Wiki/Primary documents

OpenAI – Samsung & SK Hynix Stargate LOI

Date: Announced October 1, 2025.

Why it matters: The single most aggressive disclosed memory commitment of the AI cycle — and the one with the weakest binding structure. The LOI targets up to 900,000 DRAM wafer starts per month, roughly 40% of global DRAM output if converted. It is a Letter of Intent, not a binding offtake. No cash committed, no wafers committed, no equity exchanged. If OpenAI funding falters before conversion, the capacity Samsung and SK Hynix build against the LOI is stranded — the Qimonda 2009 / Elpida 2012 failure mode in newer clothes.

Primary sources:

What the LOI says

The publicly released framing:

OpenAI, Samsung Electronics, and SK Hynix are partners in OpenAI's Stargate AI infrastructure program.
Target: up to 900,000 DRAM wafer starts per month from the combined Korean memory producers.
Coverage of Stargate AI-data-center capacity, server memory, and HBM for OpenAI's compute footprint.
No financial terms, no cash commitments, no equity stakes disclosed in the announcement.

For context: global DRAM wafer-start capacity in 2024–2025 was roughly ~2.2 million WSPM. A target of 900,000 WSPM dedicated to a single buyer is ~40% of global DRAM output.

What "LOI" means in contract terms

A Letter of Intent is a non-binding statement of intended terms, typically used to:

Signal commitment to a relationship before binding paperwork.
Allow each party to begin internal capacity planning, financing, and regulatory work.
Preserve optionality to renegotiate or exit if conditions change.

In the Stargate LOI:

OpenAI has not committed to purchase any specific quantity at any specific price.
Samsung and SK Hynix have not committed to allocate any specific capacity at any specific date.
Either party can walk before conversion to a binding agreement.

The stranded-capex problem

Where the LOI creates risk is on the producer side, not the buyer side:

Samsung Pyeongtaek P4 is converting foundry capacity → 1c DRAM (target 200K WSPM by end-2026), heavily oriented toward Stargate-style buyer demand. See supply/fab-pipeline.md.
SK Hynix M15X (Cheongju, pilot May 2026; full 50–60K WSPM mid-2027) is being built against the same booking environment.

If the LOI converts to binding offtake at full scale, the capex pays off. If OpenAI's revenue trajectory (or funding sources) cannot underwrite the buildout and the LOI does not convert, the capex sits on producer balance sheets unrecouped.

This is the precise failure mode that took Qimonda down in January 2009 — a DRAM producer with significant capacity built against demand expectations that did not materialize. See history/prior-cycles.md and documents/qimonda-insolvency.md.

Why this matters for the risk-allocation argument

The Stargate LOI is the cleanest example of the asymmetry the wiki foregrounds:

OpenAI retains optionality. The LOI commits nothing.
Producers carry the capex. Samsung and SK Hynix are spending real money against the booking.
No equity transferred. OpenAI holds no stake in the producers; the producers hold no stake in OpenAI.

This is the structure: buyers build optionality, producers absorb capex, and the loss — if the cycle turns — lands on the producers' balance sheets, not the buyer's.

The producers know this. SK Hynix's CFO Kim Jae-joon has repeated on every call that capex is being managed "based on demand visibility." Samsung has emphasized that P4 is a conversion of existing capacity, not a net add. Both responses are the producers' attempts to keep the Stargate-style risk smaller than it could be.

But the producers cannot completely refuse the build. The Stargate LOI is the largest single signaled demand event in DRAM history. The competitive logic of an oligopoly says each producer has to take some position against it — and if the LOI doesn't convert, the producer that committed least capex wins, the producer that committed most loses.

What's missing from public disclosure