"The Big Short" Chanos: There is a huge "financial mismatch" in the current AI industry chain, "AI Cloud" is actually a leasing intermediary, bearish on SpaceX.

When the "Water Sellers" under the AI tide make a fortune, what is the real return on investment in the computing power infrastructure - a money printing machine limited by physical constraints, or a feather in the wind? Wall Street's top short seller and technology bull provide completely different answers. On June 21, at the Macro Minds Symposium hosted by Jack Farley, Wall Street legendary short seller and "Enron terminator" Jim Chanos, and Val Zlatev, partner of Analog Century Capital, engaged in a deep debate on the prosperity of AI capital expenditure, the semiconductor cycle, and the business model of computing power infrastructure. The legendary short seller Chanos warned that the current technology industry is experiencing a capital expenditure boom similar to the late 1990s, where chip suppliers can immediately recognize revenue, while hyperscale cloud providers capitalize on AI capital expenditures. Hyperscale cloud providers amortize over four to seven years, rather than directly entering operating expenses. This "timing mismatch" led to a 40% profit drop in technology stocks that year. In addition, Chanos believes that compute leasing (such as CoreWeave) is fundamentally a low single-digit return financial leasing business. Val Zlatev, a Wall Street investor with a background in physics, countered by noting that there is currently a surge in demand for tensor processors, and as a result, rental prices for old GPUs have been soaring. He also mentioned that Nvidia's valuation (15 times 2027 EPS) is far from the bubble level of 1999. He further pointed out that due to the maximum annual growth rate of semiconductor equipment capacity of only 30%, the scarcity and high price cycle of storage chips will last longer than market expectations. Disparity in profits and "Depreciation time bomb" In the current AI capital expenditure boom, the market is most concerned about the flow of profits. Chanos pointed out the significant imbalance in financial statements between upstream and downstream companies in the industrial chain. Chanos warned: The current profit accounting is disconnected: those companies selling 'picks and shovels' (chips, data center equipment suppliers) are immediately recognizing revenue and profit; while those super large-scale cloud service providers spending massive amounts of money are capitalizing these costs. He recalled the period from 1998 to 2000 during the internet bubble, when operating profits in the S&P 500 increased by 30% in just two years. However, when the order book collapsed in 2001 and depreciation costs continued to appear, operating profits in the S&P 500 plummeted by 40%. To be conservative, Chanos assumes the physical life of GPUs is 10 years, but he still doubts the downstream profitability. He explained: You need to be careful, the costs of buying chips and building data centers are currently recorded in 'construction in progress.' Once they go online and start depreciating, the impact on profits is huge. Zlatev agreed with this point but added that the actual economic life of GPUs is difficult to accurately define - it may not be 10 years, it may be 6 years, but certainly not 2 years. Is compute leasing a technology company or a financial intermediary? Regarding the current hot emerging cloud service providers (such as CoreWeave and other compute leasing platforms), Chanos gave a very pessimistic evaluation, stating that this is essentially an unprofitable business model. Chanos bluntly stated: If you buy chips from Nvidia, rent data centers from others, and then sublet computing power to Microsoft or Google, you are an equipment leasing company, not a technology company, but a financial company. Chanos further stated: You should be long on what chips produce, not where the chips are located. He revealed the true return on investment for compute infrastructure buildouts (ROIC): Current transaction details show that if you currently have powered data centers and chips, the expected pretax ROIC is only in the single digits of 5%, 6%, 7%, 8%, all being single digits. If that's all you can do now (short at a time when there is a shortage), I'd rather hold other parts of the value chain. Zlatev agreed with this to some extent but pointed out that not all emerging cloud service providers are exactly the same. He gave the example that Lambda Labs has about 40% to 50% of its revenue coming from real-time inference scenario spot pricing, giving it some flexible profit margins in the current environment of surging GPU spot prices. Both agreed that the real value is not in providing physical carriers for racks and power, but in chips and their "packaging" (software and optimization layers). Bull argument: Real demand and far from a bubble valuation In response to lessons learned from the dot-com bubble