Goldman Sachs: AI Computing Power Scarcity Remains Central Theme; Bulls vs. Bears Debate: 1999 or 2000?

Privorotsky pointed out that the market is currently in a period of extreme scarcity for AI infrastructure. Datasets tracking daily rental prices of GPUs show extreme price increases in the past few months due to the worsening shortage of computing power and the continuous climb in token consumption. Notably, the resilience of the A100 stands out, with the current rental price for it being around $1.50 per hour. Despite the upcoming new generation Rubin and Vera Rubin architecture, the A100 still maintains stable demand - this is not only a microcosm of the overall shortage of computing power but also reflects the adaptability of businesses to existing computing power when optimizing token efficiency. The massive operation of intelligent agent processes demands significantly higher memory and context window requirements than computing power itself, which also supports the continuous demand for the A100 to a certain extent.

There has been a noticeable cooling in sentiment. The AAII bull-bear spread dropped to -11.9, and the NAAIM survey also weakened, indicating that the market has become more cautious overall after a recent momentum pullback. Regarding the situation in Iran, Privorotsky believes that the stock market has gradually priced in a reduced probability of renewed conflict, despite the two sides still having significant differences on core issues. However, he expects more positive statements could emerge by the weekend.

PMI data presents intriguing structural characteristics. The U.S. manufacturing PMI has risen to a multi-year high, driven mainly by inventory hoarding, cautious stockpiling by suppliers, and rising prices, rather than a genuine expansion in terminal demand; the services sector, employment, and exports have all weakened in tandem. The European service industry shows clear fatigue. The PMI report warns that current precautionary stockpiling of inventories cannot support long-term growth, and the elongation of supply chains typically takes about six months to pass through to consumer inflation, making the GDP growth rate for the second quarter unlikely to break through 1%.

The logic for both bulls and bears is equally clear. The bull case is supported by: continued repurchase efforts, active participation of retail funds, an improving trend in EPS revisions, and the bottleneck for AI has spread from computing power to memory, electricity, and industrial inputs, meaning the widening of constraints implies that supply gaps are unlikely to be bridged in the short term, much like in the early stages of 1999. The bear case, on the other hand, is concerned that some AI assets have already factored in extremely aggressive demand assumptions, bearing uncomfortably close similarities to the fiber-optic bubble of the late 1990s - when fiber optics seemed permanently scarce until engineering capabilities caught up with demand, rapidly shifting scarcity to surplus. This does not mean that demand for AI will disappear, but the speed at which supply arrives may be faster than the market anticipates. Interest rate risk is also not to be ignored, as increasing term premiums and rising bond market volatility historically have not been friendly to stock markets, even though the current market appears unusually calm about this - perhaps because the downward trend in oil prices is gradually becoming a market consensus.