Just finished watching the latest episode of All In Podcast, discussing CoreWeave's debt crisis, Anthropic's IPO financial figures being questioned, and the regulatory calculations behind doomsday theories Second tier computing power clouds like CoreWeave are experiencing a real debt crisis The beginning of this issue is a hardcore financial breakdown. GPU computing cloud companies such as CoreWeave and Nebius saw their stock prices plummet by about 20% this week, and the debt yield on CoreWeave's secondary market soared to 15%. Musk commented on Yikes, and the news quickly spread globally. Chamath pointed out that smart hedge funds price the default risk of such companies through credit default swaps (CDS). Currently, the spread has soared to 953 basis points, which translates to an annual default probability of 16%. The cumulative default probability during the five-year term is as high as 55%, which is basically the odds of throwing a coin. The core business model of such companies is a project isolation structure called The Box, which signs long-term fixed price contracts (such as 5-year terms) with large clients like Microsoft, packages the contracts and GPU assets into a separate entity, and then uses this entity to borrow asset collateral loans from banks, agreeing to prioritize the use of cash flow to repay data center rent, electricity bills, and principal and interest. This model has now exposed several fatal issues. Firstly, it has a huge financing exposure. CoreWeave has already taken on about $35 billion in debt and needs to raise an additional $35 billion in the capital market to complete the contracted data center construction. Secondly, there is a pincer attack on interest rates and interest rate spreads. The risk-free interest rate on the long end of US Treasury bonds remains high, and its average borrowing cost of about 8.5% is facing the risk of repricing. If the marginal borrowing cost really rises to 12% to 15%, it will directly offset its already thin net profit margin. According to the Chamath team's calculations, it is only $5 billion away from a true breach of financial contracts. Finally, the customer base is extremely concentrated, with 75% of revenue relying on the top three customers, all of whom have fixed price contracts and cannot rise with the computing power premium like in the spot market. The guests had disagreements on this matter. Sacks is relatively optimistic, believing that this is more of a market fluctuation in the early stages of high growth, rather than a fundamental problem. The depreciation period of the underlying GPU assets is as long as 6 to 8 years (the old generation H100 is still in short supply), and the lessee is a giant with investment grade credit such as Microsoft. He believes that the bankruptcy probability given on Polymarket is severely overestimated. Ben Horowitz reviewed the lessons of LoudCloud in 1999 from the perspective of a16z (when bandwidth prices plummeted from $1500 to $50, directly destroying a group of heavy asset borrowing companies), and explained why a16z firmly refused to invest in such second-line computing clouds. Essentially, this is an extremely dangerous capital arbitrage machine, and even a slight headwind could lead to a reversal. However, he also admitted that unlike the lack of real demand in the Internet foam in 2000, the current demand for AI is absolutely real and has been in short supply for a long time. Andy Jassy of Amazon said that the gap in computing room has been ranked after 2028. Anthropic sprints towards a $2 trillion IPO, with a growth rate mystery hidden in its financial data Anthropic plans to launch its IPO around Thanksgiving (late November). During the closed investor day at its San Francisco headquarters, it disclosed Q3 quarterly revenue of $16.8 billion (averaging over $5 billion per month), a net asset retention rate of 900%, an adjusted gross profit margin of 85%, and positive free cash flow. But the core of the market controversy lies in ARR's estimation, which is said to be between $70 billion and $100 billion. If ARR is only $70 billion, it means that there is almost no growth compared to July, and the month on month growth rate has significantly declined. This will put the IPO target of $2 trillion (about 30 times the forward market to sales ratio) to a severe test. Chamath gave an interesting reverse suggestion, calling for fake names and real gold. He firmly opposes raising only $50 billion to $80 billion at the high level of $1.9 trillion to $2 trillion, and suggests that Anthropic should proactively lower its valuation to $1 trillion to $1.5 trillion, but directly raise $200 billion to $250 billion in cash reserves from the public market. His core logic is that AI computing power will require gigawatts of electricity in the future, and $200 billion is a life and death trump card to build 5 to 10 GW of computing infrastructure and completely shake off the volatility of the spot market. Valuation is just vanity, and the cash reserves in hand can withstand potential macroeconomic tightening in the next 2 to 3 years. SemiAnalysis's calculations show that OpenAI and Anthropic will consume 60% of the global new computing power in the next year, effectively forming a duopoly monopoly of computing power and revenue. Ben Horowitz proposed a unique perspective that traditional investors have to buy these two IPOs because regardless of whether you hold stocks in the SaaS, healthcare, or finance industries, the general intelligence sold by these two companies is threatening and reshaping the commercial moat of all existing companies. Buying them is essentially paying a mandatory premium for the possibility of traditional stock pools being overturned. The software industry is experiencing deflation, but the moat is shifting to the harder core dimension of electricity Jason raised a concern that software is entering a vicious deflation, with open-source models and low-cost intelligent agent tools rapidly flattening the premium space of traditional software. Sacks and Ben's response is that although the marginal cost of single line code does tend towards zero, this would trigger the classic Jevons paradox. The cheaper the code, the greater the demand for customized software and intelligent agent workflows in society, leading to a significant increase in total output and code volume. This is not a zero sum game. Chamath made a judgment. The real moat of the future is electricity. The United States will face a power gap of 1.4 trillion kilowatt hours by 2050, equivalent to the total electricity consumption of six California states. The real constraint on the development of excellent startups in the future is no longer the algorithm itself, but whether cloud providers have surplus megawatt/gigawatt power quotas to allocate to new product launches. Sacks proposed a more practical North Star indicator: capital recovery cycle. CoreWeave relies on a 5-year long-term contract with a payback period of approximately 2.5 years. SpaceX relies on a 90 day flexible spot trading mechanism, with a payback period of only about one year (the monetization efficiency of a single gigawatt reaches $50 billion, far higher than CoreWeave's approximately $20 billion). His conclusion is that as long as the investment return cycle can be maintained within the extremely short range of 1 to 2.5 years, the underlying business model has considerable resilience against cycles. Behind the doomsday theory may be a carefully designed regulatory arbitrage Sacks and Ben Horowitz harshly criticized Anthropic's blog post this week, which warned that small language models may infect IoT devices/smart appliances, forming worm like autonomous replication. They directly described this statement as a ridiculous panic PR that crosses from Terminator to Killer toaster. They analyzed the true intention behind this. Companies like Coinbase have recently started using fine tuned open-source small models (such as Qwen-9B) for anti fraud detection, which have lower costs and faster speeds than Claude's top of the line model. The open-source small models are truly eating up the real load share of enterprise level. Several guests believe that Anthropic is trying to lobby Washington for legislation to impose restrictions on open source models by creating security panic, essentially aiming to achieve a regulatory capture. They also reflected on the deep psychological mechanisms behind the narratives of tech veterans like Bill Gates depicting AI causing one billion deaths/human survival crises, believing that this is more like these people trying to regain a non elected voice and moral high ground by establishing global governance institutions and spreading fear narratives after losing their influence in their business. The guests unanimously agreed that the current model's training computing power, time, and generalization ability have not even reached the basic threshold of recursive self evolution. To talk about the existence crisis without realistic engineering constraints not only misleads the public, but also seriously hinders the normal pace of technological innovation.