In the quiet of the bear, we count the coins. Cathie Wood just sold her SK Hynix stake and doubled down on Cerebras and Groq—two companies that build AI chips without High Bandwidth Memory. The market is pricing HBM as if it's the new oil. But Wood sees a different signal: a classic capital expenditure cycle that ends with oversupply, margin compression, and architectural disruption. For crypto miners, this isn't just a semiconductor story. It's a blueprint for how the next generation of mining hardware—and the tokens that power it—will evolve.
Context: The HBM Gold Rush and Its Discontents
High Bandwidth Memory has become the bottleneck of the AI era. Every NVIDIA H100 or B200 GPU requires multiple HBM3E stacks, and the price has surged 3x to 10x in the past year. SK Hynix, Samsung, and Micron are the three oligopolists, and they are currently enjoying record margins. But Cathie Wood's Ark Invest is systematically rotating out of these names. Why? Because she believes the architectural dependency on HBM is a structural vulnerability, not a moat.
Wood's thesis rests on two pillars. First, HBM is a commodity-like memory product with a strong cyclical history—high prices invite massive capital expenditure, which eventually leads to oversupply and price collapse. Second, alternative chip architectures—Cerebras' wafer-scale engine with on-chip SRAM, and Groq's LPU that replaces DRAM with SRAM—can bypass the HBM supply chain entirely. These designs trade raw memory bandwidth for lower latency, higher energy efficiency, and, crucially, independence from TSMC's CoWoS packaging capacity.
For the crypto space, this is directly relevant. The current mining landscape is dominated by ASICs and GPUs that rely on external memory. But as AI inference workloads become more important for decentralized compute networks (Render Network, Akash, Filecoin), the hardware that powers them will face the same HBM constraints. If Wood is right, the next wave of mining and compute nodes could be built on memory-free architectures, reshaping the entire DePIN value chain.
Core: The Architecture War—HBM vs. On-Chip Memory
Let's break down the technical differences. HBM is a stacked DRAM solution that uses through-silicon vias (TSVs) and 2.5D/3D packaging (CoWoS) to deliver enormous bandwidth to the GPU die. The current generation HBM3E offers up to 1.6 TB/s per stack, but it requires complex manufacturing processes with low yields. The price spike is not just a supply-demand imbalance; it's a reflection of the difficulty in scaling TSV alignment and thermal management across 8 to 12 layers.
Cerebras, on the other hand, builds a single wafer-scale chip—the size of an entire silicon wafer—that integrates 850,000 cores and 40 GB of SRAM directly on the die. No external memory, no HBM, no CoWoS. The key metric is not bandwidth but compute density per watt. Groq's LPU takes a similar approach but uses a tensor streaming architecture that pipelines data through SRAM banks. Both companies achieve inference latencies that are orders of magnitude lower than NVIDIA's H100 when running large language models, because they eliminate the memory bottleneck.
The alpha hides in the variance others ignore. The variance here is the capital expenditure cycle. SK Hynix and Samsung are investing billions into new HBM fabs, with a 12-24 month lead time. Micron is building a new advanced packaging facility in Boise. Meanwhile, TSMC is expanding CoWoS capacity by 60% in 2025. The result: within 18 months, HBM supply will likely catch up with demand, and prices will normalize. The classic semiconductor cycle. Wood is betting that the current euphoria is the peak, and that once the oversupply hits, the memory giants will see their margins collapse.
But there is a counterargument that Wood may be underestimating: the geopolitical dimension. The U.S. has tightened export controls on HBM to China, and the CHIPS Act is subsidizing domestic memory production. These policies artificially constrain supply, keeping prices elevated longer than a pure cycle analysis would suggest. Additionally, the AI training market is still NVIDIA's playground, and training requires massive memory bandwidth that current SRAM-based chips cannot match. Cerebras has a training version (CS-3), but it is not a drop-in replacement for an H100 cluster. The non-HBM camp is winning the inference game, but the training fortress remains unchallenged.
Contrarian: The Decoupling Thesis That Might Not Hold
Wood's core argument is that HBM dependency will decouple from AI value creation. She sees HBM as a cyclical commodity, while AI architectures can evolve to bypass it. But what if the decoupling is one-sided? What if HBM remains essential for training, and the non-HBM architectures only capture a niche of inference? In that case, the memory giants would still hold a monopoly on the highest-value part of the AI stack, and their cyclical downturns would be shallower than in the past.
Furthermore, the crypto mining industry has its own peculiarities. Bitcoin mining ASICs are already memory-light—they rely on simple logic operations. But Ethereum's shift to proof-of-stake killed GPU mining, and the remaining GPU mining is for altcoins like Kaspa, which favor high memory bandwidth. A migration to memory-free architectures could devastate these coins' hashrate dynamics. Conversely, if new AI compute tokens (like those from the Render Network) adopt Cerebras or Groq chips, the network's efficiency could skyrocket, but the pool of available hardware might shrink due to limited production.
We do not predict the storm; we build the hull. For the crypto investor, the takeaway is to watch the hardware supply chain. If HBM prices crash in 2025-2026, the cost of AI inference nodes will drop, benefiting decentralized compute networks. But if non-HBM chips gain traction, they will create a new asset class of specialized mining hardware that is not tied to the memory cycle. The real alpha is in identifying which protocols will be the first to integrate these chips and how their tokenomics will adjust.
In the meantime, Wood's pivot is a signal. She is not just short HBM stocks; she is long architectural innovation. The crypto market should pay attention because the next bull run may be powered by chips that don't need a single memory stack.
