Hook
When Cathie Wood publicly declared HBM-dependent AI chip stocks a 'red flag' last month, the market yawned. HBM prices had surged 3x to 10x, SK Hynix and Micron were printing record margins, and NVIDIA’s data center revenue was still climbing. The consensus screamed: “Buy the HBM story.” Yet Wood, the narrative hunter of Ark Invest, quietly rotated her portfolio away from the very memory stocks that analysts called the “picks and shovels” of the AI gold rush. Instead, she placed her bets on Cerebras, Groq, and other architectures that banish HBM entirely. Is she insane? Or is she the only one reading the structural tea leaves? The answer lies not in quarterly earnings, but in the uncomfortable intersection of Moore’s law, supply chain physics, and the arc of capital cycles.
Context
Cathie Wood’s Ark Invest has built a reputation as a disruptive technology oracle. Her 2020 electric vehicle bet paid off spectacularly, and her 2021 Terra/Luna skepticism (yes, she was early) added credibility. Now, her attention is on the semiconductor supply chain, specifically the high-bandwidth memory (HBM) stack that powers NVIDIA’s AI accelerators. HBM is a specialized DRAM architecture that uses through-silicon vias (TSV) to stack memory layers vertically, delivering massive bandwidth to the GPU. The market is dominated by SK Hynix, Samsung, and Micron, with SK Hynix leading in HBM3E and NVIDIA as the primary customer. Since 2023, HBM prices have skyrocketed, driven by the insatiable appetite of large language model (LLM) training. The price surge is so dramatic that some HBM packages now cost more than the GPU logic die itself. Wood’s thesis is simple: This price explosion is not a structural shift but a cyclical peak. She argues that the high prices will (a) incentivize capacity expansion, leading to oversupply, and (b) force chip designers to rethink architectures to reduce HBM dependency. Her solution? Invest in Cerebras, Groq, and other startups that use on-chip SRAM or wafer-scale integration to eliminate external HBM. The market, however, remains skeptical. Cerebras and Groq are niche players with tiny revenue compared to SK Hynix. But Wood’s counterintuitive stance forces us to examine the deeper narrative: Is HBM the new DRAM commodity, or is it a ‘toll booth’ that will extract alpha for years?
Core
I’ve spent the last decade analyzing narrative cycles in crypto, where “bottleneck creates opportunity” is a recurring theme. In 2017, I watched Ethereum’s ICO mania create a GPU shortage, only to see ASIC miners and then proof-of-stake alternatives emerge. In 2020, DeFi’s composability led to a liquidity fragmentation that I quantified in a 10,000-word deep dive. Now, the HBM cycle feels eerily familiar. The narrative is that HBM is a structural bottleneck—a toll booth that will capture AI value for years. The data, however, suggests a cyclical commodity dressed in a story. Let’s break down the mechanism.
1. The Price Signal: A Classic Cycle Top?
HBM prices have risen 3x to 10x in 18 months. This is not a typical supply-demand imbalance; it’s a panic. Cloud hyperscalers are double-ordering to secure capacity, and memory makers are racing to convert DRAM fabs to HBM. Historical data from the DRAM industry shows that such price surges are followed by a rapid increase in supply, leading to a 12-18 month correction. The capital expenditure announced by SK Hynix and Samsung is staggering—over $50 billion combined for HBM and related packaging. The lead time for new HBM capacity is 12-24 months, meaning the first wave of new supply will hit the market in late 2025 or early 2026. By then, the current demand growth rate may slow as LLM training shifts from brute-force scaling to fine-tuning, and as inference becomes dominant. The risk is that the market is pricing HBM stocks as if the current price spike is permanent, while the underlying asset is a commodity with a history of brutal cycles.
2. The Architectural Shift: SRAM as the Escape Hatch
Wood’s thesis hinges on the idea that the HBM bottleneck can be bypassed by architectural innovation. Here’s where the technical detail matters. HBM is external DRAM stacked on a GPU via TSV and CoWoS packaging. The latency is ~100 ns, and the power consumption is high. In contrast, on-chip SRAM offers latency of 1-2 ns and drastically lower power per bit. The catch is SRAM density: a state-of-the-art SRAM cell is about 100x larger than a DRAM cell, so you can’t fit the 100 GB+ of memory needed for large model training. But for inference, where you only need to load model weights once and then run many queries, SRAM can be competitive. Cerebras’ wafer-scale engine packs 40 GB of SRAM on a single 5nm wafer, enabling it to run models like GPT-3 without external memory. Groq’s LPU uses a similar philosophy, relying on SRAM and a deterministic architecture for low-latency inference. These architectures are not training replacements—they are inference accelerators. But inference is where the AI market is heading: by 2026, inference is expected to account for 70% of AI compute demand. If that shift happens, the HBM monopoly over AI memory could be broken.
3. The Supply Chain: A Complex Web of Bottlenecks
The current HBM shortage is not just about DRAM. It’s about TSV, CoWoS, and advanced packaging. There are only a few fabs that can do the interposer assembly with high yield. TSMC is expanding CoWoS capacity, but it’s a slow process. The result is that even if HBM dies are available, the packaging step can bottleneck. This is a hidden vulnerability. If the market overestimates how quickly these bottlenecks can be resolved, the current price premium could persist longer than Wood assumes. But her point is that the entire stack—TSV, CoWoS, HBM—is a commodity layer that will eventually be commoditized. The durable moat is not in the memory but in the architecture that integrates compute and memory efficiently.
4. The Geopolitical Distortion
This is where Wood’s thesis may be too clean. Export controls on HBM to China are tightening. The US government, under the CHIPS Act, is incentivizing onshore packaging capacity. If HBM becomes a geopolitical weapon, the supply chain will fragment, and the shortage could be prolonged artificially. This is not a pure market cycle; it’s a policy-driven distortion. Wood, who built her reputation on free-market disruption, may underestimate how government intervention can bend the cycle. On the other hand, the export curbs create a massive incentive for Chinese companies to develop HBM alternatives, which could accelerate the very architectural shift she predicts. Either way, the geopolitical overlay adds a layer of uncertainty that makes the HBM bet less binary than a simple cycle call.
5. Quantitative Comparison: HBM vs. SRAM AI Chips
Let’s look at the numbers. A single HBM3E package from SK Hynix offers 1.6 TB/s bandwidth and 24 GB capacity, costing roughly $300-$500 per package. A Cerebras WSE-3 has 40 GB of SRAM (not directly comparable, but bandwidth is 21 PB/s on-chip), costing an estimated $2 million per system. The total cost of ownership (TCO) for a 10,000-GPU cluster using HBM is dominated by the GPU cost (~$3 billion for H100 units), but memory is a significant fraction. For large-scale inference, the SRAM-based approach can be more power-efficient and simpler to program. The key metric is not memory bandwidth alone but the ratio of memory bandwidth to compute, and the ability to keep the compute units saturated. NVIDIA’s H100 has a memory bandwidth of 3.35 TB/s, but the GPU can only be fully utilized if the kernel is large enough. Cerebras claims that for many inference workloads, the WSE-3 achieves 10x lower latency and 5x lower power than H100. If these claims hold at scale, the narrative switch from “HBM is necessary” to “HBM is a bottleneck” will be rapid.
6. Market Sentiment as a Contrarian Indicator
I’ve seen this before. In 2020, when I tracked the DeFi liquidity fragmentation, the market was obsessed with the idea that “yield farming is the new normal.” It turned out to be a levered cycle. The same is true for HBM. The bullish narrative is that HBM is a “moat” because it’s hard to manufacture. But the history of DRAM shows that moats in memory are temporary. Every time a new memory technology (like 3D XPoint) emerged, the incumbents adapted. The real question is whether the architectural shift to SRAM or near-memory computing will happen fast enough to dethrone HBM before the cycle turns. The market is currently pricing in a bullish scenario where HBM remains the dominant AI memory for the next 5 years. Wood is betting it will be 2-3 years. The difference is a 50% drawdown in HBM stock prices.
Contrarian
Now, the contrarian angle. Wood’s bet is not without blind spots. First, she assumes that the shift to SRAM-based inference will happen smoothly. But Cerebras and Groq still face significant challenges: software ecosystem maturity, high upfront cost (the WSE-3 is a million-dollar system), and the need for custom cooling. The hyperscalers (Google, Amazon, Meta) are already building their own custom AI chips with integrated HBM. They are not likely to switch to a third-party wafer-scale engine unless the performance gap is massive. Second, the geopolitical distortion could work in the opposite direction: if the US imposes export controls on HBM to China, the premium for HBM in the West could actually increase, benefiting SK Hynix and Micron. Third, the SRAM approach is not scalable for training. The largest models (GPT-4, Gemini) require 100s of GB of memory, which cannot fit on a single wafer. So the narrative of “de-HBM” is really an inference-only story. The training market, which is the profit center, will remain HBM-dependent for years. That means the revenue of HBM suppliers is at least partially protected. Finally, the market might be mispricing the cyclical risk: HBM stocks trade at 15-20x forward earnings, which is not a bubble multiple. If the price surge persists for another year, the stocks could go higher before the correction.
Takeaway
The real insight from Wood’s bet is not that HBM is doomed, but that the AI chip market is bifurcating. Training will rely on HBM; inference will increasingly adopt SRAM or processing-in-memory (PIM). The danger is that the market is lumping all AI memory stocks together, ignoring the structural shift in the inference segment. The smart money is not necessarily short SK Hynix vs. long Cerebras. It’s about understanding the timeline: the current HBM cycle will peak in late 2025, and the alternative architectures will gain traction in 2026. The contrarian position is to sell HBM stocks into strength now, but not to short them—because the geopolitical tailwind may keep the party going. Instead, the asymmetric bet is to take a small position in the “de-HBM” plays before they are discovered by the broader market. The future of AI memory is not a single road; it’s a fork. And the fork is coming faster than the consensus thinks. So, the question for the reader is: Are you betting on the toll booth, or on the side road that bypasses it? If you’re late to the first, you’re already late to the second.