Cathie Wood sold her HBM-dependent AI positions. The market yawned. It shouldn’t have.
On the surface, this is a routine portfolio rotation. Ark Invest swapped NVIDIA for Cerebras and Groq – two companies that build AI chips without high-bandwidth memory (HBM). The move seems contrarian, even reckless. HBM demand is surging. Prices have tripled, quadrupled, in some cases tenfold. Why abandon the hottest commodity in AI?
Because Wood sees what the market ignores: capital expenditure cycles.
HBM is not a structural moat. It is a cyclical bottleneck. The current price surge is a textbook warning signal – not a confirmation of endless growth. I have seen this pattern before. In 2020, while backtesting liquidity mining strategies at Stockholm University, I learned that yield spikes attract capital, but they also attract competition. The same logic applies to HBM manufacturing. High prices incentivize capacity expansion. Capacity expansion leads to oversupply. Oversupply collapses margins. Yields attract capital, but security retains it. HBM’s supply chain is anything but secure.
Context: The HBM dependency trap
HBM is the memory backbone of AI training. It uses TSV (through-silicon via) stacking and advanced DRAM nodes (1β/1γ nm). Packaging requires CoWoS – a 2.5D/3D integration method that is already capacity-constrained. The dependency chain is fragile: SK Hynix, Samsung, and Micron control the DRAM. TSMC controls the packaging. NVIDIA controls the design. A single point of failure anywhere – a tool shipment delay, an export control, a yield hiccup – throttles the entire AI supply.
Cerebras and Groq take a different path. They replace external HBM with on-chip SRAM. Cerebras uses a wafer-scale engine (WSE) – a single giant chip with embedded memory. Groq builds LPUs (Language Processing Units) that rely entirely on SRAM. No HBM. No TSV. No CoWoS dependency. From the lab experiment to the global standard – that is the narrative Wood is buying.
But is it real?
Core: The capital expenditure cycle and the liquidity trap
Here is the analysis the market is missing. HBM prices are surging because demand exceeds supply. That is a cyclical condition, not a permanent one. The semiconductor industry has a well-documented boom-bust rhythm: high prices trigger massive capex, which eventually leads to oversupply and price collapse. We saw it in DRAM in 2018, in NAND in 2023, and we will see it in HBM.
Based on my 2024 ETF macro thesis, I constructed a liquidity model correlating Federal Reserve balance sheet changes with semiconductor capital spending. The conclusion: HBM capex is accelerating into a macroeconomic tightening cycle. Central banks are still absorbing liquidity. Real interest rates remain elevated. The cost of capital for HBM fab expansions is high. If demand softens – even slightly – the oversupply will be brutal.
Wood understands this. She is betting that the HBM price surge is a peak signal, not a structural trend. She is rotating into architectures that are immune to the memory cycle.
But there is a blind spot.
Contrarian: The geopolitical distortion
Wood’s thesis assumes a free market in memory supply. She may be underestimating the role of export controls. The United States is tightening HBM restrictions on China. The Netherlands and Japan are restricting advanced lithography and etch tools. These controls artificially constrain supply, prolonging the price surge. HBM shortages may persist longer than a pure cycle analysis predicts.
I saw this dynamic during my 2025 regulatory stress test for EU MiCA compliance. Regulatory constraints create artificial scarcity. They inflate prices for incumbents while incentivizing alternatives. The same is happening here. Export controls on HBM technology give SK Hynix and Samsung a longer runway. They also give Cerebras and Groq a stronger narrative: “We are geopolitically resilient.”
But the reality is more nuanced. Cerebras and Groq depend on advanced logic foundries (TSMC, GlobalFoundries). Those are also subject to export controls. The supply chain risk does not disappear; it shifts from memory to logic.
The market is ignoring the bifurcation that will define the next cycle.
AI training requires massive memory bandwidth. HBM is the only viable solution today. AI inference, however, is latency-sensitive and cost-sensitive. SRAM-based architectures can win there. The two markets will diverge. HBM suppliers will remain dominant in training, but they will lose share in the faster-growing inference segment.
Takeaway: Position for the divergence
Wood’s bet is not about today. It is about the next 24 months. The HBM cycle is peaking. The architecture shift is real. But the path is not linear. Geopolitics will distort the timeline. The smart money is not rotating out of HBM entirely. It is hedging: long HBM for the training monopoly, long SRAM architectures for the inference future.
From my first liquidity mining experiment in 2020 to the AI-crypto convergence analysis in 2026, I have learned one thing: capital flows where the cycle is misunderstood. Wood is early. The market is late. The signal is in the capex, not the price.