Research bank

generated 2026-08-07 · 40 theses · 3 deals · 2 priors · regenerate with python3 tools/research_bank.py

How the numbers are made →
The arithmetic from a published model config to physical demand to company revenue exposure, with the grade on every input. A model's architecture is public the day it ships; its procurement footprint is not public for quarters.

Working material for writing notes: each thesis carries its variant vs consensus, the reasoning chain with validity, a relative-performance chart, the series as CSV, and a pre-filled markdown skeleton. Everything is generated FROM the ontology — if a number here is wrong, fix the entity, not this page.

Theses

thesis	conviction	status	horizon	names	write-up
Power Caps the AI Buildout — the bottleneck moved from chips to megawatts power-caps-ai-buildout	high	playing-out	2026-2030 buildout window	GEV, VST, CEG, TLN, VRT, ETN	.md · .csv
The ABF Chokepoint Does Not Pay Its Owners — Buy the Bottleneck Fails Here abf-rent-does-not-accrue	medium	open	2027	2802.T, 4062.T, 3037.TW, NVDA, AMD	.md · .csv
ASML / China-DUV displacement is overpriced — the sell-off prices near-full substitution off a 5-machine data point asml-china-duv-overreaction	medium	playing-out	2026-2028 (near-term backlog is clear; the ramp resolves over years)	ASML	.md · .csv
The China chokepoint moves from lithography to bonding — and the substitution trade is priced in the wrong category bonding-not-litho	medium	open	2026-2028 (DF2000 Q4-2026 is the first dated test)	BESI.AS, 0522.HK, AMAT, 6146.T, 7729.T	.md · .csv
Cost Per Token Is Set Outside the Chip cost-per-token-is-set-outside-the-chip	medium	open	2027-2028	NVDA, TSM, 000660.KS, 2383.TW, COHR, CRWV	.md · .csv
Custom Silicon Pays a Different Merchant custom-silicon-pays-a-different-merchant	medium	open	2028	AVGO, MRVL, NVDA, GOOGL, AMZN, META, MSFT	.md · .csv
EMIB Shifts the Packaging Bottleneck to Substrates emib-substrate-shift	medium	playing-out	2027	TSM, 011070.KS, ATS.VI, 4062.T, 3037.TW, INTC, 000660.KS	.md · .csv
HBM moat vs. DRAM commoditization — is SK Hynix's HBM-mix 'miss' actually bullish? hbm-moat-vs-dram-commoditization	medium	playing-out	2026-2028 (China DRAM-share and HBM-qualification are multi-quarter measurements)	000660.KS, 688825.SS	.md · .csv
HBM4 Qualification Concentrates Share, Not Spreads It hbm4-share-concentration	medium	open	2027	000660.KS, 005930.KS, MU, NVDA	.md · .csv
Silicon cannot emit light — indium phosphide is the chokepoint CPO relocates but cannot remove inp-is-the-laser-chokepoint	medium	open	2026-2028 (substrate balance called for 2H28)	LITE, COHR, 2455.TW	.md · .csv
The Neocloud Rent Is Consumed By The Asset Before It Reaches Equity neocloud-rent-is-consumed-by-the-asset	medium	open	2027	CRWV, NBIS, NVDA	.md · .csv
NVIDIA's Rent Compresses Through Software, Not Silicon nvidia-rent-runs-through-cuda	medium	open	2027	NVDA, AVGO, MRVL	.md · .csv
Optical attach is set by topology, not by shipments — linear optical TAM models are mis-specified optical-attach-is-topological	medium	open	2026-2028	COHR, LITE, CRDO, ALAB, AVGO, MRVL, FN	.md · .csv
The Megawatt Is the Unit of Account the-megawatt-is-the-unit-of-account	medium	open	2027-2029	NVDA, CRWV, GEV, VRT	.md · .csv
The Upstream Is Single-Sourced and Unpriced the-upstream-is-single-sourced-and-unpriced	medium	open	2027-2028	2802.T, 6146.T, BESI.AS, 2383.TW, TSM	.md · .csv
Chipflation Is a Wafer Allocation, Not a Demand Shock chipflation-is-a-wafer-allocation	low	open	2027	005930.KS, 000660.KS, MU, 688825.SS, AAPL	.md · .csv
Memory-demand durability — does algorithmic efficiency cap the memory super-cycle? memory-demand-durability	low	open	2026-2028 (adoption is multi-quarter)	000660.KS, MU, 005930.KS	.md · .csv
Neocloud margin re-rating — do open-source deployment + rising rental rates break the thin-margin cap? neocloud-oss-margin-expansion	low	open	2026-2028 (margin re-rating is multi-quarter; depends on rental-rate persistence)	CRWV, NBIS	.md · .csv
Semicap service is an installed-base annuity, and the market still prices these names as capex cyclicals semicap-service-annuity	low	open	2027-2029 (the test requires a downturn)	ASML, LRCX, KLAC, AMAT, 8035.T	.md · .csv
Accelink: The Valuation Anomaly in China's Optical Chip Indigenization gen-accelink	none	open	—	002281.SZ	.md · .csv
Advanced Packaging: Equipment Suppliers Capture the Scarcity Premium gen-advanced-packaging-bottleneck	none	open	—	—	.md · .csv
Alphabet: The Margin Inversion Nobody Prices gen-alphabet	none	open	—	GOOGL	.md · .csv
Amazon: The Capex-Margin Inversion gen-amazon	none	open	—	AMZN	.md · .csv
Applied Optoelectronics: Hyperscaler Procurement Arbitrage Driving Negative Spreads gen-applied-optoelectronics	none	open	—	AAOI	.md · .csv
Arista: Customer concentration is diluting, not concentrating, as AI scales gen-arista	none	open	—	ANET	.md · .csv
Co-packaged optics transfers margin to substrate integrators, not optics specialists gen-co-packaged-optics	none	open	—	—	.md · .csv
Constellation: Hyperscaler Monopsony Masks Nuclear Fuel Transmission gen-constellation-energy	none	open	—	CEG	.md · .csv
CoWoS Constraint Prices Into the Wrong Companies gen-cowos-capacity	none	open	—	—	.md · .csv
Eaton: Margin expansion telegraphs demand scarcity more than growth can show gen-eaton	none	open	—	ETN	.md · .csv
Electrical Steel Asymmetry: GE Vernova Captures Transformer Scarcity Without Steel Exposure gen-electrical-steel-supply	none	open	—	—	.md · .csv
HBM5: The Signal Integrity Tax Inverts the Memory Supplier Power Balance gen-hbm5	none	open	—	—	.md · .csv
Innolight: Negative Operating Leverage Hiding in Hyperscaler Concentration gen-innolight	none	open	—	300308.SZ	.md · .csv
Long-context inference: memory constraint compresses cloud margin before GPU-poor plays feel it gen-long-context-inference	none	open	—	—	.md · .csv
Microsoft: Burning Cash to Rent Margin It Cannot Own gen-microsoft	none	open	—	MSFT	.md · .csv
MoE inference: converted miners capture memory scarcity, hyperscalers leak it gen-moe-inference	none	open	—	—	.md · .csv
Transceiver margin expansion is a mirage: buyer concentration at 1.6T resets pricing power the market prices as durable gen-optical-transceiver	none	open	—	—	.md · .csv
Package Perimeter Arbitrage: Optics Suppliers Capture Value NVIDIA Cannot Retain gen-package-shoreline	none	open	—	—	.md · .csv
Palantir: The Margin-Protected Infrastructure Play gen-palantir	none	open	—	PLTR	.md · .csv
Rack Power Density: The Liquid Cooling Adoption Mirage gen-rack-power-density	none	open	—	—	.md · .csv
SK Hynix: The Customer Concentration That Validates the Moat gen-sk-hynix	none	open	—	000660.KS	.md · .csv

Deals — bindingness

deal	archetype	consideration	score	read
AMD–OpenAI 6GW deployment + 160m-share warrant amd-openai-warrant	warrants-tied-to-deployment	$77.9bn	81/81	HIGH — a real dependency; carry weight on the edge
NVIDIA–OpenAI letter of intent — up to $100bn against 10GW nvidia-openai-loi	circular-vendor-financing	$100.0bn	27/81	MED — partially binding; watch for terms
NVIDIA $5bn stake in Intel + custom x86 co-development nvidia-intel-stake	competitor-lifeline-plus-product-commitment	$5.0bn	3/81	LOW — option premium; do NOT move dependency weights

prior: contingency × offtake × materiality × instrument · python3 tools/deal_metrics.py scan

Generative priors

Contingency × Offtake × Materiality — the deal-bindingness generator conviction medium

A GENERATIVE mechanism, not a deal note: an announced partnership actually binds the parties when the money is CONTINGENT on performance, a VOLUME commitment is attached, and the check is MATERIAL to the payer — and it is MULTIPLICATIVE, so a zero on any factor collapses the bindingness regardless of the headline. High score predicts a real dependency the graph should carry weight for; low score p…

Incentive × Capacity — the indigenization / margin-compression generator conviction medium

A GENERATIVE mechanism, not a thesis: an actor breaks a bottleneck or enters a fat-margin business when incentive AND capacity are both high — and it is MULTIPLICATIVE, so both are required. High I×C predicts entry → the bottleneck widens → the incumbent's rent compresses. Low on either factor → the moat persists even under enormous pressure. Analysts called China domestic DUV from this before it …