The memory cycle has killed profits every five years for decades. 2019 saw a 50%+ collapse in DRAM margins. 2023 brought a near-repeat. The narrative is etched: oversupply follows demand euphoria, prices crash, losses mount. But the numbers now whisper a different story. AI is not the typical end market. And the elasticity of demand for compute might just be the structural break that rewrites the ledger.
Consider the data from the latest Citrini analysis. It models a scenario: by 2028, HBM capacity will flood the market. Traders fear a margin massacre. But the report assigns a price elasticity of demand for AI compute at 1.42. For every 1% drop in effective price, demand rises 1.42%. At a 30% price decline, demand expands 42%. The result: profit contraction of only 15%, not 50%. The cycle, in this telling, is dampened. The question for any data detective is not whether the number is elegant — it is whether the causal chain holds.
The chain of causality is the only truth.
Let me break down the transmission mechanism as I would audit a smart contract. Step one: HBM suppliers cut prices. Step two: NVIDIA, AMD, and other chip designers see lower memory costs. Step three: they pass some of that saving to cloud providers and enterprise customers via lower GPU prices or more capacity per dollar. Step four: developers of AI applications see cheaper inference. Step five: they build more products; usage scales up. Step six: demand for AI chips rises. Step seven: chip designers order more HBM. The elasticity of 1.42 measures the final link: API price to developer usage. The loop closes.
But the ledger reveals three breaks in the chain. First, the price cut at step one does not automatically transmit to step three. NVIDIA's gross margins hover above 70%. Their incentive is to keep margins high, not to deflate the entire stack. If they absorb the lower memory cost as profit, the demand response at step five never happens. The elasticity becomes moot. Second, the cost reduction assumption—15% from process node migration and yield improvement—is an optimistic projection. Based on my audit experience with HBM yield data at Samsung and SK Hynix, the gap between theoretical and realized yield is wide. EUV tool delivery delays have pushed node transitions by quarters. If cost falls only 5%, the profit impact of a 30% price drop is a 40% margin contraction, not 15%. Third, the customer concentration. Three major buyers — NVIDIA, AMD, and a handful of cloud hyperscalers — control over 80% of HBM demand. In such an oligopsony, price cuts are often absorbed by the supplier to win an RFP, not passed through to end users.
Whales don't get fooled by aggregate elasticities. They negotiate individually.
My own work on the MakerDAO stability fee during the 2020 crash taught me to stress-test elasticity assumptions. The fixed fee didn't account for sudden liquidity crunches. Models failed because they assumed users would always respond to price signals. In reality, the largest holders — the whales — didn't adjust their positions until the system was near collapse. Similarly, the largest HBM buyers — the chip makers — don't react to a 10% price change in memory by immediately scaling orders. They respond to their own product cycles, which are sticky. The true elasticity of HBM demand from the supplier's perspective is likely closer to 0.5–0.7 during the first 18 months of a price decline, because GPU design cycles lock in memory specs 12–24 months in advance. By the time demand responds, the price has already moved past the inflection.
Correlation is a whisper; causation is the shout. The correlation between falling memory prices and rising AI compute demand is real. The causation is indirect and delayed.
Now, the contrarian angle. The market is pricing memory stocks with a deep cyclical discount. Forward P/E for Samsung and SK Hynix is around 10x, compared to 6x during pre-AI cycles. That implies investors already expect a partial softening of the cycle. But what if the discount is still too wide? What if the AI demand elasticity is real, and the chain holds because NVIDIA is forced to compete with a surge of alternative AI chips from Amazon, Google, Microsoft, and startups? In that competitive landscape, every chipmaker must reduce cost to win workloads. They will pass through memory savings. The elasticity then becomes a self-fulfilling prophecy. _In the absence of noise, the signal screams._ The signal here is the exponential growth in AI inference. If inference grows 3x per year, price elasticity becomes a mathematical certainty.
The ledger never lies, only the interpreter does. The interpretation hinges on whether the supply side overbuilds uniformly. The three HBM suppliers are not a cartel. They are locked in a race for design wins on NVIDIA's Rubin and AMD's next platform. Each fears losing market share. So each invests heavily, hoping to capture even a 5% higher share. This internal competition creates a prisoner's dilemma: even if total demand grows 42%, individual supply may grow 80% per player. The result is a price decline deeper than the aggregate elasticity would predict. The cycle, in that scenario, survives.
My takeaway after two decades in this industry: never trust a single elasticity number to rewrite a decades-old cycle. But do trust the structural shift in the end market. AI compute is not cyclical — it is secular. The replacement of traditional servers with AI clusters is a decade-long trend. The memory cycle will not die, but its amplitude will shrink. The key signal to watch is not the wholesale HBM price. It is the gross margin of NVIDIA's data center segment. If that margin holds above 65% through 2026, the cycle remains. If it dips below 60%, the transmission chain is working, and memory stocks deserve a growth-stock multiple. Until then, I will keep stress-testing the assumptions. The ledger never lies, but I will make sure I am reading the right entry.
