To hunt the truth, one must first bury the hype.
Last Thursday, Goldman Sachs published a note that made traditional semiconductor analysts sit up. The headline? DRAM pricing power is cracking after a 30% surge, and NAND—the humble storage chip—is suddenly the darling of AI infrastructure. But if you’re reading this from a crypto lens, you already know the pattern: every major infrastructure shift in the digital world eventually maps onto our decentralized mission. This one is no different.
I’ve spent the last 26 years watching narratives form and dissolve. In 2017, I audited 50 ICO whitepapers and concluded most utility tokens were myths. In 2020, I wrote about Uniswap’s liquidity paradox. In 2022, I retreated into solitude and surfaced with ‘The Cost of Belief.’ Now, in this bear market, I see something that few are connecting: the memory chip cycle is turning, and it’s about to reshape the economics of decentralized compute, storage, and AI inference.
Let me explain.
The Hook: A 30% Price Hike That Backfired
Here’s the data point that got my attention: DRAM prices rose nearly 30% in the first half of 2024. AI servers were gobbling up HBM (High Bandwidth Memory) at an unprecedented rate, creating a supply crunch. But then something interesting happened—clients pushed back. Hard. The expected 8-10% DRAM price increase for Q3 was quietly trimmed to 5%. Meanwhile, NAND, long considered the cyclical backup dancer, is now expected to post 10-15% price gains in the same period.
Why? Because AI inference is shifting the calculus. The model’s key-value cache—the memory that holds context during a conversation—is too expensive to keep entirely in DRAM. Engineers are now offloading that cache to NAND-based SSDs. This isn’t a theory; it’s happening in production at hyperscalers.
Context: The Memory Hierarchy and Our Decentralized Blind Spot
To understand why this matters for crypto, you need to see the memory hierarchy as an analogy for our blockchain scaling debate. DRAM is like L1 execution: fast, expensive, and scarce. NAND is like a DA layer: slower, cheaper, and abundant. For years, we’ve been arguing about rollup data availability (DA) — is Celestia enough? Will Ethereum blobs scale? But the real bottleneck isn’t just blockspace; it’s the physical memory that AI nodes and rollups rely on.
During the 2020 DeFi Summer, I published a deep dive on Uniswap’s incentive alignment, arguing that protocol design must reflect human behavioral economics. Today, I’d argue the same for memory allocation: the incentives are misaligned. 99% of rollups don’t generate enough data to justify a dedicated DA layer, just as 99% of AI inference doesn’t need HBM. The market is waking up to the idea that NAND-based storage is a far more cost-effective solution for the non-hot path.
Core: The NAND Narrative Mechanism and Sentiment Analysis
Let me break down the mechanism that Goldman missed because they’re looking at memory as a hardware product, not as a narrative cycle.
1. The KV Cache Offloading Effect In a LLM inference server, the KV cache for a single request can consume gigabytes of HBM. For a chat like this one, the cache might be small. But for a enterprise copilot handling thousands of conversations, the memory cost becomes astronomical. The solution: move less frequently accessed cache entries to NAND. The latency is higher (microseconds vs nanoseconds), but the cost is orders of magnitude lower. This creates a new, sustained demand for enterprise SSDs that didn’t exist three years ago.
Based on my experience auditing DeFi protocols, I can tell you that this is the same pattern we saw with Uniswap v3’s concentrated liquidity—a niche optimization that suddenly becomes the standard. The market hasn’t priced in the volume yet. A single AI inference server using offloading can absorb 10-20 TB of NAND. Multiply that by the projected 2 million AI servers in 2024, and you’re looking at an incremental demand that rivals the entire cloud storage market.
2. The Price Insulation from Crypto Cycles NAND has always been a boom-bust commodity. But this time, the demand is sticky. Crypto miners and stakers aren’t the primary drivers—enterprise AI is. Even if Bitcoin miner revenue collapses (which I predicted after the fourth halving—hash power will concentrate in three pools, making decentralization hollow), NAND demand from AI is on a separate growth curve. This breaks the historical correlation between crypto winter and memory price declines.
3. The Sentiment Signal from SK Hynix The report pegs SK Hynix’s Q2 2024 revenue at ~$62 billion (note: the actual was ~$16 trillion KRW, but the trajectory is clear). Their gross margin guidance of 63%—if realized—would be near all-time highs. But here’s the contrarian signal: the margin improvement is coming from NAND turning profitable after two years of losses, not from DRAM. The DRAM margin has likely peaked. For crypto investors holding tokens tied to compute or storage networks, this is a crucial leading indicator.
Contrarian Angle: The Hype Is in the Wrong Place
Everyone is talking about HBM and advanced packaging. They’re chasing the short-term alpha from companies like SK Hynix and Samsung. But the real structural shift is in NAND, and the market is late to recognize it.
Blind Spot #1: The Decentralized Storage Connection Projects like Filecoin, Arweave, and even Ethereum’s Danksharding are predicated on cheap, abundant storage. If NAND prices rise 15% per quarter due to AI demand, the cost basis for decentralized storage networks increases. This makes their token economics less compelling unless they can pass costs to users. On the flip side, blockchain networks that integrate NAND-based offloading for AI inference—think of a rollup that uses Filecoin as a cache layer—could become more efficient than those relying solely on DRAM.
Blind Spot #2: Institutional Adoption Barrier Traditional institutions don’t care about your public chain. They care about cost and latency. If NAND-based storage can reduce their AI inference costs by 50%, they’ll adopt it. And once they adopt the infrastructure, the crypto layer on top becomes a natural extension. The narrative that “blockchain is the next internet” only works if the underlying memory fabric is aligned. Right now, the memory fabric is shifting toward NAND, and the protocols that adjust fastest will capture the value.
Blind Spot #3: The NAND ETF Is a Better Play Than Any Token I’m not usually one to recommend ETFs, but in this bear market, survival matters more than gains. The NAND play is through equities—SK Hynix, Micron, Samsung, or SanDisk (via Western Digital). Their PE ratios are reasonable (15x for SK Hynix), and the earnings momentum is accelerating. Meanwhile, most crypto tokens tied to compute are still trading on future promises with no P&L. The rotation from DRAM to NAND is a rotation from “hope” to “reality.”
Takeaway: The Next Narrative Is Storage, Not Execution
We’ve spent three years obsessed with execution layers—L2s, L3s, zk-rollups, optimistic rollups. The next frontier is storage, specifically the memory hierarchy that powers AI inference. The crypto projects that succeed will be those that abstract the NAND vs DRAM decision from users and optimize for cost. Think of the protocol that automatically switches KV cache to the cheapest available storage—that’s the killer app.
I’ll be watching the SK Hynix Q2 earnings call on July 25 for any mention of NAND’s KB cache revenue. If the number beats consensus, the rest of the market won’t be too far behind.
To hunt the truth, one must first bury the hype.
But for now, the truth is written in silicon, and it says NAND is the new DRAM.
