Tom BilyeuDid the AI Bubble Just Pop?! Anthropic's Leaked Numbers are INSANE
Tom Bilyeu
This Time Is Different, Again 0:00
The conversation opens with a warning about the familiar refrain that "this time is different," a phrase people use every market cycle right before things go wrong. The speaker argues the economy runs on mechanisms that are complicated but not magic, and that ignoring those mechanisms, like the way debt eventually has to be paid back through either a hard default or a soft default through inflation, is dangerous. He says this current moment does involve real structural change, but the failure patterns are still mechanical and predictable, and he expresses hope that an AI bust happens sooner rather than later, because waiting only makes the eventual damage worse.
Anthropic's Leaked IPO Numbers 4:00
Reuters reported on a leaked version of Anthropic's S1, the filing a private company must make when preparing to go public, and the numbers are described as staggering. Revenue grew twelvefold to nearly 4.6 billion dollars last year, which counts as genuinely fast growth even though the dollar figures look small next to the losses. Operating losses, meaning real cash spent on things like staff and equipment, hit 8 billion dollars, so the company spent far more than it earned. Including capital expenditures, the total loss for 2025 reached 42 billion dollars against that 4.6 billion in revenue. Much of that 42 billion, though, is explained as a non-cash paper loss tied to rising share values and buyback obligations rather than money actually leaving the bank today, which tempers some of the alarm even as the 8 billion real cash loss remains the number to watch.
Half a Trillion Already Committed 9:00
Anthropic has already committed 518 billion dollars in spending over the next few years, a sum that cannot be walked back regardless of performance. The scale is tied to the broader AI race to spin up intelligence at massive scale, limited mainly by compute, data centers, energy, and the capital to build them, which is pulling enormous amounts of money into the sector. Anthropic was last valued near a trillion dollars in equity markets and is reportedly seeking a two trillion dollar valuation despite the losses, with its own filing disclosing the risk that AI could end humanity. The segment closes by comparing this moment to the twenty years of pain that followed the dot-com bust, cautioning that hoping for a quick crash may be naive since there is no simple mechanism to stop or slow the momentum, only individual investors, banks, and lenders choosing whether to keep feeding it.
A regulatory capture fight ahead 14:32
The discussion turns to why Anthropic and OpenAI disclosed existential risk in their filings, noting this is a legal requirement from the SEC to flag material risks rather than pure alarmism. The hosts argue a bigger political collision is coming, with Democrats and Republicans becoming avatars for competing views on AI, Democrats likely warning of reckless spending and pushing for a nanny state, Republicans framing their approach as beating China and putting AI gains directly into Americans' hands through things like Trump accounts. Whoever wins in 2028 could determine whether Anthropic and OpenAI achieve regulatory capture that locks out competitors, or whether open source and competition continue, which changes where investment value ends up accruing, either to the dominant model companies or to the infrastructure layer underneath them.
The parabolic growth bet 21:01
Anthropic has floated spending commitments around half a trillion dollars, and the hosts explain that every technology and asset bubble runs on the same logic: spending is fairly predictable, but revenue gets extrapolated as parabolic, eventually crossing above costs to produce extreme profitability. They walk through rough numbers, suggesting maybe 100 billion in equity and 50 billion in revenue against that half trillion commitment, leaving roughly 350 billion that would need to come from debt, with the real total likely to grow well beyond half a trillion over time.
Debt, delay, and default risk 24:31
The core danger is that if revenue growth lags even slightly behind the extrapolated curve, the company needs far more debt, and investors grow less confident about ever being paid back. The hosts stress that a company can keep growing at a record pace and still go out of business if it cannot service its debt, pointing to historical patterns where exuberant spending outruns revenue and a later generation simply inherits the built infrastructure. Rising equipment costs, competition for limited hardware, and uncertainty about when profitability actually arrives all add friction, echoing past skepticism such as IBM's CEO doubting the numbers and comparisons to the dot-com era, when Cisco's CEO admitted investors had priced in a future, roughly ten times smaller than today's AI projections, that could never realistically have come to pass.
Growth That Defies Realistic Projections 29:00
The speaker draws on personal experience growing Quest Nutrition by 57,000 percent in three years, describing the terror of committing to months of inventory purchases based on continued doubling that might not happen. He uses this to explain why the AI numbers, while not realistic by normal standards, could still theoretically occur, though the risk of stalling out mid-growth is severe. He then compares the situation to Lehman Brothers in 2008, noting that no single dramatic event killed the firm, just a sudden collective unwillingness among lenders to extend overnight money. He warns that a similar dynamic could hit the AI boom, where no one person decides to pop the bubble, but credit simply becomes unavailable.
Everything Must Go Right 31:01
For Anthropic's valuation to make sense, spending must stay controlled, revenue must arrive as projected, and several risks must not materialize, including businesses switching to cheaper lower tier models or a macroeconomic downturn delaying revenue. Rising 10 year interest rates add pressure, since AI companies and the Treasury are competing for the same pool of cash, potentially forcing a standoff between the Federal Reserve's fight against inflation and the AI industry's determination to keep borrowing, framed around arguments about beating China. He suggests 518 billion dollars in commitments could become a trillion within a year, and that the longer the boom runs, the harder these numbers become to justify, making any eventual reckoning more severe.
Tangled Debt Echoes the Dotcom Era 34:01
The speaker argues this boom relies far more heavily on debt and credit structures, including special purpose vehicles, than the dotcom bubble did, and that everyone involved is financially entangled, citing Nvidia's reported guarantee to cover up to 25 percent of value declines in its own chips. He describes circular financing, such as Nvidia lending money to Anthropic that flows back to Nvidia through chip purchases, questioning whether any real new wealth exists. He details a specific Anthropic deal involving Broadcom's residual value support, Morgan Stanley, Apollo, and Blackstone, and notes Oracle's recent credit downgrade as an early warning sign. Drawing a direct parallel to Nortel and Lucent, which lent money to undercapitalized dotcom customers and eventually went bankrupt when revenue failed to materialize, he points to early distress at Blue Owl and BlackRock as signs the same pattern may be repeating, with guarantees rather than direct loans as the main structural difference this time.
Shadow Banks and Private Credit Demand 43:01
After the 2008 crash, regulators clamped down on banks, but that never made the underlying demand for risky lending disappear, it just pushed it into private credit and shadow banking. Asset managers now set up special purpose vehicles, raise money from wealthy investors and institutions, sell bonds, and pile on leverage to chase the hot opportunity of AI, since traditional lenders see it as too uncertain to touch. The result is a complex, financialized web where money gets packaged and sold into places it probably shouldn't be, leaving many people unaware of how exposed they actually are.
Collateral, Guarantees, and Hidden Risk 47:01
Companies like Anthropic buy chips from suppliers such as Broadcom or Nvidia and pledge that equipment as collateral to special purpose vehicles, which lets them get cheaper loans despite having no track record or real revenue, much like dot-com era startups. Because chips lose value fast, the chipmakers offer residual value support, essentially promising to take the equipment back and resell or repurpose it if the borrower defaults, so the lender feels protected. This only works as long as fast-growing AI revenue and steady debt obligations cross paths early enough to stay profitable, which recreates the same circular financing structure seen in the dot-com boom.
Confidence, Jobs, and Where to Place Bets 50:00
The 2008 crisis showed how quickly confidence can vanish and freeze overnight lending, and with AI companies now carrying obligations like one firm's half a trillion dollars in debt, a similar loss of faith could unravel things over a longer timescale. There is real uncertainty about how AI will affect jobs, with questions about white collar work left largely unanswered even by figures like Elon Musk, who shifts toward talk of an age of abundance and universal high income rather than addressing immediate job fears. Given a government running two trillion dollars a year in debt and a Federal Reserve raising rates on the wrong assumptions about inflation, the advice is to keep some exposure to AI without going all in, since history shows the real winners are those who kept cash ready through a crash and still believed in the long term thesis, and durable value may ultimately sit with infrastructure providers rather than the models themselves, though even that could shift if things move toward space-based computing.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
