Cathie Wood doesn’t buy the HBM story. Over the past year, Ark Invest has quietly rotated out of positions tied to high-bandwidth memory—the very component that made NVIDIA’s Blackwell possible. The price of HBM has surged 3x, 4x, even 10x in some contracts. But Wood sees a red flag, not a green light. She’s pivoted to Cerebras and Groq, companies that build chips with on-chip SRAM, not external DRAM stacks. The question: is she early, or just wrong?
HBM is the backbone of every major AI accelerator. It stacks dozens of DRAM dies vertically, connected by through-silicon vias, then bonded to the logic chip via CoWoS packaging. Without it, a GPU spends most of its cycles waiting for data. The current market is a three-player oligopoly: SK Hynix, Samsung, and Micron. Together, they’ve driven prices to levels that make even NVIDIA blush. Wood’s thesis is that this is a cyclical commodity peak, not a structural shift. She points to Cerebras’ wafer-scale engine—which packs 40 GB of on-chip SRAM—and Groq’s LPU, which uses 230 MB of SRAM per core, as proof that architecture can bypass the HBM stack entirely.
Let’s dissect the technical claim. HBM’s advantage is density: a single stack can deliver 1 TB/s of bandwidth with 16 GB of capacity. SRAM, by contrast, is fast but expensive—it takes up to 150x more die area than DRAM for the same capacity. Cerebras’ solution is to build a chip the size of a wafer, integrating 850,000 cores and 40 GB of SRAM. Groq, meanwhile, uses a systolic array architecture that keeps data in local registers, minimizing memory access. Both trade raw capacity for latency and energy efficiency. That’s a valid trade for inference, but for training, where models require 100+ GB of weights, HBM remains the only practical option.
The supply chain story is where Wood’s skepticism gains teeth. HBM’s production chain is a house of cards: DRAM wafers from Korea, TSV etching from Tokyo Electron, CoWoS from TSMC, all under export controls and geopolitical tension. Yield is a sedative; volatility is the needle. When HBM prices spike, every link in the chain screams for more capacity. SK Hynix is building a new plant in Cheongju, Samsung is doubling its HBM output, and TSMC is expanding CoWoS capacity. But capital expenditure cycles in memory are brutal: five years of depreciation, two years of profit, and a crash. Wood has seen this play before. In 2018, memory prices collapsed 50% after a similar boom. The difference this time is AI demand, but even that can’t defy the laws of supply and demand.
I’ve watched this cycle before. In 2020, I audited Yearn Finance’s vault strategies and noticed a pattern: when yield curves steepen, people forget that risk is always priced in. HBM’s spot price vs. contract price spread is widening—a classic sign of speculative double ordering. Assets don’t lie—only their shadows do. The shadow price of HBM is the cost of substituting it with SRAM. Today, that substitution cost is astronomical. But if HBM stays expensive for another two years, the incentive to innovate becomes irresistible. Cerebras and Groq are the proof of concept, not the endgame.
Now, the contrarian angle. Wood might be underestimating the stickiness of the HBM supply chain. Geopolitics cuts both ways: if the US further restricts HBM exports to China, the shortage will worsen, not ease. Meanwhile, NVIDIA’s next-generation Rubin architecture will likely use HBM4, which is already in development. The architecture shift Wood is betting on may take a decade, not a year. She’s right that the cycle will turn, but timing is everything. Her bet on non-HBM chips is a long-term option, but the market is pricing it as if the shift is imminent.
Cold hands dissect the heat of a hype cycle. Wood’s move is a bet on architecture diversification, not on HBM’s demise. The fork isn’t in the code—it’s in the memory. For now, HBM remains the king of AI training. But the next generation of inference chips will look more like Cerebras than NVIDIA. The question is when the market will realize that the current price surge is a signal, not a signal of strength, but of a system under stress. We audit the code, but we mourn the users—and in this case, the users are the investors who buy at the top of the cycle.
