BlockBeats report: On June 21, Bernstein’s renowned chip analyst Stacy Rasgon openly stated that this is the first time in his 18-year career that he has genuinely witnessed a semiconductor supercycle. With a Ph.D. from MIT and a background as an engineer, Rasgon’s data is staggering: the global semiconductor industry generated over $800 billion in revenue last year and is now racing toward $1.3 trillion this year. Every segment—from accelerators and memory to semiconductor equipment, network optical communications, power chips, and CPUs—is experiencing severe supply shortages. “The only consensus we’re hearing is that no one has enough computing power. Take memory as an example: HBM may account for over 85% of the silicon area in AI chips, and the silicon area required to produce 1GB of HBM is roughly four times that of standard DRAM—meaning even if fabs expand production at full capacity, the actual increase in storage capacity remains extremely limited. This supply-demand imbalance has even benefited Intel—its inventory, previously written down to zero value, has been completely sold out, with customers saying, ‘We don’t care—just sell it to us.’”
Rasgon noted that the industry’s core focus is shifting from model training to AI inference, which is key to achieving commercial monetization—training models themselves does not generate revenue; it is the use of models that creates value. According to Anthropic’s data, annualized revenue surged from approximately $9 billion in December last year to $30 billion in April this year, nearly vertical growth. In the chip competitive landscape, custom ASICs represented by Broadcom and NVIDIA GPUs are not a zero-sum game: “The right question is whether the opportunity is still growing—if it’s large enough, both will thrive.” Broadcom currently expects its AI revenue to reach $100 billion next year; ASICs currently account for around a few percentage points of AI chip market revenue and are expected to rise to 25%-30% in the future, but will not fully replace GPUs. Regarding AI inference chip startups like Groq, recently acquired by NVIDIA, Rasgon cited Jensen Huang’s assessment: not all tokens are equal—low-latency tokens hold higher value, and GPUs are not always the optimal choice for every task.
When asked about the most overlooked risks in the industry, Rasgon shifted focus from silicon back to the physical world—electricity. Calculations show that if NVIDIA’s projected annual infrastructure investment of $3 to $4 trillion materializes, the U.S. power grid would need to expand by approximately 5% per year—a growth rate that power industry analysts consider nearly impossible to achieve. This means the next wave of bottlenecks will emerge in energy generation, cooling, and nuclear power. “But never underestimate human ingenuity—engineers always find a way when there’s money to be made.” Regarding Intel, new CEO Chen Liwu’s pragmatic, low-expectation strategy, along with better-than-expected yields for its new 18A process node and substantial funding from the government and NVIDIA, has significantly alleviated prior market concerns over its balance sheet. Rasgon concluded that as long as AI demand doesn’t collapse, the full-chain supercycle will continue, and capital markets must closely track capacity bottlenecks across every stage of the supply chain.