Bernstein Analyst Stacy Rasgon: Semiconductor Enters Super Cycle, All Segments in Short Supply
Bernstein star chip analyst Stacy Rasgon stated that this is the first time in her 18-year career that she has witnessed a semiconductor super cycle. Last year's industry revenue exceeded $800 billion, and this year it is heading towards $1.3 trillion, with shortages across accelerators, memory, equipment, optical communication, power chips, and CPUs.
In AI chips, HBM occupies over 85% of the wafer area, and the silicon wafer required to manufacture 1GB of HBM is four times that of standard DRAM. Even with capacity expansion, the increase in storage capacity remains limited; the industry's focus has shifted from training to inference, with Anthropic's annualized revenue soaring from $9 billion last December to $30 billion this April.
Broadcom's ASIC and NVIDIA's GPU are in a non-zero-sum game, with the ASIC market share expected to rise from a few percentage points to 25-30%. Power is the next bottleneck, as expanding the U.S. grid by 5% annually is extremely challenging, but engineers always find a way.
Source: Public Information
ABAB AI Insight
Stacy Rasgon, an analyst with an engineering background, bases her judgment of this super cycle on supply and demand data, continuing the cyclical characteristics of the semiconductor industry, similar to past AI-driven shortages in memory and logic chips.
On the capital front, the supply-demand mismatch drives funds to concentrate on HBM, ASIC, and inference chips, with resources leaning towards power, cooling, and nuclear energy infrastructure to alleviate bottlenecks.
Similar to the chip shortage in 2021, the current AI computing power is in a phase of commercial transformation from training to inference, highlighting bottlenecks across the entire industry chain.
Essentially, this is about capital concentration: under the super cycle, capacity bottlenecks push capital towards key segments, accelerating technological substitution and reshaping the pricing power of the semiconductor and energy supply chains.
ABAB News · Law of Cognition
Comprehensive short supply indicates a super cycle: all segments, AI demand is limitless.
HBM wafer area is 4 times that of DRAM: storage bottleneck constrains the speed of computing power expansion.
Inference commercialization outpaces training burn rate: Anthropic's revenue sharply rises, with a clear monetization path.