The Civic Node

AI DRAM Crisis

7 min read

Definition

The AI DRAM Crisis is the global memory shortage that emerged in late 2025 when OpenAI signed simultaneous non-binding letters of intent (LOIs) with Samsung and SK Hynix for approximately 900,000 DRAM wafer starts per month — roughly 40% of global supply. Neither supplier knew about the other’s deal. Combined LOI value was estimated by analysts at $71.3 billion over four years. The market treated these non-binding signals as binding demand, triggering a supply panic: competitors rushed to lock in multi-year contracts at peak prices, producers shifted capacity toward DRAM and HBM, and prices spiked 171% year-over-year. The original demand signal was subsequently abandoned — OpenAI cut compute spending by 57%, cancelled Stargate, and shuttered Sora — but the contracts signed during the panic run through 2027–2028, sustaining elevated prices despite collapsed demand. The crisis is not merely a supply-demand imbalance; it is the result of information asymmetry in a three-firm oligopoly responding to a signal that was never meant to be permanent.

Why It Matters for the Newsletter

The AI DRAM Crisis sits at the intersection of every theme this wiki covers. It is simultaneously a story about market power (three companies control 91.5% of global DRAM; a single buyer can move the entire market), monetary policy (prices for a commodity embedded in every consumer device tripled in under a year), infrastructure (AI buildout is now structurally competing with consumer computing for the same physical production capacity), and political economy (the costs of AI infrastructure speculation are being borne by ordinary PC buyers and small businesses, not by the hyperscalers who triggered the panic). The mechanism — non-binding LOI creating binding downstream effects through panic — is a case study in how signals propagate through concentrated markets and why “non-binding” is often a legal fiction when the counterparty is large enough.

Evidence & Examples

The Trigger

Price Impact

Downstream Consumer Damage

Producer Windfall

Demand Signal Collapse

Relief Timeline

  • Structural relief from new fab capacity: mid-2027 at earliest
  • Full market correction: 2028, when CXMT and YMTC (Chinese DRAM entrants) reach scale and introduce meaningful competition
  • Contracts signed at peak prices during the panic run through 2027–2028, sustaining elevated prices even as spot demand collapses

Published Synthesis

  • The $71 Billion Bluff (The Civic Node, April 11, 2026) — the definitive published synthesis of this crisis; coins the “non-binding LOI → binding downstream panic” framing

Tensions & Counterarguments

Intent vs. effect. OpenAI likely did not intend to create a market panic. The LOIs may have reflected genuine demand projections at the time of signing. But in a three-firm oligopoly with no excess capacity, even earnest signals carry manipulative effects — the market cannot distinguish a sincere projection from a strategic feint. The absence of intent does not reduce the damage.

Samsung’s early knowledge. Samsung’s shift of 80,000 wafer starts from HBM back to DDR5 in December 2025 occurred months before OpenAI publicly announced demand cuts. This could indicate that Samsung had early intelligence about OpenAI’s retreating demand — and continued to accept new binding contracts from other customers anyway. If true, this would reframe the crisis as partly a knowing exploitation of information asymmetry rather than a shared mistake.

Jevons Paradox risk. Google’s TurboQuant paper (March 2026) demonstrated 6x memory compression for LLM inference. This is a demand-side force that could accelerate the price correction — but it could also enable more AI deployments at lower per-unit memory cost, potentially sustaining total demand even as per-model memory needs fall. The net effect on DRAM demand is genuinely uncertain.

Non-binding ≠ non-consequential. The legal structure of the LOIs gave OpenAI no binding obligation, but the market’s response was binding. This is a structural feature of concentrated markets: in a three-firm oligopoly with long capital planning cycles, a non-binding LOI from a buyer of OpenAI’s scale is functionally indistinguishable from a binding order — suppliers must act as if it is real or risk being locked out.

  • Chokepoint Control — the three-firm DRAM oligopoly is itself a chokepoint; the crisis illustrates how control of concentrated supply enables both windfall and vulnerability
  • Chokepoint Economics — the mechanism by which market concentration turns information asymmetry into price power
  • Infrastructure Warfare — the AI buildout and the DRAM crisis demonstrate how competition for critical infrastructure components can damage civilian computing as collateral
  • Jevons Paradox — the risk that memory efficiency gains (Google TurboQuant) enable more AI deployments rather than reducing total demand
  • Semiconductor Supply Chain — DRAM is one node in a larger supply chain; the crisis reveals how demand shocks in one segment (AI HBM) propagate to all others (consumer DDR5)

Key Sources

Graph View

Backlinks