Who eats the loss

The companion to ltas-and-prepayments.md. That page collected what's disclosed. This page walks the chain top-to-bottom and asks where a demand miss actually lands.

The exercise is the memory analogue of the ai-capex "Follow the dollar through Hyperion" walkthrough: trace a single HBM stack from die through accelerator through hyperscaler through end-user, and at each step ask who signed what.

The chain

DRAM die  →  TSV-packaged HBM stack  →  CoWoS-bonded GPU  →  Accelerator OEM (Nvidia/AMD)
                                                                ↓
                                                  Hyperscaler / neocloud / OpenAI
                                                                ↓
                                                  End enterprise / consumer AI buyer

At each interface, an entity signs a commitment. The strength of that commitment determines who is exposed if demand softens.

Interface 1 — DRAM producer → HBM packaging

Who: Samsung, SK Hynix, Micron, internal to themselves (and to TSMC for HBM4 base dies and CoWoS).

Binding paper: None to outside parties. Each producer is making its own capex decision. The decision is to add packaging capacity (SK Hynix Cheongju $13B, West Lafayette $3.87B) more aggressively than fab capacity, because TSV is the binding constraint (see supply/hbm-packaging-bottleneck.md).

Loss exposure if demand misses: Producers retain the depreciation on this capex. But — critically — the capex levels remain disciplined relative to past cycles. SK Hynix's 2024 capex was still below the 2022 peak. Micron's $20B FY26 is the largest in its history, but allocated to HBM/1γ, not to commodity DRAM wafer additions. The producers have specifically structured this cycle so a demand miss doesn't kill them.

Interface 2 — HBM stack → Accelerator OEM

Who: SK Hynix / Micron / (now) Samsung selling HBM3e and HBM4 to Nvidia and AMD.

Binding paper: This is the strongest interface. Mehrotra Q1 FY26: "specific commitments and stronger contractual structures" with locked price and volume for calendar 2026. Six customers. SK Hynix: ~70% of Nvidia HBM4 secured. Samsung qualified only October 10, 2025 for GB300 — the marginal supplier.

Loss exposure if demand misses: This is the first place a demand miss bites. The marginal supplier is cut first — that means Samsung. SK Hynix and Micron, with the bulk of pre-booked volume locked in, are partially protected. But the next miss bites SK Hynix's Nvidia exposure (27% of H1 2025 consolidated revenue from a single customer per the SK Hynix DART filing).

Interface 3 — Accelerator OEM → Buyer

Who: Nvidia and AMD selling GPUs to hyperscalers, neoclouds, sovereigns, OpenAI.

Binding paper from OEM upstream (purchases):

• Nvidia: $50.3B in manufacturing/supply/capacity commitments (substantially all paid through FY27) + $26B in multi-year cloud-service commitments + $3.5B prepaid to suppliers.
• Nvidia: $2.77B already accrued as excess-inventory purchase obligations, plus a $4.5B Q1 FY26 charge for H20 excess inventory and purchase obligations after USG export-license requirements collapsed H20 China demand. The H20 charge is the proof-of-concept: a single regulatory action triggered $4.5B in writedowns on commitments Nvidia could not recover.
• AMD: $12.2B unconditional purchase commitments ($8.5B due fiscal 2026).

Binding paper from OEM downstream (sales): Far weaker. Coverage of hyperscaler-Nvidia purchase agreements suggests reservation deposits and allocation, not binding offtake matching Nvidia's $50.3B upstream commitment.

This is the asymmetry. Nvidia and AMD have signed $62.5B of unconditional purchase obligations to their suppliers. Their customers — Microsoft, Google, Amazon, Meta, OpenAI — have not signed matching unconditional obligations downstream. The OEMs are the load-bearing balance sheet in the chain.

Loss exposure if demand misses: Most. Nvidia's already-booked $3.7B/year inventory writedowns are the leading indicator. AMD's $12.2B is the next layer. If AI demand softens, the OEMs eat the obligations they have signed before any other tier feels it.

Interface 4 — Accelerator buyer → End enterprise / consumer

Who: Hyperscalers selling AI compute to enterprises and consumers; neoclouds (CoreWeave, Lambda, etc.) selling capacity to Microsoft, OpenAI, anchor tenants; OpenAI converting compute into ChatGPT / API revenue.

Binding paper: Effectively none in disclosed form. Enterprise AI contracts are typically annual. ChatGPT subscriptions are month-to-month. Anchor-tenant agreements at neoclouds (CoreWeave's 67% Microsoft concentration, per the ai-capex CoreWeave page) are multi-year — but the disclosed take-or-pay terms are weak relative to the upstream commitments.

Loss exposure if demand misses: The buyers in this tier hold the option to slow purchases. They have built optionality precisely by not signing the binding paper that exists upstream. The asymmetry is the design.

Interface 5 — OpenAI → Samsung / SK Hynix

The Stargate LOI sits outside the main chain because it is a direct end-buyer → producer commitment, bypassing Nvidia.

Binding paper: An LOI, not a binding offtake. 900,000 WSPM = ~40% of global DRAM output if converted. No cash committed. No wafers committed. Just a deferred negotiation.

Loss exposure if Stargate funding falters before conversion: Samsung and SK Hynix have begun signaling capex plans against Stargate (Samsung's P4 1c ramp; SK Hynix's M15X). If the LOI doesn't convert, the capex sits on producer balance sheets. Qimonda 2009 / Elpida 2012 mechanics, dressed in newer clothes.

The ledger

If AI demand misses meaningfully — say, foundation-model revenue grows slower than the obligations Nvidia and AMD have signed assume — the loss lands roughly in this order:

Tier	Who	Why they bleed first
1	Nvidia	$2.77B already accrued + $4.5B Q1 FY26 H20 charge on $50.3B in obligations + $26B cloud commitments
2	AMD	$12.2B unconditional purchase commitments, 3× YoY growth
3	Samsung HBM (marginal supplier)	First cut; qualified at Nvidia only Oct 2025
4	OpenAI Stargate stranded capacity	If LOI doesn't convert, producers eat unrecouped P4 / M15X capex
5	Commodity DRAM buyers (PCs, phones, auto)	Already priced out — DDR5 +95% Q1 2026; demand destruction is the relief mechanism
—	Hyperscalers	Optionality intact; LTAs with prepayments but no equity, no binding offtake
—	Memory producers (commodity DRAM)	Capex-disciplined; protected unless cycle is severe enough to break LTAs

The argument the wiki lands on

The producers learned from Qimonda and Elpida. The hyperscalers learned from Apple's TSMC playbook — buy optionality, not equity. Nvidia and AMD are the two parties in the chain who actually signed binding paper. Nvidia's H20 episode — a $4.5B charge in a single quarter when USG export rules killed demand for one product — is the proof that the contractual paper isn't theoretical. If AI demand misses more broadly, the loss lands first on Nvidia's inventory writedowns and AMD's purchase obligations, then on stranded capacity built against the OpenAI Stargate LOI, then on commodity DRAM users (PCs, phones, auto) priced out of the wafer pool — not on the hyperscalers who built the optionality, and not on the producers who refused to speculate.

This is the asymmetry the wiki exists to surface. It does not predict whether AI demand will miss. It identifies who has signed what — and therefore who is positioned to absorb the loss if it does.