Senior Analyst at Bernstein: The First True Chip Supercycle is Coming, 'Constraint' is the Wealth Creation Machine

BlockBeats News, June 21st. Bernstein's star semiconductor analyst Stacy Rasgon recently stated that this is the first time in his 18-year career that he has truly witnessed a semiconductor supercycle. Rasgon, who holds a Ph.D. from MIT and has an engineering background, provided staggering data: the semiconductor industry's total revenue surpassed $800 billion last year and is sprinting towards $1.3 trillion this year. All subsectors, from accelerators to memory, semiconductor equipment, network optics, power chips, and even CPUs, are facing an unprecedented supply shortage. "The only consensus we hear now is that no one's computing power is sufficient. Taking memory as an example, HBM in AI chips may account for over 85% of the die area, and producing 1GB of HBM requires approximately four times the die area of standard DRAM, meaning that even with insane wafer fab expansions, the actual storage capacity increment remains very limited. This supply-demand mismatch has even benefited Intel—its previously written-off inventory has been completely scooped up, with customers expressing, 'We don't care, sell it to us.'"

Rasgon pointed out that the industry's core focus is transitioning from model training to AI inference, which is key to achieving commercial monetization. Training models themselves do not make money; it is the utilization of models that generates revenue. According to Anthropic data, annualized revenue soared from around $9 billion in December last year to $30 billion in April this year, almost a vertical surge. Regarding the chip competition landscape, the custom ASIC represented by Broadcom and Nvidia's GPU are not in a zero-sum game. "The right pain point is whether the opportunity is still expanding—if it is significant enough, both will thrive." Currently, Broadcom expects its AI revenue to reach $100 billion next year, with ASICs accounting for about a dozen percentage points of the AI chip market revenue share, expected to rise to 25%-30% in the future, but they will not completely replace GPUs. For inference chip startups like Groq, recently acquired by Nvidia, Rasgon quoted Huang Renxun's judgment: "Not all tokens are alike; low-latency tokens have higher value, and GPUs are not the optimal choice for all tasks."

When asked about the industry's most overlooked risk, Rasgon shifted the focus from silicon back to the physical world—electricity. It is estimated that if Nvidia's projected annual $3 to $4 trillion infrastructure investment materializes, the U.S. power grid would need to expand by about 5% annually, a rate that power industry analysts see as nearly an insurmountable task. This implies that the next bottleneck will be in energy generation, cooling, and nuclear power. "But never underestimate human creativity; engineers always find a way out when it's profitable." Regarding Intel, the new CEO Chen Liwu's pragmatic strategy with low expectations, the new 18A chip process yield is better than expected, and government and Nvidia's funding have significantly alleviated previous market concerns about the balance sheet. Rasgon concluded that as long as AI demand does not collapse, the full industry supercycle will continue, and the capital markets need to focus on the capacity bottlenecks at various stages.