Bernstein senior analyst: The first true chip supercycle is coming; the "bottleneck" is the real wealth-creating machine.

According to Mars Finance, on June 21, Bernstein's star chip analyst Stacy Rasgon stated that this was the first time in his 18 years in the industry that he had truly witnessed a semiconductor supercycle. Rasgon, who holds a PhD from MIT and is an engineer by training, provided some astonishing data: the semiconductor industry's total revenue exceeded $800 billion last year and is heading towards $1.3 trillion this year, with all sub-sectors, from accelerators to memory, semiconductor equipment, network optical communications, power chips, and even CPUs, experiencing a comprehensive shortage of supply. "The only consensus we're hearing right now is that no one has enough computing power. Take memory as an example: HBMs (High-density Memory) may account for over 85% of the silicon wafer area in AI chips, and the silicon wafer area required to manufacture 1GB of HBM is about four times that of standard DRAM. This means that even if wafer fabs expand production at breakneck speed, the actual increase in storage capacity will still be extremely limited. This supply-demand mismatch has even benefited Intel—its inventory, which had already been written off to zero, was snapped up. Customers' attitude was, 'We don't care, please sell to us.'" Rasgon pointed out that the industry's core focus is shifting from model training to AI inference, which is the key to commercial monetization—training the model itself doesn't generate revenue; only using the model can generate income. Anthropic data shows that annualized revenue surged from approximately $9 billion in December last year to $30 billion in April this year, almost a vertical increase. In the competitive chip landscape, the competition between custom ASICs, represented by Broadcom, and Nvidia GPUs is not a zero-sum game. "The real pain point is whether the opportunity is still growing—if it's big enough, both will thrive." Broadcom currently expects its AI revenue to reach $100 billion next year, with ASICs accounting for about ten percent of the AI chip market revenue, potentially rising to 25%-30% in the future, but not completely replacing GPUs. Regarding inference chip startups like Groq, recently acquired by Nvidia, Rasgon quoted Jensen Huang's assessment: not all tokens are created equal; low-latency tokens are more valuable, and GPUs are not the optimal choice for all tasks. When asked about the most overlooked risk in the industry, Rasgon shifted the focus from silicon wafers back to the physical world—electricity. It is estimated that if Nvidia's predicted annual infrastructure investment of $3 trillion to $4 trillion comes to fruition, the US power grid will need to expand by about 5% annually, a 5% annual growth rate that power industry analysts consider almost impossible. This means the next bottleneck will fall on energy generation, cooling, and nuclear power. "But never underestimate human creativity; engineers can always find a way when there's profit to be made." Regarding Intel, the new CEO Chen Liwu's pragmatic strategy of low expectations, the better-than-expected yield of the new 18A chip's process, and government and Nvidia's investment have significantly alleviated previous market concerns about its balance sheet. Rasgon concluded that as long as AI demand doesn't collapse, the supercycle across the entire industry chain will continue, and the capital market's focus should be on the capacity bottlenecks that exist at each stage.