Not Chasing HBM! Cathie Wood Avoids Memory Stocks Despite Bullish AI Outlook, Citing Tesla's Example to Sidestep SK Hynix and Micron

The AI boom has sharply increased demand for high-bandwidth memory (HBM), making memory manufacturers like SK Hynix and Micron the focus of market attention. However, Cathie Wood, founder of ARK Invest and known as "Wood Sister," is choosing to avoid this popular trade. She argues that tech industry history repeatedly shows that when supply chain bottlenecks drive up component prices, companies do not accept high costs as a permanent norm but instead redesign products through engineering to ultimately reduce reliance on expensive and scarce resources.

Wood recently stated on the Insightful Investor Podcast that Tesla's (TSLA-US) past experience dealing with battery material supply issues serves as a key reference for evaluating the current HBM market. She emphasized that technology companies tend to bypass expensive supply constraints through engineering rather than accept them as long-term realities.

She cited Tesla as an example, noting that electric vehicle batteries were once heavily dependent on cobalt, whose supply and cost issues posed major industry challenges—especially given labor exploitation concerns in the Democratic Republic of Congo's cobalt supply chain. Tesla CEO Elon Musk did not simply accept this supply constraint but significantly reduced cobalt dependency by adjusting battery chemistry.

Wood points out that the current AI industry's challenges with HBM supply tightness and soaring prices may follow a similar path. When HBM becomes excessively expensive, she argues, this very condition will incentivize chip designers and AI companies to rethink system architecture and find ways to reduce memory demand, rather than treating the current supply-demand imbalance as a permanent state.

This is the main reason she is currently avoiding investments in SK Hynix and Micron (MU-US). Wood describes memory as one of the "most cyclical and commoditized" sectors in the semiconductor industry. Therefore, she believes the market's current optimistic expectations for HBM supply-demand imbalances and significant price increases may not be sustainable in the long term.

Wood notes that when product prices rise threefold, fourfold, or even tenfold, this is not a normal state for the technology industry. While the market often views price surges as positive for supplier profits, from a technological evolution perspective, abnormally high prices can become a negative factor, as they provide stronger incentives for competitors and technology developers to seek alternatives or change existing technical architectures.

Wood's investment logic does not negate the AI industry but intentionally distinguishes between "sustained AI demand growth" and "component suppliers currently benefiting most from AI demand." She believes that even if the AI market expands rapidly over the next few years, it does not guarantee that memory manufacturers currently benefiting from HBM supply tightness will permanently maintain pricing power and excess profits.

In Wood's view, today's HBM bottleneck could easily become tomorrow's engineering challenge. She expects AI chipmakers to gradually reduce reliance on external HBM through different processor architectures, memory designs, and system integration methods—especially in the AI inference market.

She specifically highlights Cerebras Systems (CBRS-US) and Groq, noting that these inference-focused architectures offer alternative technological pathways. Cerebras and Groq's inference architectures heavily utilize high-speed on-chip memory, which in some workloads can reduce or even eliminate the need for traditional HBM.

Wood states that the market is already beginning to show trends of reducing AI's dependence on high-bandwidth memory through engineering techniques—particularly in the inference domain. AI inference refers to the computational phase where a trained model generates responses based on user input, such as large language models processing text queries, generating code, or producing other content.

She emphasizes that the inference market's importance deserves special attention. As AI applications shift from model development to large-scale commercialization, industry observers widely expect AI inference to become a computing market far larger than model training. If inference demand grows rapidly while new chip architectures successfully reduce HBM demand per unit of computation, it could significantly alter the market's long-term expectations for memory supply, demand, and pricing.

Therefore, Wood is not dismissing the AI boom but holds a different view on who will ultimately capture the greatest economic benefits from AI capital expenditures. She acknowledges that the AI industry will indeed generate massive computing demand and infrastructure investment but argues that value in the tech industry does not necessarily remain permanently concentrated among the currently most constrained component suppliers.

She points out that abnormally high profits typically attract more capital, prompting companies to expand capacity, ultimately leading to new supply and competitive pressures. Rising HBM prices will similarly serve as a strong incentive for SK Hynix and Samsung Electronics to expand production.

Wood says that when she sees massive capital rapidly flooding into a highly cyclical industry in an extremely short time, her approach is not to chase the popular trade but to focus on how the problems caused by high prices can be solved. In other words, she prefers to seek out companies that can transform cost structures and overcome supply bottlenecks through technological innovation, rather than bet on suppliers currently earning excess profits due to scarcity.

This thinking reflects a clear divergence between Wood and the current market's popular trade. Investors generally expect continued growth in demand for AI servers, accelerators, and large models, which will drive up HBM demand, making SK Hynix and Micron the primary beneficiaries in the AI memory supply chain.

However, Wood argues that one of the most important rules in the tech industry is that "today's bottlenecks often become tomorrow's engineering challenges." Once a technology or component becomes too costly, chip designers and system developers will invest more resources in finding alternatives to reduce dependence on that resource.

Her judgment ultimately hinges on whether AI inference architectures can truly reduce HBM demand. Currently, there is no definitive answer, especially as AI models continue to grow larger and more complex, increasing demand for computing power and memory bandwidth simultaneously.

But if Cerebras, Groq, and other emerging AI chip architectures successfully demonstrate that high-speed on-chip memory, application-specific chips, or other system designs can perform large-scale inference tasks at lower cost, HBM's current market perception as irreplaceable core technology could be challenged.

Thus, Wood's decision to avoid SK Hynix and Micron is not a rejection of AI's long-term growth prospects but a warning about the memory cycle within the AI supply chain. She argues that when prices surge due to scarcity, the market should not only see supplier profit gains but also consider whether the high prices themselves are accelerating the birth of next-generation technologies and whether today's most popular supply bottlenecks will gradually be eliminated by engineering innovation in the future.