Cathie Wood Warns High Memory Prices Could Backfire! HBM Demand Being 'Killed' by New Architectures

ARK Invest founder Cathie Wood recently shared her views on the semiconductor and memory markets during a discussion. She believes that as AI chipmakers reduce their dependence on HBM through architectural and engineering innovations, the market's perception of high-priced memory should be reevaluated. A sharp rise in component prices does not necessarily signal positive industry momentum—it could instead drive companies to seek alternative engineering solutions.

Wood noted that in the tech industry, when a component’s price increases threefold, fourfold, or even tenfold, it is not a normal phenomenon but potentially a negative signal. Yet, most market participants still view it as a major bullish indicator.

She pointed out that the industry is already engineering ways to reduce HBM demand, citing AI chipmakers like Cerebras (CBRS-US) and Groq, which integrate memory and computation more tightly to lessen reliance on traditional HBM architectures.

Wood revealed that her venture fund holds shares in Groq and stated that NVIDIA (NVDA-US) has already begun collaborating with the company.

She also referenced Tesla’s (TSLA-US) past example to illustrate how engineering design can reduce dependency on specific materials. When cobalt supply chains were disrupted and prices surged, Tesla CEO Elon Musk reduced cobalt usage in batteries through innovative engineering.

Additionally, Wood observed that from a free cash flow perspective, chipmakers and hyperscale cloud enterprises are currently showing completely divergent trends—though she believes this situation is temporary.

On the 12th of this month, Groq announced it had become a NVIDIA Cloud Partner, offering users training and inference services based on large-scale NVIDIA-accelerated computing clusters.

NVIDIA plans to launch hardware based on Groq later this year and will develop more advanced chips built on this foundation.

On the 17th, Groq completed a $350 million Series A funding round, valuing the company at $3.5 billion. NVIDIA is also planning to participate in this round.

According to Groq’s official website, its AI inference infrastructure is powered by the LPU. The LPU is a processor designed specifically for AI workloads, integrating memory and computation functions on a single chip to improve data storage and retrieval speeds while eliminating timing variability. By using on-chip memory, the LPU achieves faster processing speeds.

Cerebras, on the other hand, employs the Wafer Scale Engine (WSE) architecture, turning an entire silicon wafer into one massive chip, integrating compute cores and memory onto a single die to minimize data transmission distance and latency between components.

Currently, OpenAI’s GPT Codex Spark is already running on Cerebras chips, achieving near-zero-latency response times.