What a Burst AI Bubble Looks Like Day by Day: CoreWeave, Nvidia and the $700 Billion Question

From the outside Nvidia looks like the market's safest harbor; its coffers are full, demand has exploded, and the shares have climbed for years. But the video opens with a sharp claim: somewhere between $27.8 and $33 billion of receivables sit on the balance sheet waiting to be collected. The chips have been built, packed and shipped, yet the money has not arrived. More critically, some of the large buyers placed the very same hardware up as loan collateral. In this scenario, the fate of a single payment sets the direction of the whole story.

The story starts on a Sunday night at 9:47 p.m. in Livingston, New Jersey, with the lights still on at the CoreWeave operations center. The company went public that year as an AI infrastructure player worth tens of billions. Yet one ratio in the financial statements ruins the night shift's sleep: debt to equity at 5.27. That means more than $5 of debt against every $1 of real value. Levels above 2 already count as elevated in many industries, while AI infrastructure uses leverage far more aggressively.

CoreWeave's origins break the stereotype; the founders are not cloud engineers but three commodities traders who joined forces in 2017. The team that tried Ethereum mining with a single GPU in a garage in 2016 grew through the crypto boom under the name Atlantic Crypto. When Ethereum moved away from mining in 2019, they were left with warehouses of cards whose mining value was melting. They chose to rent out what they still had: machines, power infrastructure, cooling, and know-how in running dense GPU clusters. The timing proved near-perfect; within four years an unknown mining shop had become backbone material for the generative AI wave, followed by one of the largest US tech listings since 2021.

According to the video this model is the industry norm rather than the exception: borrow billions, buy Nvidia GPUs, rent the cards to AI firms, then pledge the same cards against fresh loans. When product and collateral are the same hardware, the structure rests on two bets: the cards will hold their value and tenants will keep paying. The H100 is described as changing hands near $50,000 at the 2024 scarcity peak, and that figure is what banks wrote into leverage tables and loan contracts. Once the valuation assumption cracks, the collateral chain cracks with it.

Nvidia's revenue is not spread across thousands of buyers; two anonymous purchasers logged only as Customer A and Customer B make up 39% of quarterly sales. CNBC, citing the SEC filing, breaks it down as 23% for A and 16% for B. Leaning on analyst and supply-chain estimates, the video suggests the pair are not cloud giants but Taiwanese design-manufacturers in the Quanta mold or system integrators of the Dell type. In other words, middlemen that buy chips and sell finished racks. The distinction deepens the risk, because Nvidia cannot always know who the final user is; the filing says so openly. The silicon changes hands several times before reaching a data center run by someone like CoreWeave.

The backbone of the video is Sequoia Capital's arithmetic: the gap between money buried in infrastructure and revenue produced by AI. An analysis that began as a $200 billion question in 2023 had grown into a $600 billion question in the 2024 update. Expected AI capital spending for 2026 alone is cited above $700 billion, with Microsoft, Google, Meta, Amazon and Oracle all building at once. The video reaches for a bold image: a single year's spend could fund hundreds of Burj Khalifas. The question is simple and brutal: will these data centers ever pay for themselves? At that point, few AI firms earned enough to justify what they were spending.

Monday opens at 9:30 a.m.; 45 minutes pass in ordinary chatter, Nvidia slightly red, CoreWeave flat. At 10:17 a.m. a single line never lands in a bank's settlement system: CoreWeave misses a scheduled debt payment. Not a bankruptcy on its own, just a delay, yet risk models start firing at once. The account given is that quarterly interest reached $311 million in the third quarter, tripling in under a year. Planned IPO proceeds of $2.7 billion are said to have shrunk to $1.5 billion after investors studied the balance sheet. By midday CoreWeave is down 11% and Nvidia 4%; on a $5 trillion valuation, 4% means $204-216 billion gone before lunch. In the first hours the market files the event as an isolated glitch and chooses denial.

What the denial papers over is the current value of the cards. The loans were signed when the H100 stood near $50,000; as fresh chips arrived and demand cooled, second-hand prices slid for months. The seizure values discussed in the scenario sit in the $5,000-10,000 band, a 90% fall from the peak. The parallel offered is a Bitcoin miner who bought rigs at the 2021 top and tried to sell in 2022. The difference is scale: this is not a niche hobby market but the physical backbone of the AI revolution. At 2:14 p.m. a private credit firm holding a large share of CoreWeave's debt instructs its lawyers to start margin-call proceedings; the countdown accelerates.

At 6:00 a.m. on Tuesday the phones ring at the New Jersey data center; a uniformed guard arrives with instructions from a law firm that makes no mistakes, ready to log and remove H100 racks that ran AI workloads for 18 months. The numbers describe the physical job: 16 kilos per rack unit, 181 kilos for a fully loaded H100 server; industrial cooling, dedicated power and networking, fiber and copper to be pulled in the right order, steel frames, lifting gear. Values run to $30,000 per rack and up to $8.8 million for a full multi-GPU build, with around 4,000 servers in the facility. Removal turns into a multi-day military-style operation, handled by an asset manager that did similar work after the 2022 crypto collapse. Before 9:00 a.m. thousands of servers in New Jersey are dark. AI firms that planned against that compute see error screens; products running on CoreWeave have to stop. A legal-document startup emails customers about the outage, an AI marketing platform goes dark for 6 hours. At Nvidia the $27.8-33 billion receivables question grows, with the 39% share of Customers A and B named in the same breath as the missed payment. In the afternoon the CFO holds a calm-toned investor call; the video suggests the subtext is anything but calm.

Wednesday at 7:45 a.m. brings the ordinary index investor into frame: a teacher, a small-business owner, a health worker who fed an S&P 500 fund for years. Promised diversification brought sound sleep, yet the top 10 holdings now carry 40% of the index. In nearly 60 years of index history there is no denser lineup; even at the dot-com peak the top 10 stopped at 26%. A 6.4% morning drop recalls March 2020. The math is plain: $9,600 erased in 11 minutes from a $150,000 retirement pot. By noon the loss stretches to 8.1%. The video's thesis is blunt: when nearly all of the top 10 ride the AI story in one form or another, the fund is not diversified but indexed to a single idea. Nvidia, Microsoft, Apple, Amazon, Alphabet, Meta; above all of them hangs whether 2026's $700 billion build will convert into revenue.

Thursday morning a mid-sized AI software firm in Austin summons 400 staff to a virtual meeting; the team that raised $200 million 18 months earlier on a compelling demo and a rising chart cuts 35%, with 140 people receiving calendar invites the same day. The scene repeats across the country: expanding startups in San Francisco hand back offices and tighten belts, hundreds of firms in Microsoft's AI orbit around Seattle wobble as the center contracts, funds in New York that built on top of other companies' AI products learn the price of dependence. More than 50,000 AI-tech jobs were already gone in the first quarter of 2026, past 120,000 by mid-year; IBM handed hundreds of HR roles to chatbots, Salesforce cut 4,000, and Microsoft, after shedding 15,000 in 2025, framed the new era through AI. The video recalls the 2001, 2008 and 2022 cycles and asks what is different this time: unlike a software subscription, this clear-out cannot be cancelled. Chips, data centers, power contracts and fiber cannot be sent back. Seventeen thousand GPUs cannot be returned to the store, and hundreds of thousands of careers built on these skills cannot pivot overnight.

Friday at 9:00 a.m. the owner of an 11-year-old cafe on Austin's South Congress sees a routine rush running at half strength; the campus across the street is 35% smaller and a January catering contract is about to vanish. In San Francisco, Peet's, serving since 1966, has closed around 30 spots, with Starbucks shuttering nearby corridors. Office attendance in the city ranks lowest among large US metros in Placer.ai data, under half the pre-pandemic level and still falling. This is what economists call the second shock: people who never placed the bad bet pay through proximity and dependence. The dot-com crash erased $5 trillion yet the internet survived and something larger grew from the ashes. The video asks whether the same sentence can be spoken when the collateral is not domain registrations and vaporware plans but physical hardware collected at dawn with a clipboard.

The following Monday the reckoning lands: decades of savings down 15-22% depending on tech weight, workers in their fifties recalculating how many more years they must work. Nvidia sits 28% below its peak, but the video adds a caveat: the company still makes real chips and mints real cash, a different fate from dot-com wreckage. The true unknown is how much of up to $33 billion in receivables will ever be collected, an answer that takes months or years. Banks that funded the GPU build run stress tests, private credit shops reprice their books, venture funds field questions from endowments and retirement money. The example given is CalPERS, America's largest public pension fund, stewarding nearly half a trillion dollars for close to 2 million teachers, officers and state workers, dragged into the fall by index weight whether it chose the trade or not. A 20% slide in the shares writes off the output of a mid-sized American state. For the 63-year-old retired teacher in Fresno who planned to stop work this year, what remains is a benefits administrator reciting the long-horizon line. Modern contagion does not travel in straight lines; it echoes through every connected system.

The most unsettling passage is saved for last: there is nobody to blame. The year 2008 had its brokers, loan officers and rating agencies; this scenario has no fraud and no single bad call. It totals thousands of separately rational choices: the fund manager tracked the benchmark, the benchmark followed the money, the money believed the growth, the growth rested on real revenue. Debt was stacked on the assumption that revenue would grow faster still; collateral looked solid because the chips seemed scarce and precious. Every link adapted to the one before it, leaving no slack anywhere. The video expects assets to change hands, a bottom to form, postmortems to be published and reforms to be discussed and shelved. It concedes the technology works, the compute is real, and the models are extraordinary; perhaps only the schedule was wrong, not the thesis. But a wrong schedule looks very different to a founder with a 15-year fund than to a 51-year-old electrician nine years from retirement. Just as 2001 was not the end of the internet, this would not be the end of AI; what follows would likely be built on cleaner, more honest math. It closes on one of the oldest lines in economic history: gains concentrate, losses distribute.