AI Demand Elasticity Might Break the Memory Cycle – But The Real Alpha Is in the Timeline

The alpha isn't in the spot price of HBM3E. It's in the price elasticity of AI demand. That's the core takeaway from a recent Citrini report that's been quietly doing the rounds in the Tallinn analyst circles I run with. The report argues that a 30% drop in memory prices could trigger a 42% surge in AI-related demand, softening the blow of the 2028 supply glut that everyone's bracing for. But let's be real: the chain from a memory chip price cut to a developer paying for API inference is long and full of friction. And as someone who spent the 2017 ICO boom auditing whitepapers at breakneck speed, I've seen how these elasticity narratives can get overly romanticized. The real story isn't the number itself – it's what happens inside the timeline when you zoom in on who's capturing the margin. The context is straightforward: the memory cycle is driven by HBM (High Bandwidth Memory), the critical component in NVIDIA's AI GPUs. Since 2023, HBM demand has been insatiable, pushing Samsung, SK Hynix, and Micron into a capex race. By 2028, all that new capacity is expected to land, threatening a classic oversupply crash. Historically, memory downturns have wiped out 50-60% of profits. Citrini's contrarian take? This time is different – because AI demand is more price elastic than traditional server or PC demand. Specifically, they estimate an elasticity coefficient of 1.42: for every 1% drop in HBM price, AI demand rises 1.42%. That implies a 30% price fall would spur 42% more volume, limiting the profit decline to around -15% rather than the -50% markets are pricing in. Sounds neat. But the alpha isn't in the headline number – it's in the transmission mechanism. During my DeFi Summer meetups in Tallinn, I learned that the most interesting data is often the unspoken granularity of who actually eats the price change. The 1.42 elasticity is derived from API call behavior – how much more developers will up their inference usage when OpenAI or Anthropic cuts prices. That's a different customer than Samsung's procurement team negotiating with NVIDIA. When HBM prices fall, the benefit doesn't pass straight through to API users. NVIDIA pockets most of the margin as incremental profit before it ever touches a developer's wallet. So the effective elasticity seen by memory makers is much lower – maybe 0.7 or 0.8 – because the upstream player absorbs the price change as fat margin, not as stimulus to volume. I've been tracking on-chain AI token activity for the past 18 months as a proxy for real compute demand. The correlation between memory spot prices and AI token staking yields is surprisingly weak. Render Network's utilization rate, for instance, has stayed flat around 60% even as HBM prices tripled in 2024. That suggests developers are not price-bound – they're bottlenecked by chip availability and cooldown times from NVIDIA's allocation logic. Cutting HBM price won't suddenly unlock a wave of new GPU rigs if Blackwell chips are still on backorder. So the demand elasticity that matters isn't price elasticity; it's capacity elasticity. And capacity is controlled by TSMC's CoWoS packaging, not by memory pricing. Let me double-click on the real constraint: CoWoS. In 2024, TSMC more than doubled its CoWoS capacity to 30k wafers per month, yet NVIDIA still sold out every GPU for the next three quarters. Why? Because the bottleneck moved from the logic die to the memory stack itself – HBM takes up a huge portion of the CoWoS interposer area. Every Blackwell GPU requires 8 HBM3E stacks, each of which needs to be tested, binned, and bonded. The yield of those HBM stacks is currently around 60-70% for the best suppliers. That means for every 100 HBM stacks produced, only 60-70 make it into a GPU. The scrapped stacks are a hidden cost that doesn't show up in the price elasticity model. If prices fall, suppliers might cut corners on testing to maintain margins, raising the defect rate and actually reducing effective supply. I've seen this play out in crypto mining rigs during the 2021 bull run – as ASIC prices soared, manufacturers lowered quality standards to hit shipment targets, flooding the market with subpar hashers. Now, bring in the geo-political layer – my final hidden signal from the Citrini report. The analysis assumes supply overshoot in 2028 happens as planned. But based on my experience covering the China-export-control saga since 2022, I believe the risk is skewed to the supply side being delayed, not accelerated. Samsung and SK Hynix are both reliant on ASML's EUV lithography machines to scale their advanced DRAM nodes for HBM4. The lead time for a new EUV order is 18-24 months, and ASML is already running at capacity. If the US expands export controls to include memory-specific equipment – a very real possibility given the tightening semiconductor rules – the 2028 supply wave could be delayed by a year or more. That would invert the whole thesis: instead of a glut, we'd see a prolonged shortage, keeping HBM prices elevated and memory stocks richly valued. The counter-intuitive angle? The market might be pricing in too much of a 'crash' right now. The current valuation of Samsung and SK Hynix – around 15x forward P/E – already discounts a 30% profit decline by 2028. If the actual decline is only 15% due to demand elasticity plus supply delays, these stocks could rerate as growth stocks. I've seen this movie before: in the DeFi summer, every protocol was valued as a high-growth tech stock until liquidity mining ended, and when the subsidies stopped, the real users vanished. Here, the 'subsidy' is the enormous AI infrastructure spending by hyperscalers. As long as Microsoft, Google, and Amazon keep buying GPUs, memory demand stays sticky – but that spending is itself tied to the ROI of AI applications. If the elastic API pricing does stimulate a wave of new consumer AI apps, the flywheel works. If not, we get the old memory cycle with a new coat of paint. So what should we watch? The signal is not in HBM spot prices or CPI reports. It's in the timeline – specifically, NVIDIA's next-generation Rubin platform announcement, expected in 2026. Rubin is rumored to require 12-16 HBM4 stacks per GPU, a 50-100% increase in memory content per chip. That alone could absorb the capacity expansion that everyone fears. But if Rubin is delayed or downscaled, the 2028 oversupply scenario becomes real. That's the alpha: the timeline of product launches, not the elasticity coefficient. The market is currently pricing a 2028 problem – but the decisions that determine it are being made in 2025 and 2026. Those who track NVIDIA's tapeout schedules, TSMC's CoWoS expansion, and ASML's order backlog will have the edge over those staring at a linear demand curve. My takeaway from this report, seasoned with 38 years and a master's in blockchain engineering? Don't let a single elasticity number distract you from the systemic friction. The AI memory cycle will be different from the past – not because demand is elastic, but because supply is now entangled with geopolitics, advanced packaging bottlenecks, and a single dominant customer (NVIDIA) that can suppress or amplify margin at will. The real question is whether 2028's profit dip will be 15% or 50%. Given the structural capacity constraints and the stubbornness of my own experience watching ASML lead times stretch, I'm leaning toward the milder scenario. But I'm not betting on that yet. I'm waiting for the Rubin timeline to drop. That's the next big signal in the timeline. Keep your eyes on 2026 H2.