⚠️ Breaking: Cathie Wood just dumped her SK Hynix position. The reason? A 10x price surge in HBM memory that she calls a 'cycle top' signal. But beneath the surface, Wood is betting on a radical architecture shift—one that could reshape how AI inference runs on blockchain networks.
Context: Why HBM matters and why Wood is bearish
High Bandwidth Memory (HBM) is the backbone of modern AI training chips. It stacks DRAM dies vertically, using TSV (Through-Silicon Via) and advanced packaging like CoWoS to deliver massive bandwidth to GPUs. SK Hynix, Samsung, and Micron dominate this market. Over the past 18 months, HBM prices have surged 3x-10x, driven by NVIDIA's insatiable demand for H100 and B200 chips.
Cathie Wood, founder of Ark Invest, has a contrarian take: this price explosion is a warning sign, not a growth opportunity. She argues that HBM is a cyclical commodity, not a structural moat. History backs her—memory chips have always boomed and busted. But this time, she's taking it a step further: she's not just avoiding HBM stocks; she's actively betting on companies that eliminate the need for HBM altogether.
Core: The technical case for de-HBM chips
Wood's portfolio now includes Cerebras and Groq—two startups that replace external HBM with on-chip SRAM. Let's break down the architecture.
Cerebras: Wafer-Scale Engine (WSE)
Cerebras builds a single chip the size of a wafer. Instead of routing data to off-chip HBM, it integrates 40 GB of SRAM directly on the die. This eliminates the bandwidth bottleneck of external memory. In my early 2020s work on DeFi arbitrage, I ran into similar latency issues when trying to execute cross-chain trades—every millisecond counted. Cerebras solves that for AI inference by keeping the entire model on-chip. The trade-off? Yield is a nightmare: one defect can kill the entire wafer. Cerebras uses redundant cores to compensate, but that eats into efficiency.
Groq: Language Processing Unit (LPU)
Groq takes a different approach. Its LPU architecture uses a deterministic tensor streaming processor with massive on-chip SRAM (up to 230 MB per chip). No HBM, no DRAM at all. The chip is designed for low-latency inference—think of it as a hardened pipeline for transformer models. During the 2021 BAYC floor crash, I traced whale wallets using on-chain analytics; the speed of data processing was critical. Groq's deterministic execution means predictable latency, which is a godsend for real-time applications like blockchain-based AI oracles.
Why Wood is right, but also wrong
Wood's thesis is that HBM's price spike will accelerate architectural innovation. If memory becomes too expensive, chip designers will find ways to use less of it. That's a reasonable long-term bet. But here's the blind spot: HBM's manufacturing moat is deeper than she thinks. The combination of advanced DRAM nodes, TSV stacking, and CoWoS packaging is a triple barrier. SK Hynix and Micron have spent decades perfecting this. A startup can't replicate that overnight.
Moreover, the geopolitical factor cuts both ways. US export controls on HBM to China could actually prolong the shortage—restricting supply while demand remains high. Wood's cycle model assumes free market dynamics, but government intervention distorts that. I've seen this play out in crypto: when regulators banned KYC-free exchanges, usage didn't drop; it just moved to decentralized platforms. Similarly, HBM restrictions may push Chinese AI firms to develop domestic alternatives, but that won't impact global supply for years.
Contrarian: The anti-HBM thesis is overhyped for now
Speaking at a conference in 2022, I argued that on-chain analytics would replace traditional financial surveillance. That was true eventually, but the adoption curve was slower than I predicted. The same applies to de-HBM chips. Cerebras and Groq are targeting inference, which is a smaller slice of the AI market. Training still demands HBM, and NVIDIA's dominance there is unchallenged. Wood's portfolio is a bet on the future, but the present belongs to HBM.
Another hidden risk: SRAM is expensive. On-chip memory consumes die area and power. For a wafer-scale chip like Cerebras, the cost per chip is astronomical. Groq's LPU requires multiple chips to handle large models, which increases system cost. In contrast, HBM is relatively cheap per gigabyte. The total cost of ownership (TCO) for de-HBM solutions may not beat HBM-based systems for another 2-3 generations.
Takeaway: What to watch next
If HBM supply eases in 2025-2026 as new fabs come online, memory stocks could drop 30-50%. But if export controls tighten, that scenario gets delayed. For blockchain-focused readers, the real opportunity is in decentralized AI inference. Projects like Bittensor or Akash Network could benefit from low-latency, deterministic chips like Groq. I'm tracking the 'memory-less' chip trend—not as a replacement for HBM, but as a complement for specific use cases.
