I used to think the HBM shortage was just a supply chain hiccup. Then I read the code of a Cerebras whitepaper, and I saw the architecture of a future where memory is no longer a leased commodity.
Cathie Wood's Ark Invest recently made headlines for avoiding stocks tied to High Bandwidth Memory (HBM) — the expensive, stacked DRAM that powers NVIDIA's AI chips. Instead, she is betting on Cerebras and Groq, companies that design chips without HBM, using on-chip SRAM or wafer-scale integration. As a crypto education platform founder with a background in economics, I see this as more than a sector rotation. It is a signal about the fragility of centralized hardware dependencies, and a lesson that echoes in blockchain's own quest for trustless computation.
Here is the context you need: HBM is the backbone of modern AI training. It sits next to the GPU, providing the massive bandwidth needed to feed large language models. But HBM is a bottleneck — literally. It requires TSV (Through-Silicon Via) stacking, CoWoS advanced packaging, and a supply chain dominated by three players: SK Hynix, Samsung, and Micron. Prices have surged 3x to 10x in the past year, driven by AI demand. Wood sees this as a cycle top: high prices trigger capacity expansion, which leads to oversupply and collapse. She is betting that architecture innovation will bypass HBM entirely.
Let me take you through the technical core. The 'no-HBM' approach is not about shrinking transistors; it is about coupling memory and compute on the same die. Cerebras uses a wafer-scale engine — a single giant chip with 400,000 AI cores and 40 GB of on-chip SRAM. Groq uses a Language Processing Unit (LPU) with SRAM-based memory. Both eliminate the need for external HBM. This is a radical shift from the NVIDIA paradigm, where the GPU is a tenant renting memory from a separate DRAM landlord. In crypto terms, it is like moving from a centralized cloud service to a self-sovereign node.
From my experience auditing smart contracts in 2017, I learned that hidden points of centralization are the most dangerous. HBM is such a point. If NVIDIA cannot get enough HBM, its GPU shipments stall. But Cerebras and Groq control their own memory supply. This is analogous to the multi-sig problem in DAOs: when governance keys are held by a few, the system is not truly decentralized. Similarly, a chip that depends on a handful of HBM suppliers is not truly resilient.
The contrarian angle, however, is that Wood may be underestimating the stickiness of the HBM ecosystem. Geopolitical factors, such as US export controls on HBM to China, are artificially prolonging the shortage. The US government is tightening HBM restrictions, which actually benefits HBM incumbents by limiting supply. Wood's cyclical view assumes a free market, but the semiconductor industry is now a tool of geopolitics. Meanwhile, non-HBM chips like Cerebras face their own scaling challenges: wafer-scale yields are notoriously low, and system-level costs for cooling and power are high. They are not a drop-in replacement for NVIDIA's training dominance.
Yet, for the crypto world, the implications are profound. AI tokens like Render, Bittensor, and Akash rely on decentralized compute. If HBM remains scarce and expensive, the cost of running AI inference on decentralized networks could skyrocket, favoring architectures that are less memory-dependent. Non-HBM chips could become the go-to hardware for decentralized AI, especially for inference tasks where latency and total cost of ownership matter more than raw training throughput. This is a long-term bet on architectural divergence: training stays with HBM, inference moves to SRAM.
Follow the fear, not the chart. The fear here is not just about HBM prices — it is about the centralization of the entire AI compute stack. If you can understand the technical architecture of your hardware dependencies, you can anticipate where the next bottleneck will form. For blockchain builders, the lesson is to design for modularity and resilience, not just peak performance.
The takeaway is not a buy or sell signal. It is a lens: Wood's bet is a hedge against the commoditization of memory. Whether she is right or wrong, the underlying trend — architecture innovation to escape supply chain traps — is real. In crypto, we call that 'sovereignty.' In chips, it is called 'the end of the HBM era.'