Hook: HBM prices are up 3x, 4x, even 10x. Cathie Wood just dumped her HBM stocks. She's betting on Cerebras and Groq—chips that don't use high-bandwidth memory. Is this a visionary pivot or a crash waiting to happen? I've seen this pattern before. In 2020, during the DeFi yield farming sprint, I watched the same herd mentality drive up liquidity mining yields. The crowd chased the hottest returns, and the smart money rotated into infrastructure. Wood's move smells like that same playbook—but this time, the battlefield is silicon, not smart contracts.
Context: The HBM market is dominated by a triopoly: SK Hynix, Samsung, and Micron. Their products are essential for AI training chips like NVIDIA's GPUs. The current price surge is driven by a genuine demand explosion from large language model training. But Wood sees a trap. She argues that the price jump is a cyclical peak, not a structural trend. She's shifted her focus to Cerebras—which uses a wafer-scale engine with massive on-chip SRAM—and Groq, with its LPU architecture that relies on SRAM instead of external memory. These companies are building AI chips that don't need HBM. They are the architectural rebels.
Core: Let's cut through the noise. The real bottleneck isn't HBM itself—it's the advanced packaging ecosystem. HBM requires TSV (Through-Silicon Via) stacking and CoWoS (Chip-on-Wafer-on-Substrate) integration. This is a complex manufacturing process that takes 12-24 months to scale. The current price surge is partly due to these capacity constraints, not just DRAM scarcity. Yields are transient; infrastructure is permanent. The HBM price spike is a signal that the supply chain is stretched, but it's also a catalyst for architectural change. From my experience auditing smart contracts in Mumbai, I learned that the most crowded trades often hide the biggest vulnerabilities. The same applies here: the consensus that HBM is indispensable for all AI workloads is a vulnerability.
Wood's thesis is grounded in the capital expenditure cycle. When prices rise, manufacturers expand capacity. That expansion eventually leads to oversupply and price collapse. HBM is a commodity memory product—its historical margins are cyclical. The current high prices are not sustainable. But the deeper insight is that the AI chip market is bifurcating. Training workloads require massive memory bandwidth, which HBM provides. But inference workloads—especially those demanding low latency and high energy efficiency—can benefit from on-chip memory. Cerebras and Groq are not competing with NVIDIA for training dominance; they are carving out a niche in inference. Speed is a feature, not a bug, until it breaks. HBM's speed is a feature for training, but its cost and supply volatility are bugs for inference.
I've run the numbers on the technical trade-offs. A single Cerebras wafer-scale engine has 2.6 trillion transistors and 40 GB of on-chip SRAM. That's enough for many inference models without touching external memory. Groq's LPU can achieve 80 TOPS/Watt, far better than a typical GPU with HBM. The catch? These chips are limited by logic process yield and power dissipation. They are not a drop-in replacement for NVIDIA's ecosystem. But they don't need to be. The market is big enough for multiple architectures. Art is the metadata of human emotion. The art here is the engineering trade-off: choosing between a universal memory solution (HBM) and a specialized on-chip solution (SRAM). Wood is betting that the market will reward specialization.
Contrarian: Wood's contrarian stance has a blind spot: geopolitics. The US is tightening export controls on HBM to China. This artificially prolongs the shortage by restricting supply to a major market. If HBM remains scarce for longer, the price cycle may not turn as quickly as she expects. I've seen this in DeFi—regulation can distort market mechanics. The SEC's regulation-by-enforcement created artificial scarcity in certain tokens. Similarly, export controls are a man-made variable that defies pure cyclical analysis. Furthermore, Cerebras and Groq face their own scaling challenges. Their wafer-scale and SRAM-based designs depend on TSMC's advanced logic capacity, which is also constrained. They are not immune to supply chain risks.
But Wood's weakness is her strength: she is early. The market is still pricing HBM stocks as if the current earnings are sustainable. They are not. The average depreciation cycle for memory fab equipment is 5-7 years. The new capacity coming online will depress margins exactly when demand growth slows. The protocol is neutral; the user is the variable. Here, the protocol is the HBM supply chain. The user—the AI chip designer—is the variable. When HBM becomes abundant and cheap, designers will still use it. But the architecture shift will happen at the edge: inference chips will increasingly ditch HBM for on-chip memory. Wood is betting on that shift.
Takeaway: The AI chip market is not a monolith. It's a dual economy: training (HBM-dependent) and inference (architecture-independent). Wood's bet is that the next $100 billion in AI value will come from inference, not training. If she's right, the HBM triopoly will become a footnote in the AI story. The question is not whether HBM is overvalued today—it's whether the architecture of the future will need it at all. I don't predict trends; I ride the volatility. And right now, the volatility is screaming that the old infrastructure is being rebuilt.