Summary
TrendForce reported in October 2025 that OpenAI signed preliminary agreements (LOIs) with Samsung and SK Hynix to supply memory chips for its Stargate data center initiative. OpenAI’s projected demand is approximately 900,000 DRAM wafer starts per month — up to 40% of total global DRAM output. Meeting this demand would require both suppliers to restructure their production mix, necessarily tightening supply of mainstream DRAM for consumer and enterprise markets.
Key Points
- OpenAI’s projected monthly DRAM requirement under Stargate: ~900,000 wafer starts.
- 900K wafers/month represents up to 40% of total global DRAM output.
- This figure “would more than double the current global HBM production capacity.”
- SK Hynix’s current HBM capacity is ~160,000 wafers/month — the Stargate demand would vastly exceed this alone.
- Samsung’s role: supply graphics DRAM (GDDR), high-capacity SSDs, and LPDDR5X-PIM in addition to HBM.
- SK Hynix fab expansion: Cheongju M15X fab completion end of 2025; Yongin Semiconductor Cluster Phase 1 targeted May 2027.
- Supply chain implication: meeting Stargate demand requires adjusting the production mix for both HBM and commodity DRAM, tightening mainstream supply across the board.
- Agreements described as preliminary (LOIs) — not binding contracts.
Newsletter Angles
- The 40% figure is the load-bearing number in “The $71 Billion Bluff” thesis: a single buyer reserving 40% of global output is not a market — it’s a chokepoint capture. If the LOIs don’t convert to binding contracts, the price signal they sent to the whole DRAM market is still real and still irreversible.
- The gap between SK Hynix’s current HBM capacity (~160K wafers/month) and OpenAI’s stated demand (~900K wafers/month) illustrates that this demand cannot be met without years of new fab construction — raising the question of whether the LOIs were ever physically achievable or were always aspirational signals.
- Samsung and SK Hynix restructuring their production mix to chase Stargate demand creates a systemic risk: if Stargate contracts fall through or are reduced, both companies have already committed capacity, leaving mainstream DRAM supply disrupted with no corresponding demand.
Entities Mentioned
- OpenAI — the buyer; signed preliminary LOIs for 900K wafers/month of DRAM
- Samsung — one of two primary suppliers; agreed to provide GDDR, SSDs, LPDDR5X-PIM, and HBM
- SK Hynix — one of two primary suppliers; current HBM capacity ~160K wafers/month; expanding via Cheongju M15X and Yongin Semiconductor Cluster
- Stargate — OpenAI’s data center initiative driving the DRAM demand projections
- TrendForce — the market research firm authoring this analysis
Concepts Mentioned
- AI DRAM Crisis — this source is the primary quantitative basis for the crisis: 900K wafers/month demand vs. ~400K wafers/month global HBM capacity
- Chokepoint Control — a single buyer potentially controlling 40% of global DRAM output represents infrastructure chokepoint dynamics applied to semiconductor supply
- LOI vs. Binding Contract — the preliminary nature of these agreements is central to evaluating whether the demand signal is real or performative
Quotes
“OpenAI’s projected demand of ~900,000 DRAM wafer starts monthly could represent up to 40% of global DRAM output.” “This would more than double the current global HBM production capacity.”
Notes
TrendForce is a credible Taiwanese semiconductor market research firm with strong sourcing in the DRAM supply chain. However, this analysis is based on LOI terms rather than binding contracts — a material distinction for evaluating the demand signal. The 900K wafer figure should be read as OpenAI’s stated intention at the time of signing, not confirmed production orders. Cross-reference with OpenAI Orders 71B in Korean Memory Chips — Light Reading and TrendForce DRAM Market Share Q3 2025 for the broader supply picture.