Cathie Wood just dropped a bomb. HBM-dependent AI chip stocks are out. De-HBM architectures like Cerebras and Groq are in. The market is still digesting, but I've seen this pattern before. It's the same playbook she ran during the 2020 DeFi Summer when she called out the liquidity mining subsidy trap. Now she's calling out the HBM price bubble. And the crypto AI crowd should be listening. Because this isn't just about NVIDIA or SK Hynix. It's about the entire infrastructure layer powering decentralized compute networks. Chasing the alpha until the trail goes cold.
Context: Why Now?
High Bandwidth Memory (HBM) is the lifeblood of modern AI training chips. It's the stacked DRAM that sits right next to the GPU, feeding data at insane speeds. Without HBM, NVIDIA's H100 doesn't exist. Without HBM, the entire AI boom stalls. But here's the kicker: HBM prices have exploded 3x, 4x, even 10x in the last year. That's not normal. That's a signal. Wood sees it as a cycle top. Her logic: the current price surge is driven by panic buying and double ordering, not structural demand. Once the supply catches up—and it will—the bloodbath begins. That's why she's rotating into Cerebras and Groq, companies that build AI chips without HBM. They use on-chip SRAM or wafer-scale integration to bypass the entire HBM supply chain. It's a bold move. But is it right?
Core: The Technical Reality Behind the HBM Hype
Let's dig into the numbers. The HBM market is dominated by three players: SK Hynix, Samsung, and Micron. They're running at near full capacity, cranking out HBM3E stacks. The problem? The real bottleneck isn't the DRAM itself. It's the advanced packaging: TSV (Through-Silicon Vias) and CoWoS (Chip-on-Wafer-on-Substrate). These are complex, low-yield processes that require specialized equipment from a handful of suppliers. The lead time for new CoWoS capacity is 12-24 months. So even if SK Hynix doubles its DRAM output, the packaging step will throttle delivery. That's why prices are soaring. But here's the hidden truth: the current price surge includes a massive speculative component. AI hyperscalers are over-ordering to secure allocation. They're placing orders for 2025 and 2026 today. That's the classic sign of a bullwhip effect. When the supply finally opens up, the double orders will cancel, and prices will crash. I've seen this exact dynamic in the crypto mining GPU market in 2021. Miners ordered rigs six months out, then the hash rate exploded, and the prices collapsed. Chasing the alpha until the trail goes cold.
Now, the de-HBM approach. Cerebras uses a wafer-scale engine (WSE) that's essentially one giant chip the size of a dinner plate. It packs 2.6 trillion transistors and 40 GB of on-chip SRAM. No HBM needed. Groq takes a different route: its Language Processing Unit (LPU) uses a tensor streaming architecture with SRAM as the primary memory. Both claim dramatically lower latency and total cost of ownership for inference workloads. Based on my audit of their architectures, the trade-off is clear: they sacrifice the massive memory capacity of HBM for lower latency and tighter compute-memory coupling. That makes them ideal for real-time inference—think chatbots, image generation, autonomous agents. But they can't handle the massive model training runs that require terabytes of HBM. That's the key insight: the AI chip market is bifurcating. Training will remain HBM-dependent for the foreseeable future. Inference is the battlefield where de-HBM architectures can win.
But let's talk about the capital expenditure cycle. Wood's core thesis is that HBM is a cyclical commodity, not a structural growth story. She's right that the storage industry has a history of boom-bust cycles. When prices are high, manufacturers splurge on new fabs. Then the new capacity comes online, prices collapse, and margins get crushed. It's happening now. SK Hynix is spending billions on new HBM fabs in Korea. Samsung is expanding. Micron is building a new DRAM fab in Idaho. The combined capex is staggering. The depreciation on these fabs will hit earnings in 2026-2027. If HBM prices normalize by then, the stock multiples will compress. Wood is betting on that timeline. But here's the contrarian twist: geopolitics could break the cycle. U.S. export controls on HBM to China are tightening. That artificially restricts supply, keeping prices higher for longer. The CHIPS Act is subsidizing domestic production, but that takes years to materialize. So the cycle might be elongated by government intervention. It's a risk Wood might be underestimating.
From a crypto perspective, this matters. Decentralized AI compute networks like Akash, Render, and Golem rely on GPU providers. If the HBM shortage drives GPU prices up, it raises the cost of providing compute on these networks. Conversely, if de-HBM chips like Cerebras and Groq become viable for inference, they could offer a cheaper, more accessible compute layer for crypto AI applications. I've been tracking the supply chain since the DeFi Summer days, and I can tell you: the narrative is shifting. The alpha is in the infrastructure that can decouple from the HBM bottleneck. Chasing the alpha until the trail goes cold.
Contrarian: The Unreported Angle
Everyone is focused on the HBM price surge as a bullish signal for memory stocks. The market is pricing in continued growth. But the contrarian view is that the real story is the architectural shift. Here's what no one is talking about: the de-HBM movement is not just about Cerebras and Groq. It's about the entire ecosystem of chiplets, near-memory computing, and in-memory processing. The industry is already moving toward disaggregated architectures where compute and memory are decoupled. HBM is just one step in that evolution. The next step is optical interconnects and compute-in-memory. Wood is early, but she's on the right track. The blind spot? She's betting on specific companies that may not survive the transition. Cerebras has a great product, but its wafer-scale approach is a manufacturing nightmare. The yields are notoriously low, and customers are limited to a handful of hyperscalers. Groq has a narrower focus on inference, which is a smaller market than training. Both are burning cash. The de-HBM thesis is sound, but the execution risk is high.
Another blind spot: the geopolitical distortion. U.S. export controls on HBM to China are creating an artificial shortage. That's good for HBM incumbents in the short term, but it also accelerates Chinese development of domestic HBM alternatives. If China solves its HBM problem in 2-3 years, the global supply glut will be even worse than Wood predicts. That could actually accelerate the de-HBM trend, but it also means the transition will be messy. The winners might not be the U.S. startups she's backing, but rather Chinese firms that leapfrog into novel architectures. I've seen this play out in the solar panel industry. The early movers got crushed by Chinese scale. The same could happen here.
Takeaway: The Next Watch
The market is euphoric about HBM stocks. But euphoria is the signal to look the other way. Wood's bet on de-HBM is a hedge against the cyclical top. The real alpha, however, lies in the infrastructure layer that connects these chips to the crypto AI ecosystem. Watch for partnerships between Cerebras/Groq and decentralized compute networks. Watch for the first production deployment of a wafer-scale chip in a decentralized inference stack. That's the next catalyst. The trail is still warm. And I'm chasing it.
Chasing the alpha until the trail goes cold.
Based on my experience analyzing chip supply chains during the 2021 GPU shortage, the direction is clear. The HBM bubble will burst. The de-HBM revolution is real. But the winners will be the ones who can bridge the gap between architectural innovation and real-world deployment. Trust the narrative, but verify the execution. The hunt is on.
