We didn't see this one coming. Cathie Wood, the oracle of disruptive innovation, just quietly rotated her AI chip portfolio away from the HBM-dependent giants. Her Ark Invest funds are now betting on Cerebras, Groq, and other 'de-HBM' architectures. The market is still pricing in the HBM shortage as a bull run for memory stocks. But Wood? She's reading the tea leaves differently.
Here's the context. HBM — High Bandwidth Memory — is the secret sauce powering NVIDIA's AI dominance. It's the DRAM stack that sits next to the GPU, feeding data at insane speeds. But HBM prices have exploded 3x, 4x, even 10x in the past year. That's not a normal supply-demand curve. That's a warning flare. Wood sees it as a classic commodity cycle peak: high prices trigger massive capex, which eventually leads to oversupply and a crash. She's been around long enough to know that memory stocks are traps at the top.
Root: The HBM supply chain is a brittle masterpiece. TSV (through-silicon via) stacking, CoWoS packaging, and DRAM process nodes at the bleeding edge. It's a three-company oligopoly — SK hynix, Samsung, Micron. And they're all ramping capacity like there's no tomorrow. But the bottleneck isn't the DRAM itself. It's the advanced packaging. TSMC's CoWoS capacity is already spoken for by NVIDIA. Any new entrant trying to scale HBM will face a 12-24 month lead time. That's a long time for a market that could flip from shortage to glut.
This is where Wood's contrarian bet gets interesting. She's not just shorting memory stocks. She's going long on architectures that don't need HBM at all. Cerebras's wafer-scale engine and Groq's LPU (Language Processing Unit) both rely on on-chip SRAM instead of external HBM. No TSV, no CoWoS, no dependency on a three-company supply chain. In a world where HBM prices are squeezing margins, that's a structural advantage.
Cerebras's Demo at the last AI hardware conference was a spectacle. They ran a 70-billion-parameter model entirely on a single wafer, with zero HBM involved. The latency numbers were jaw-dropping. But the crowd was still obsessing over NVIDIA's next GPU. The party doesn't stop for the incumbent until it's too late.
Based on my experience building a real-time transaction indexer during the 2017 ICO frenzy, I've learned that the market's narrative often lags reality by a full cycle. When everyone is singing the same song about HBM being the only path, that's when the disruption sneaks in. Wood's move is a bet on a narrative shift: from 'bigger memory is better' to 'memory dependency is a liability.'
Let's get into the core technical analysis. The HBM camp argues that training large models requires the massive bandwidth that only HBM can provide. That's true today. But the inference market is where the growth is — and that's where de-HBM architectures shine. Inference doesn't need the same memory bandwidth as training. It needs low latency and low total cost of ownership. SRAM-based chips can deliver that at a fraction of the power. Groq claims its LPU can outperform a cluster of NVIDIA H100s on certain inference workloads, using 90% less power. If true, that's a game-changer for AI data centers, and by extension, for crypto AI projects that rely on decentralized compute.
But here's the contrarian angle that most pundits are missing. Wood might be underestimating the geopolitical distortion of the HBM cycle. The US is actively tightening HBM export controls to China. That artificially restricts supply, keeping prices high longer than a pure commodity cycle would predict. The semiconductor industry is no longer a free market — it's a weapon. So HBM prices could stay elevated for another 18-24 months, allowing memory stocks to keep printing money. Meanwhile, Cerebras and Groq still need advanced logic process nodes from TSMC or GlobalFoundries, which are also constrained. The 'de-HBM' camp isn't free from supply chain risk.
s Demo of Groq's LPU at Oracle's cloud showed impressive throughput, but the unit economics are still opaque. How many chips can they actually ship? Their customer concentration is high — a few cloud hyperscalers. That's a fragile base for a disruptive thesis.
Yet, the direction is clear. The AI chip market is about to bifurcate: training stays HBM-dependent, inference goes memory-independent. That's the structural change Wood is betting on. And if she's right, the crypto AI sector — which is largely inference-based — could be the biggest beneficiary. Projects like Bittensor, Render Network, and Akash Network rely on distributed compute for inference. If de-HBM chips become the standard for inference, that creates a new equilibrium between hardware cost and token economics.
Takeaway: The next big crypto AI narrative won't be about which GPU has the most TFLOPS. It'll be about which architecture can run inference without touching a memory bottleneck. Wood is playing the long game on this one. The question is: are you going to wait for the 'HBM crash' to buy in, or will you follow the architecture shift before the crowd does?
From my seat, the signal is clear. The market is still pricing HBM stocks as AI winners. But the real winner might be the architecture that makes HBM obsolete. Fast enough to break things.