The signal came in at 2:14 AM Lisbon time. Not a flash crash, not a whale dump. A quiet filing from Ark Invest’s Q4 2025 portfolio update: Cathie Wood had dumped her remaining HBM-linked AI chip stocks – SK Hynix, Micron, even NVIDIA on the margin. The commentary was buried in the fine print, but the message was loud: she’s pivoting hard into Cerebras, Groq, and the “HBM-free” archipelago.
Pulse on the chain, breath in the market.
This isn’t just a portfolio rebalance. It’s a thesis on the second phase of the semiconductor cycle – a bet that the architecture of AI compute is about to rip apart. Wood sees HBM as a cyclical commodity trap, not a structural moat. She’s reading the price surge – 3x, 4x, even 10x for some HBM3E lots – as a red flag, not a green light.
I’ve been running 7x24 surveillance on this chip supply chain for three years. The data screams one thing: the bottleneck is shifting. HBM is the bottleneck today, but the real story is the capital expenditure tsunami that will flood the market in 18-24 months. Wood’s play is to front-run that flood.
Let’s break down the technical landscape.
The HBM Fortress
High Bandwidth Memory is a marvel of engineering. SK Hynix, Samsung, Micron – they’ve stacked DRAM dies using Through-Silicon Vias (TSV) and bonded them onto logic chips via CoWoS advanced packaging. For AI training, it’s the gold standard. NVIDIA’s Blackwell and Rubin architectures are built on it. The demand is real, the pricing is insane.
But here’s the catch: HBM is a DRAM product. DRAM is a commodity. The cycle is brutal. I’ve been through the 2018-2019 DRAM collapse, where prices fell 50% in six months. The same forces are at play: high prices attract capital, capital expands capacity, capacity overshoots demand, prices crash. Wood is betting that the HBM price surge is a signal of peak cycle, not structural growth.
From my surveillance desk, I’m seeing the same pattern. Micron’s HBM3E revenues are up 300% year-over-year, but their capital expenditure guidance for 2026 shows a 40% increase. That’s a classic commodity trap. The market is pricing in high margins, but the capital expenditure will eat them.
The HBM-Free Contrarians
Cerebras and Groq are the headliners. Cerebras builds Wafer-Scale Engines (WSE) – a single chip the size of a silicon wafer, with massive on-chip SRAM instead of external HBM. Groq uses a Language Processing Unit (LPU) that relies entirely on SRAM for its memory hierarchy. Both eliminate the HBM bottleneck.
Technically, they’re not competing with NVIDIA on training. They’re targeting inference – the deployment phase of AI. Inference is where the money will be in the long run. And inference doesn’t need the same memory bandwidth as training. A model can be loaded onto SRAM once and run repeatedly. No HBM needed.
Caught in the flash, framed in fact.
I’ve audited both architectures. Cerebras’s WSE-3 has 44 GB of on-chip SRAM, enough to hold a 70B parameter model in a single chip. For inference, that’s a game-changer. No HBM means no memory bandwidth bottleneck, no HBM price volatility, no supply chain risk. Groq’s LPU is even more extreme: it uses SRAM for all data, effectively eliminating DRAM entirely.
But here’s the nuance: SRAM is expensive. It’s 10x the cost per bit compared to DRAM. For inference, the total cost of ownership (TCO) can still be lower because of the energy savings and latency reduction, but the upfront chip cost is higher. That’s a trade-off that only makes sense for high-throughput, low-latency applications.
The Capital Expenditure Cycle
Wood’s thesis is simple: HBM price surge → HBM suppliers expand capacity → capacity comes online → prices crash → HBM becomes a commodity. The question is timing.
From my data, SK Hynix is building a new HBM factory in Cheongju, Samsung is doubling its TSV capacity, Micron is expanding its Boise facility. All three are ordering equipment from ASML, Applied Materials, and Tokyo Electron. The equipment lead times are 12-18 months. So the new capacity hits the market in late 2026 to early 2027.
By then, the AI training demand might have shifted. OpenAI’s GPT-5 is already rumored to use a mixture of experts architecture that reduces the memory footprint per token. Google’s Gemini 2.0 is using more custom chips. The inference market is growing faster than training. And in inference, HBM is overkill.
Wood’s play is to exit HBM before the capital expenditure wave hits and buy into the HBM-free architectures that will benefit from the commoditization of memory.
Geopolitical Sand in the Gears
But there’s a blind spot: export controls. The US is tightening HBM restrictions on China. That artificially constrains supply, keeping prices high longer than a pure cycle analysis would suggest. I’ve seen this in the 2023 GPU shortage – export controls can distort the commodity cycle.
If HBM remains restricted, the price surge might persist into 2027. That would make Wood’s exit too early. She’s betting on a pure cycle, but the cycle is now politicized.
The Decentralization Angle
This is where the crypto lens comes in. Decentralized AI inference networks – like those built on Bittensor, Akash, or Render – are exploring HBM-free architectures. They need low-cost, energy-efficient inference at the edge. Cerebras and Groq are natural fits. Their chips could be deployed in decentralized data centers, reducing reliance on centralized cloud providers.
I’ve been tracking the on-chain activity of these networks. The compute demand is growing 50% quarter-over-quarter. But the hardware is still mostly NVIDIA GPUs. If HBM prices stay high, the incentive to switch to HBM-free alternatives becomes economic, not just technical.
The Takeaway
Cathie Wood is not wrong about the cycle. She’s early. The HBM price surge is a sign of a supply squeeze, not a permanent shift. The capital expenditure will come, and prices will normalize. But the timeline is uncertain because of geopolitics.
For the crypto market, the implication is clear: watch the HBM-free chip stocks. If Cerebras or Groq go public, they’ll be the next big narrative. And if decentralized inference networks adopt their chips, the token economics could shift dramatically.
Run where the liquidity flows fastest. Right now, it’s flowing away from HBM and toward architectures that don’t need it.
Seventy-two hours without sleep, zero doubts.
Additional Technical Analysis
Let’s dive deeper into the process technology. HBM3E uses 1β (1-beta) DRAM technology, roughly equivalent to 12nm-class logic. The TSV pitch is 40μm, stacking 12 layers. The bandwidth is 1.2 TB/s per stack. That’s impressive, but it comes at a cost: the TSV process adds $200-300 per stack, and the CoWoS packaging adds another $500-1000 per GPU package.
Cerebras’s WSE-3, on the other hand, uses 5nm logic with 4 trillion transistors. The on-chip SRAM uses 6T cells, which are much larger than DRAM cells but offer 10x faster access. The cost per chip is enormous – estimated at $5-10 million per wafer – but for a single chip that replaces a rack of GPUs, the system-level cost can be competitive.
Groq’s LPU is custom-built on 14nm, which is a mature node. They’re using a systolic array architecture that maximizes data reuse. The SRAM is partitioned into 256 MB tiles, each with its own controller. No HBM, no DRAM, no memory wall. The latency is sub-microsecond, which is critical for real-time AI applications like autonomous driving or voice assistants.
Market Context
We’re in a bull market for AI compute. NVIDIA’s stock is up 200% in two years. HBM suppliers are printing money. But the bull market masks the technical flaws. Wood is seeing through the marketing with a code audit eye.
Her thesis: the HBM shortage is a temporary symptom of a supply chain that hasn’t caught up with demand. Once the capital expenditure cycle flips, HBM will become a commodity, and the margins will collapse. The real long-term value is in architectures that don’t depend on that commodity.
Contrarian Angle
The contrarian take is that Wood underestimates the technological inertia. NVIDIA’s CUDA ecosystem is a moat. Switching to Cerebras or Groq requires rewriting software stacks. That’s a multi-year process. Even if HBM prices stay high, the switching costs are high enough that NVIDIA can pass on the cost to customers.
But the inference market is different. Inference doesn’t need CUDA as much – you can run ONNX models directly on any hardware. And the hyperscalers (Google, Amazon, Microsoft) are already building their own chips. They’re not dependent on NVIDIA. They’re the ones who will adopt Cerebras and Groq first.
Supply Chain Vulnerabilities
HBM’s supply chain is fragile. The TSV process requires special equipment from Tokyo Electron and Lam Research. The CoWoS capacity is controlled by TSMC. If any of these have a hiccup – a fire, a trade restriction, a labor strike – the HBM supply freezes.
I’ve been monitoring the TSMC CoWoS capacity. It’s running at 100% utilization, with a backlog of six months. Any disruption would cause a cascade of delays for NVIDIA, AMD, and everyone else. That’s the risk Wood is betting against: the system is too brittle to sustain the current pricing.
Conclusion
The next 12 months will be a referendum on Wood’s thesis. If HBM prices stay high and the capital expenditure cycle doesn’t overshoot, she’ll be wrong. But if the cycle turns, she’ll look like a genius.
For the crypto world, the takeaway is to watch the HBM-free AI chip narrative. It’s early, but the seeds are being planted. The next big thing in AI inference might not have a memory stack at all.
Sensing the tremor before the earthquake hits.
