue was fake. It's that the debt was built on the assumption  that the revenue would grow even faster. The debt looked safe because the collateral supported it.

30:04 The collateral looked solid because the chips were scarce, valuable, and in extraordinary demand.  Every step made sense. Every link in the chain was responding rationally to the one before it. That’s what made it dangerous. This is what systematically unstable structures  look like from the inside. Tightly coupled, in the same direction, with no slack or wiggle room.

30:23 Eventually, the assets will change hands and the market will find a bottom. Experts will publish  postmortems. Commentators will outline the reforms that should have happened. They won’t happen. Because the people with the power to implement those changes are often the same people who  did very well under the old system. And by the time everyone agrees on what went wrong,  they're already busy building the next one.

30:45 Silicon Valley was not stupid about any of  this. It was, in a specific and important sense, the right way to approach the AI  revolution. The technology works. The compute is real. What the models can  do are extraordinary human achievements. We’re not hallucinating any of this; it  is the future. And it may truly be that the rationalization that AI will generate the  revenue it promised eventually is actually not crazy.

31:08 It might turn out that the timing  was the only thing wrong, not the thesis. But that lands very differently depending  on whether you’re a venture capitalist with a 15-year fund life, or a 51-year-old  electrician 9 years from retirement. The eventual AI crash, if that’s what we’re  calling it, is not the end of artificial intelligence any more than 2001 was the end of  the internet.

31:26 What comes after will probably be better, built on cleaner and clearer foundations  with more honest math. The people who build the next version won't be the same people who  built this one. And the people who suffer the consequences usually won't be the people who  made the original decisions. That's one of the oldest stories in economic history. The gains are concentrated.

31:48 The losses are distributed. But what if the bubble has already burst and we just haven’t realized it yet? Watch “$115  Billion Burn Rate: The AI Bubble Just Popped” to see how the entire industry may be balanced  on a knife edge. Or watch this next video.