From Leopold to the White Haired Stock God to Chen Liwu, the common trump card of AI stock gods: a list of the top ten bottlenecks Recently, there has been a highly consensus framework in the semiconductor investment circle: identifying bottlenecks and investing in them. Leopold Aschenbrenner founded Situational Awareness LP with $225 million, achieving $5.5 billion in 12 months, with all core holdings focused on AI physical infrastructure such as power, computing power, memory, and optical connectivity. The number of paid subscribers on the X platform exceeds that of Musk's Serenity (the "white haired stock god"), who relies on the "bottleneck theory" to select small cap stocks, claiming an annualized return of 225 times. He called for a two-day increase of 30% in A-shares and a direct limit up in Yishite. Intel CEO Chen Liwu's recent interview on the No Priors podcast further emphasized this theory. Prior to taking over Intel, Chen Liwu served as the CEO of Cadence for twelve years, during which his stock price increased 32 times. He is also one of the most active venture capitalists in the semiconductor industry, having personally invested in over 200 semiconductor companies, including 159 IPOs. He dismantled the bottleneck list of AI infrastructure diffusion from GPUs to the entire supply chain one by one in the interview. The following is the complete bottleneck list mentioned by Chen Liwu in the interview. Interconnected. Large model training requires thousands or even tens of thousands of GPUs for collaborative computation. No matter how powerful a single GPU is, if the data transmission speed between chips cannot keep up, the actual utilization rate of the entire cluster will be dragged down. The current mainstream copper cable interconnection solutions are approaching the physical bandwidth limit, which means that high-speed interconnection chips and new interconnection architectures have become capital intensive investment directions. Chen Liwu has personally invested in the internet chip company Credo and the optoelectronic fusion architecture company Celestial AI. He mentioned in the interview that Huang Renxun has almost invested in all companies that do optical interconnection, which in itself indicates the strategic priority of this link. Photons. This is the next generation technological solution to the bottleneck of interconnectivity. Electrical signals suffer from signal attenuation and heat generation in long-distance and high-density transmission scenarios, while optical signals have physical advantages in these two aspects. However, photonic chips are not yet mature in terms of manufacturing processes, packaging integration, and cost control, and there is a clear gap between production capacity and the demand growth rate of AI clusters. EDA。 The entire process of chip design to chip fabrication relies on EDA (Electronic Design Automation) software. With the increasing complexity of chips (more transistors, smaller processes, more heterogeneous integrations), the computational workload that EDA tools need to handle is growing exponentially. The current global EDA market is mainly dominated by Synopsys and Cadence. During his twelve years as CEO of Cadence, Chen Liwu personally boosted the company's stock price by 32 times and has first-hand knowledge in this field. He said in the interview that several startups in the EDA field are digging deep into it, and 'this is a gold mine'. Advanced packaging. GPU and HBM (High Bandwidth Memory) are not independently installed on the motherboard, but are interconnected through advanced packaging technology on the same silicon intermediate layer. TSMC's CoWoS (Chip on Wafer on Substrate) is currently the mainstream solution, while Intel's EMIB (Embedded Multi die Interconnect Bridge) is a competing solution. Starting from 2024, CoWoS production capacity continues to be tight, and TSMC has been expanding production but still faces supply shortages. The direct consequence of insufficient packaging capacity is that even if GPU and HBM chips are produced, they cannot be assembled. Power consumption conversion. The voltage transmitted by the power grid is much higher than the voltage required for chip operation (about 1V), and multiple levels of conversion are required in between, each of which generates energy loss. Chen Liwu specifically pointed out the issue of losses during the process of reducing from 40V to 1V. The power density of AI servers is extremely high, and every one percentage point decrease in power conversion efficiency means that a large amount of electricity is wasted as waste heat at the scale of data centers, while increasing the burden on the cooling system. Silicon carbide (SiC) and gallium nitride (GaN) power devices have better conversion efficiency than traditional silicon-based devices, but their production capacity is also in a state of scarcity. dissipate heat. The direct physical consequence of increased power consumption is an increase in heat density. When the chip temperature exceeds the design threshold, the system will automatically downshift or even shut down to protect the hardware. The power consumption of Nvidia's next-generation GB200 NVL72 cabinet can reach 120kW, and traditional air cooling solutions are no longer effective in dissipating heat at this power density, making liquid cooling a rigid requirement. But the components involved in the liquid cooling system, such as cold plates, pipelines, CDU (cooling distribution unit), coolant, etc., form a brand new industrial chain, and production capacity is in the stage of climbing from scratch. New materials. The process miniaturization of traditional silicon-based semiconductors is approaching the physical limit, and the performance improvement brought by further shrinking the process is becoming increasingly limited. Breakthroughs must be found at the material level. Chen Liwu revealed that he has invested in companies in the fields of GaN (gallium nitride), SiC (silicon carbide), and InP (indium phosphide), some of which have been acquired by large semiconductor companies such as Analog Devices. He also invested in a company that manufactures diamond wafers and is optimistic about the potential of diamonds as thermal conductive materials in chip packaging (diamonds are currently one of the materials with the highest known thermal conductivity). In addition, he invested in 3DGS, a glass substrate company. Compared to traditional organic substrates, glass has better heat dissipation performance and warpage control, making it suitable for the next generation of large-sized packaging needs. These materials correspond to bottlenecks in different stages: GaN and SiC are used for power conversion, InP is used for optical communication and interconnection, artificial diamonds are used for heat dissipation, and glass substrates are used for advanced packaging. In his own words, "This is the spirit of an engineer. You keep hitting bottlenecks and then find ways to jump over or bypass them Memory. The full utilization of GPU computing power depends on whether the memory can continuously supply data with sufficiently high bandwidth. If data cannot be fed in, the actual utilization rate of GPU will be significantly lower than the theoretical peak. Chen Liwu directly said in the interview that 'memory is the biggest shortage' (memory is currently the biggest shortage), and everyone is competing for memory production capacity. The latest export data from South Korea also provides side validation: not only are HBM exports growing, but the export value and unit price per kilogram of ordinary DRAM, NAND, and SSD are also strengthening synchronously, indicating that AI demand is overflowing from a single category of HBM to the entire storage system. SK Hynix, with a market share of approximately 59% in the HBM field, surpassed Samsung in market value for the first time today, with an operating profit margin of 72%. Helium gas. In an interview with Zhongyuan, Chen Liwu said, "Many people are not aware that helium also has a significant impact on semiconductors. Multiple key processes in the semiconductor manufacturing process, including photolithography, etching, vapor deposition, cooling detection, etc., require the use of high-purity helium gas. Helium is a non renewable and scarce resource, with a high concentration of global supply in a few natural gas fields, large price fluctuations, and unstable supply. If the helium supply is interrupted, the production of the wafer fab will be directly affected. Electricity. This is the lowest constraint of the entire bottle neck chain. Chen Liwu pointed out that some countries simply do not have the power capacity needed to support the growth of AI infrastructure. The electricity consumption of a single large AI data center can reach several hundred megawatts, equivalent to the total electricity consumption of a small or medium-sized city. The power grid in many regions does not have additional capacity, and building new power generation and transmission infrastructure usually takes several years. Without electricity, all upstream chip, memory, packaging, and interconnect investments cannot be converted into actual computing power output. In the past, the market only focused on Nvidia, but now bottlenecks are spreading along the supply chain to every link. That's also why almost all semiconductor analysts and investment influencers have been telling stories using the same framework recently: finding bottlenecks, investing in bottlenecks. But bottlenecks are not eternal. The link with pricing power will attract capital to expand production. Around 2028, Micron, Hynix, and Samsung will release new production capacity, and the bottleneck may become overcapacity, and pricing power will disappear. At least until the end of 2027, there is no sign of easing the supply-demand imbalance on this chain.