Hook:
Cathie Wood just drew a line in the sand. Ark Invest is actively avoiding HBM-dependent AI chip stocks, betting instead on Cerebras and Groq. The signal is clear: the price of high-bandwidth memory has surged 3x to 10x in the past 18 months. That's not a sign of strength. It's a yellow flag on a cycle that historically ends in a liquidity drain. For crypto investors, this isn't a sidebar. The same chips that power AI training also hash Bitcoin and validate Ethereum. If Wood's thesis is right, the next hardware bottleneck in crypto won't be ASICs or GPUs—it will be the memory stacked on top of them.
Context:
HBM (High-Bandwidth Memory) is the DRAM solution that sits directly on top of AI accelerators like NVIDIA's H100 and B200. It delivers the bandwidth needed to feed massive parallel compute units. Without HBM, even the most advanced GPU is just a paperweight. The production of HBM requires TSV (Through-Silicon Via) stacking, CoWoS (Chip-on-Wafer-on-Substrate) packaging, and advanced DRAM process nodes. Three companies dominate: SK Hynix, Samsung, and Micron. Their current capacity is fully booked, and lead times are stretching into 2026. Crypto miners, who rely on the same GPU supply chain, are already feeling the pinch. Wait times for new GPU clusters have doubled, and the secondary market for H100s is at a premium.
Wood's alternative thesis is architectural: replace HBM with on-chip SRAM as in Cerebras's wafer-scale engine or Groq's LPU. These designs eliminate the need for external high-bandwidth memory, bypassing the TSV and CoWoS bottleneck. The idea is simple: if the memory is on the chip, you don't need to wait for a packaging line. But the implications for crypto are deeper than just mining rigs. Decentralized compute networks that rely on GPU sharding, like Render Network or Akash, depend on the availability of HBM-equipped GPUs. If HBM becomes a strategic choke point, those networks will see supply constraints and pricing volatility.
Core:
Let's break down the failure modes. My background in auditing DeFi protocols taught me to look for hidden dependencies. HBM is the hidden dependency of the entire AI-Crypto hardware stack. I built a model in 2022 to simulate the feedback loop between TerraUST's algorithmic stability and LUNA's inflation. The death spiral was mathematically predictable. The HBM cycle has a similar structure.
Wave 1: Demand surge. AI training and crypto mining both require massive memory bandwidth. The demand for HBM3E has outstripped supply by a factor of 2.5, based on industry capacity estimates. Wave 2: Price explosion. HBM contracts have been renegotiated at 3x to 10x previous prices. That's not just inflation—it's a signal of scarcity. Wave 3: Capacity expansion. SK Hynix and Samsung are pouring capital into new HBM fabrication lines. The typical DRAM fab takes 18-24 months to come online. Wave 4: Overcorrection. The same cycle has played out in DRAM and NAND for decades. High prices lead to overinvestment, which leads to a glut, which leads to a crash. Math doesn't lie—the HBM cycle is a textbook commodity cycle dressed in AI hype.
But here's the twist for crypto. The crypto mining industry is not just a consumer of GPUs; it's a price taker in a market where HBM supply is allocated first to hyperscalers (AWS, Google, Microsoft) and then to everyone else. When HBM is scarce, GPU manufacturers prioritize data center contracts over consumer cards. That means fewer GPUs available for retail miners and smaller decentralized compute operators. The network effect is toxic: higher HBM prices reduce GPU availability, which increases mining centralization, which weakens the decentralization thesis of proof-of-work and proof-of-stake nodes that rely on GPU compute.
What about the "de-HBM" architectures? Cerebras's wafer-scale engine has 40 GB of on-chip SRAM, enough to run large language models without external memory. Groq's LPU uses a similar SRAM-heavy design. The technical advantage is latency: no need to pass data through TSV and CoWoS. The disadvantage is cost: wafer-scale chips have lower yields and higher defect rates. In my 2020 audit of Aave v1, I modeled how oracle latency could cause cascading liquidations. The same principle applies here: the latency advantage of SRAM is real, but only if the architecture can scale to support the memory requirements of frontier models and crypto workloads. Current SRAM-based chips are optimized for inference, not training. Crypto mining, which is essentially a compute-intensive inference task (hashing), could benefit from inference-optimized chips. But the hash rate of a Cerebras chip per watt is still an open question. I'd want to see a benchmark before calling it a crypto mining disruptor.
— Scenario: When debunking a project's hype, I focus on the gap between the whitepaper and the deployed code. For Cerebras and Groq, the code is the architecture. The question is whether the architecture can be economically deployed at scale. The HBM alternative is not a free lunch—it's a trade-off between supply chain independence and manufacturing complexity.
Contrarian:
Wood's thesis is elegant, but it ignores two critical distortions. First, geopolitics. The US government has tightened export controls on HBM to China, and there are proposals to extend those controls to allied countries. Export restrictions artificially prolong the HBM shortage by segmenting the market. This means the price cycle may not peak as quickly as Wood assumes. The HBM producers are operating in a politically protected oligopoly, not a free market. Second, the "de-HBM" chips are not immune to supply constraints. Cerebras and Groq rely on advanced logic process nodes (5nm/3nm) from TSMC, which is also at capacity. The bottleneck just shifts from memory packaging to logic fabrication. Code is law, until it isn't—the packaging code that binds HBM to GPUs is being rewritten, but the new code runs on a different layer of the stack that is equally constrained.
For crypto, the contrarian angle is that HBM dependence may actually be a stabilizing force. If HBM supply is captive to hyperscalers, then the remaining GPU supply for decentralized networks becomes more predictable—because it's smaller. Smaller supply means the remaining GPUs are held by committed miners, not speculators. That could lead to less volatility in hash rate and more stable network security. I'm not saying it's good, but it's a scenario Wood doesn't address.
Another blind spot: the transition timeline. Wood's fund is known for early exits and missed inflection points. She sold NVIDIA in 2022 before the AI boom. She's now betting against the HBM ecosystem. But the installed base of HBM-equipped GPUs is enormous. Any shift to SRAM-based architectures will take years of capital expenditure, retooling, and software optimization. Crypto miners cannot simply swap GPUs for wafer-scale chips. The mining software is optimized for CUDA and specific memory architectures. Rewriting the firmware for a new chip is a multi-year effort. The liquidity of the used GPU market would collapse if miners tried to offload their HBM-reliant hardware before the new architectures are proven.
Takeaway:
Wood's bet is a macro play on the commoditization of HBM. But crypto is not a macro asset class that moves in lockstep with commodity cycles. It's a system of interconnected dependencies. The failure mode I'm watching is not HBM price collapse—it's the opposite. If HBM shortages persist due to geopolitical lock-in, then the cost of GPU compute for decentralized networks will stay high, making it harder for small players to participate. That is a structural threat to the trustless ideal. The architecture shift to non-HBM chips is the long-term solution, but the transition period is where the risk lives. The question for crypto investors is not whether Wood is right, but whether the timeline aligns with the next halving cycle. If HBM prices stay elevated through 2025, the next bull run will be dominated by those who can afford the hardware. Code is law, until it isn't—and the law of hardware scarcity is the hardest to break.