AI computing power bottleneck migrating upstream! Behind Micron's investment in GlobalWafers: going from a "chip shortage" to a "silicon wafer grab" in the storage super cycle.

As part of the expansion of the US domestic chip manufacturing investment plan, Micron Technology, Inc. (MU.US) has invested in the leading silicon wafer manufacturer headquartered in Taiwan, GlobalWafers, to provide critical financial support for its large wafer manufacturing facility in Sherman, Texas. The well-known Wall Street investment firm, Wedbush Securities, believes that the supply of core raw materials for chip production, silicon wafers, may ultimately become another "key AI computing hardware" bottleneck faced by the AI computing industry and even the global technology industry. According to senior stock market analysts from Wall Street giants such as Goldman Sachs Group, Inc., Morgan Stanley, and Nomura, the main theme of the global AI super bull market, including US stocks, is shifting from "NVIDIA Corporation's AI GPU demand surge" to "bullish wave on AI computing infrastructure across the entire industry chain," which will continue to support the AI-centric super bull market trends in US stocks and global stock markets. The most typical example undoubtedly lies in the latest performance of NVIDIA Corporation, which clearly highlights that the global AI computing infrastructure construction craze is far from over. It is expanding from AI GPU/AI ASIC to data center CPUs, high-performance network infrastructure, enterprise-level HBM/DRAM/NAND storage, complete machine-level server clusters, AI super factories, and enterprise-level large-scale AI cloud computing systems. On Wall Street, analysts' bullish sentiment towards the "global AI leader" NVIDIA Corporation is growing hotter, with the average target price on Wall Street suggesting that NVIDIA Corporation's market value is expected to exceed $7 trillion. The latest research report released by the globally renowned research firm IDC shows that the world's highest-valued company, the AI chip superpower NVIDIA Corporation (NVDA.US), has become the largest supplier of global data center Ethernet switch market measured by revenue. The latest report by IDC is consistent with the views of Wall Street giants such as Morgan Stanley, Goldman Sachs Group, Inc., and Bank of America Corp., that leaders in the AI computing industry chain like NVIDIA Corporation are expanding the scope of computing power coverage from "GPU/AI chip single-point control" to "GPU/CPU cabinet clusters + storage chips + Ethernet switch clusters + DPU + optical interconnect systems + software ecosystem + data center power chain + AI server manufacturing" and other aspects of the "AI factory system level" closed loop. Therefore, the "cake" of the AI super bull market is expanding and continuously growing from GPU core assets to the hardware bottlenecks related to full-stack AI computing infrastructure. While NVIDIA Corporation still sets the tone for the AI bull market, other components such as PCBs, MLCCs, ABFs, ODMs, liquid cooling, power supplies, silicon wafers, CMP consumables, photoresists, glass substrates, SOI/InP, data center optical interconnect equipment, optical modules, and advanced packaging equipment may all undergo a new round of revaluation. The core supply capabilities of global advanced semiconductor silicon wafers are highly concentrated in the hands of a few manufacturers such as Shin-Etsu Chemical and Sumco in Japan, GlobalWafers in Taiwan, Sekisui Electronic Materials in Germany, and SK Siltron in South Korea. Technologically, silicon wafer factories need to melt semiconductor-grade high-purity polycrystalline silicon, grow almost defect-free single-crystal silicon rods through the Czochralski method or float-zone method, and then complete doping, slicing, grinding, etching, polishing, and epitaxial growth. The final silicon wafer is only about 0.7-0.8 millimeters thick but must meet extremely high flatness, surface cleanliness, and trace metal contamination standards. Silicon wafers serve as the physical substrate for all transistors and memory units, as well as the "upper limit" of wafer fab yield: any tiny crystal defects, warping, or particle contamination can be magnified in the subsequent hundreds of processes, reducing the number of qualified chips that can be produced on a single wafer. Therefore, without a sufficient and certified supply of high-quality silicon wafers, even if the lithography machines and advanced packaging capacity of wafer fabs are in place, capital expenditure cannot be converted into real chip output. Wedbush: Micron's investment in GlobalWafers locks in silicon wafer capacity lifeline "We believe that Micron's investment in GlobalWafers may indicate that Micron views wafer supply as another potential AI computing hardware bottleneck. Considering that the investment scale in storage and logic chips may continue to expand in 2028, 2029, and 2030, leading semiconductor-grade wafer/silicon wafer demand is likely to increase dramatically, and Micron's management's judgment on this is very reasonable," Matt Bryson, a senior financial market analyst at Wedbush Securities, wrote in a report to clients. Micron stated in a statement that the company plans to invest up to $3 billion to strengthen the US domestic semiconductor supply chain ecosystem, including investing in GlobalWafers in Taiwan. The analyst team led by Bryson further analyzed that the latest commitments made by Micron largely indicate that "storage chip customers expect future AI computing-related demand to be stronger, which may significantly prolong the current storage super cycle, even if additional capacity may eventually bring additional supply surplus risks." He also pointed out that as the US federal government actively collaborates with the chip manufacturing industry to address storage chip shortages, related supply surplus or supply chain shortage risks have been somewhat limited. Driven by the surge in storage demand in the AI era, Micron currently expects its total investments to exceed $250 billion by 2035, higher than the previous estimate of $200 billion. The storage chip giant expects that the increased investment will support its long-term growth target of producing 40% of its dynamic random-access memory (DRAM) products in the US while creating more lucrative direct and indirect job opportunities. In June 2025, Micron's management stated that they planned to invest approximately $200 billion in the US to meet the growing demand for HBM/DRAM/NAND storage chips and eventually produce 40% of the company's dynamic random-access memory product line in the US. Is the storage super cycle spreading to "silicon base shortages"? The grand investment narrative of global capital seeking "silicon-based inflation, weakening carbon-based" assets so far this year fundamentally represents a shift of capital from traditional manufacturing, automotive, consumer, real estate, and energy sectors, which rely on population, resources, and linear economic growth, to the high-end manufacturing chain around silicon wafers related to AI computing infrastructure. This is not just a narrative of chasing tech stocks, but a reassessment of "the most core carrier of future growth": whoever controls the resources related to AI training/inference of AI computing infrastructure will receive higher valuation premiums. The relative allocation weight of funds is shifting from old-economy assets relying on population, oil and gas, real estate, and consumption cycles to infrastructure assets capable of supporting AI training/inference and automated physical AI productivity expansion. The core logic behind global capital embracing "silicon-based inflation" is undoubtedly that in the AI era, the most scarce resources are not traditional labor, real estate, or general types of production and manufacturing capacity, but GPU/ASIC, HBM/DRAM/NAND storage chips, data center CPU components, high-performance Ethernet infrastructure, advanced packaging capacity, EUV, and other cutting-edge semiconductor manufacturing equipment, data center power chains, and data center optical interconnection/optical communication, these "silicon-based production materials." The 300mm semiconductor-grade silicon wafers required for global advanced chip processes, including AI chips and high-end DRAM storage chips, may become the latest structural bottleneck in the AI computing industry chain. Micron plans to provide $500 million in financing to GlobalWafers and sign a ten-year silicon wafer supply agreement, indicating that the core purpose is not just financial investment but pre-locking key inputs for future US DRAM, NAND, and high-bandwidth memory production capacity expansion. This suggests that Micron has already regarded silicon wafer supply security as a long-term production constraint on par with wafer manufacturing equipment, advanced packaging, and electric power. If storage and advanced logic wafer fabs expand production simultaneously by 2028-2030, the increase in chip output must first be converted into higher "wafer inputs." Dynamic random-access memory and flash memory mainly use high-quality 300mm polished silicon wafers, while advanced logic chips heavily rely on epitaxial silicon wafers; advanced nodes not only require more silicon wafers but also demand lower crystal defect densities, higher flatness, more stringent metal contamination control, and more stable resistance values. It is worth noting that such materials require long-term customer certification, and wafer fabs cannot switch suppliers as quickly as replacing ordinary industrial materials, so even if the global supply of silicon wafers is not in short supply, high-spec products that meet the requirements of high-bandwidth memory and advanced logic may still face local supply shortages. Micron's move undoubtedly significantly strengthens the judgment that the AI supply chain bottleneck is shifting from GPUs, HBM, advanced packaging to silicon wafers and other basic materials. Advanced 300mm silicon wafer suppliers may gain longer order visibility, higher capacity utilization, and stronger long-term agreement negotiation capabilities, becoming the beneficiaries upstream in the AI capital expenditure diffusion market. However, the ten-year long-term contract and supplier financing may also be aimed at avoiding future shortfalls; if various silicon wafer fabs expand production simultaneously in the most optimistic demand period, and the growth rate of chip capital expenditure declines around 2029, today's bottleneck may also turn into tomorrow's oversupply.