Bernstein senior analyst: This is the first true "chip super cycle" in my career, "bottleneck" is the wealth building machine.

Bernstein's star analyst Stacy Rasgon believes that, as investment in AI infrastructure approaches a staggering 4.4% of the US GDP, the semiconductor industry is experiencing an unprecedented "super cycle". On June 21st, the Tech Surge Deep Tech Podcast, focused on cutting-edge technology, released a new interview transcript featuring Michael Marks, Managing Partner of Celesta Capital, and Stacy Rasgon, a renowned chip analyst at Bernstein. During the nearly one-hour conversation, the two delved deep into topics such as the revenue growth of semiconductors being driven by AI, the transition from model training to inference in AI, capacity bottlenecks at various stages of the supply chain, and the sustainability of future industry growth. Unlike most Wall Street analysts, Rasgon holds a PhD from the Massachusetts Institute of Technology and has an engineering background, which makes him value mature physical laws and capital flow more. Rasgon clearly stated that the semiconductor industry is currently experiencing the largest demand surge he has seen in his career. Last year, the total revenue of the semiconductor industry surpassed $800 billion, and this year it is racing towards a size of $1.3 trillion. In the interview, Rasgon exclaimed: "Throughout my career, I have always heard the term 'super cycle'. This may be the first one I have truly witnessed. The only voice we hear now is that no one has enough computing power." Rasgon emphasized that the market focus is shifting from "model training" to "AI inference", which is the core of achieving commercial monetization. At the same time, capacity bottlenecks are spreading from GPUs to HBM storage, semiconductor equipment, and even power supply. In the future, custom chips (ASICs) represented by Broadcom and GPUs from NVIDIA will coexist in the ever-expanding incremental market, collectively meeting the insatiable demand for computing power. The game of "whack-a-mole" in the supply chain is being forcibly driven by AI With the bottomless pit of AI computing power demand opened, the market is showing a peculiar "whack-a-mole" effect - capacity bottlenecks are erupting throughout the industry chain. Rasgon detailed this phenomenon: "Everything is being dragged by this insatiable demand for AI computing power. In my career, I have never seen such a scale of demand. The situation is spreading from accelerators to storage, to semiconductor manufacturing equipment, networks and optical devices, power semiconductors, and now even CPUs are in short supply." Using memory as an example, the industry is currently experiencing the strongest upturn in history, with prices doubling every quarter. The core driver behind this surge is HBM (High Bandwidth Memory). Rasgon revealed a key data point: "In a silicon wafer of an AI chip, over 85% may be HBM." Even more crucial is the issue of the "trade ratio". He said: "Due to yield losses in stacking technology and the space taken up by logic naked chips, producing 1GB of high-bandwidth memory requires about four times the silicon wafer area of standard DRAM." This means that even if wafer factories are ramping up production like crazy, the actual increase in storage capacity (in bits) remains severely limited. This extreme demand has even inadvertently benefited struggling companies. When discussing Intel's server CPU business, Rasgon bluntly pointed out that the current server demand is exceptionally strong, to the point where Intel even saw an increase in profit margins: "The demand is so strong that they even sold off inventory that had been written off as junk and thrown in a corner of the warehouse. Customers now say, 'We don't care, we want it, please sell it to us." The turning point is coming: "You can't make money just by training models" Although billions of dollars are flowing into the market, the biggest concern in the market is: Is this growth sustainable? Where is the space for imagination? Rasgon pointed to the turning point directly at "inference". He emphasized that while a significant amount of funding has been used for large model training, this is not the ultimate monetization of the business. Rasgon said: "You can't make money just by training models... you have to be able to use the model, that's inference." This transformation has already begun to be reflected in the astonishing data of startup companies. Rasgon cited data in the interview, such as companies like Anthropic, whose annual revenue run rate is showing a vertical rise: In December of last year it was $9 billion, in January it reached $14 billion, and most recently (April) it reached $30 billion." Furthermore, with NVIDIA's recent acquisition of Groq, the segmented demand for inference markets is becoming evident. Rasgon pointed out that not all data "tokens" have the same value. For specific inference tasks that require very low latency and fast response times, customized chips or dedicated inference architectures often have better economy than general-purpose GPUs. ASICs and GPUs are not a zero-sum game With the explosion of inference demand, the momentum of customized chips (ASICs) is challenging the absolute dominance of GPUs. Broadcom has become the biggest beneficiary of this trend. When mentioning Broadcom, Rasgon said: "Before all of this began, Broadcom said that semiconductors were a mature industry with a growth rate in the mid-single digits. But now everything has exploded. (Broadcom) they said they think they can achieve $100 billion in AI revenue next year." Why do major cloud services insist on developing their own ASICs? Rasgon believes that this is not only for performance optimization, but also to have a bargaining chip in negotiating with NVIDIA's high 75% gross profit margin. Rasgon said: "At least when you sit at the negotiating table with Jensen Huang to discuss next year's contracts, youll want to have a backup plan in your pocket." However, Rasgon emphasized that this is not a game of who replaces whom. If ASICs occupy a larger share, it is because the entire pie has grown bigger. For large, stable, and internally developed workloads, ASICs can provide lower total cost of ownership; but if the model structure changes, the programmability advantage of GPUs is irreplaceable. Rasgon believes: "The real pain point is: is the opportunity in front of us still growing? If it's big enough, they will both thrive; if it's not, everyone will be in trouble." The ultimate ceiling of the future: The power grid may not hold up When asked about the risks that the market may be overlooking, Rasgon shifted the focus from code and silicon chips back to the physical infrastructure of the real world - electricity. Currently, the capital expenditures of the technology giants this year have reached $600 billion. If future infrastructure spending grows according to NVIDIA's expectations of $3 to $4 trillion annually, the existing human energy system will face collapse. Rasgon shared a modeling he had previously developed: "Do we really have enough power to do this? The power grid may not be able to withstand it. The US's electricity capacity needs to grow by about 5% annually over the next ten years. And, in the eyes of power equipment analysts, a 5% annual growth rate is simply unattainable." This means that the next breakthrough and bottleneck in AI innovation will inevitably fall in areas such as energy generation, cooling, and nuclear power. As he has always believed: "Never underestimate human ingenuity. If it's profitable, engineers will always find a way out." Overall, as long as the demand for AI does not undergo a cliff-like collapse, the "super cycle" of the entire semiconductor industry will continue, and the focus in the equity capital markets must closely follow the "capacity bottlenecks" that are constantly surfacing in various aspects of the industry.