The semiconductor narrative is breaking. While most analysts pile into HBM (High Bandwidth Memory) stocks—SK Hynix, Samsung, Micron—Cathie Wood quietly exits. She’s rotating into Cerebras, Groq, and other architectures that reject the HBM dependency. Most see a cyclical play. I see a liquidity trap disguised as a growth story.
Context: The Plumbing of AI Compute
Let’s deconstruct the HBM stack. HBM is not just DRAM; it’s a complex assembly of TSV (Through-Silicon Via) stacking, advanced packaging (CoWoS), and thermal management. The supply chain is brittle: only three players control HBM3E production, and every NVIDIA GPU depends on their output. Prices have surged 3x to 10x in 18 months. That’s not a signal of structural demand—it’s a signal of bottleneck rents.
Wood’s bet is on architecture that bypasses this bottleneck. Cerebras uses a wafer-scale engine with on-chip SRAM. Groq’s LPU relies on SRAM as the primary memory. Both eliminate the need for external HBM. This is not a niche experiment; it’s a fundamental shift in compute-memory coupling. The question is not whether HBM is superior, but whether its supply chain fragility will force a redesign of AI hardware.
Core: The Macro-Liquidity Lens
I’ve spent 27 years watching cycles. The HBM price surge is a textbook indicator of a commodity cycle top. When prices spike 10x, capital expenditure floods in. SK Hynix and Micron are building new fabs with 12-24 month lead times. TSMC is expanding CoWoS capacity. The result: supply glut, price collapse, and margin compression. This is the same pattern we saw in DRAM cycles in 2018 and 2022. The only difference is that AI hype masks the recurrence.
But here’s the structural twist: HBM’s manufacturing moat is deeper than standard DRAM. TSV + advanced packaging require specialized equipment and decades of process expertise. New entrants can’t just build a fab; they need to master multi-layer stacking and yield management. So the cycle might be longer, but it’s not broken. What Wood sees is that the peak of price power is the peak of narrative. Once supply catches up, the incumbents lose pricing leverage.
I’ve audited enough smart contracts to recognize when a system is over-leveraged. HBM’s current pricing is a debt ponzi: it extracts rent from AI chip makers, who pass it to hyperscalers, who eventually pass it to end users. The moment AI training demand softens (or a new architecture reduces memory needs), the whole tower collapses. That’s why Wood is selling memory stocks and buying architecture disruptors.
Contrarian: The Decoupling Thesis
Most analysts argue that HBM is essential for AI training, and therefore demand is inelastic. They’re wrong. The real inelasticity is in compute, not memory. If a chip can achieve the same training throughput with on-chip SRAM, the HBM premium disappears. Cerebras’ WSE-3 already matches NVIDIA’s H100 in certain benchmarks without HBM. The decoupling has begun.
But there’s a blind spot: Wood may underestimate the geopolitical distortion. HBM is now a national security asset. The U.S. is restricting HBM exports to China, which artificially prolongs the shortage. This means the cycle peak might be delayed by 12-18 months. If you’re shorting HBM stocks, you need to hedge against export control extensions. The plumbing is not just economic; it’s political.
Additionally, the “non-HBM” camp has its own supply constraints. Cerebras depends on TSMC’s advanced logic nodes. Groq needs foundry capacity for its SRAM-heavy design. They are trading one bottleneck for another. The difference is that logic foundry capacity is more diversified than HBM packaging. Still, the risk is real.
Takeaway: How to Position
I’m not betting against HBM entirely. I’m betting that the market is mispricing the architectural shift. The next 18 months will see a bifurcation: training chips will continue to use HBM, but inference chips will increasingly move to SRAM and near-memory compute. The winners will be those who own the proxy for this transition—not the incumbents who profit from the bottleneck.
Watch the plumbing, not the price. The yield on HBM stocks is a mirage. The real yield is in architecture that breaks the dependency. Code is law, but incentives are god. The incentive to escape HBM is now stronger than the incentive to hold it.
⚠️ This article is for deep analysis. Do not trade on emotion; trade on structural shifts. The market will eventually price in the decoupling—but only after the cycle peaks. Position ahead of the curve, not after it.
"Bubbles don't burst when the narrative is strongest. They burst when the last buyer runs out of liquidity." — Chris Lopez