←The mapResearch bank
Theses
Transformer lead times gate hyperscaler returns, not start datesPackage Perimeter Arbitrage: Optics Suppliers Capture Value NVIDIA Cannot RetainNAND Flash: Datacenter Crowding-Out Masked by Low Consumer PullCoWoS Binds the Buyer, Not the LandlordLong-context inference: memory constraint compresses cloud margin before GPU-poor plays feel itCo-packaged optics transfers margin to substrate integrators, not optics specialistsSilicon Wafer: Duopoly Capturing Nothing as Downstream Rents CompoundAdvanced Packaging: The Scarcity Premium Flows Upstream, Not DownAmazon: The Capex-Margin InversionMicrosoft: The High-Multiple Cloud Business Is Already SubscaleFoxconn: Margin Squeeze Masked by AI Server Revenue MixSK Hynix: The Packaging Bottleneck Inverts the Margin NarrativeGPU collateral decay gates buildout faster than physical supplyOracle: Monopsony Rent Capture Masked by Consolidated AccountingInnolight: Negative Operating Leverage Hiding in Hyperscaler ConcentrationInterconnection queue converts hyperscaler capex into generator rentEaton: Margin expansion telegraphs demand scarcity more than growth can showElectrical Steel Asymmetry: GE Vernova Captures Transformer Scarcity Without Steel ExposureConstellation: Hyperscaler Monopsony Masks Nuclear Fuel TransmissionABF substrate dual-gate arbitrage: substrate converters capture the spread the market assigns upstreamRack Power Density: The Liquid Cooling Adoption MirageArm: The Royalty Stream That Scales With Everyone's Margin but Its OwnHBM5: The Signal Integrity Bottleneck Transfers Margin to Silicon PHY DesignersShin-Etsu: Margin Expansion Masked by Commodity PVC ConsolidationSMIC: Subsidy Converts to Capacity Under Obscured Margin PressureSemicap service is an installed-base annuity, and the market still prices these names as capex cyclicalsAlphabet: The Capex Ratchet Liquidity Mispriced as Platform OptionalityPalantir: The Margin-Protected Infrastructure PlayAccelink: The Valuation Anomaly in China's Optical Chip IndigenizationTransceiver margin expansion is a mirage: buyer concentration at 1.6T resets pricing power the market prices as durableQualcomm's low capex masks an IP-only future neither consensus nor bears have pricedMurata: The Margin That Doesn't MoveApplied Optoelectronics: Hyperscaler Procurement Arbitrage Driving Negative SpreadsGlobalFoundries: Customer Concentration Masks Structural Insulation from AI Capex WhiplashSilicon capacitors expose Intel's package productivity deficitArista: Hyperscaler capex intensity conceals margin compression riskSilicon cannot emit light — indium phosphide is the chokepoint CPO relocates but cannot removeMoE inference: converted miners capture memory scarcity, hyperscalers leak itTesla: AI Pivot Insulated from Datacenter BottlenecksCustom Silicon Pays a Different MerchantThe ABF Chokepoint Does Not Pay Its Owners — Buy the Bottleneck Fails HereNVIDIA's Rent Compresses Through Software, Not SiliconThe China chokepoint moves from lithography to bonding — and the substitution trade is priced in the wrong categoryThe Upstream Is Single-Sourced and UnpricedOptical attach is set by topology, not by shipments — linear optical TAM models are mis-specifiedHBM moat vs. DRAM commoditization — is SK Hynix's HBM-mix 'miss' actually bullish?Power Caps the AI Buildout — the bottleneck moved from chips to megawattsEMIB Shifts the Packaging Bottleneck to SubstratesASML / China-DUV displacement is overpriced — the sell-off prices near-full substitution off a 5-machine data pointCost Per Token Is Set Outside the ChipChipflation Is a Wafer Allocation, Not a Demand ShockThe Neocloud Rent Is Consumed By The Asset Before It Reaches EquityMemory-demand durability — does algorithmic efficiency cap the memory super-cycle?HBM4 Qualification Concentrates Share, Not Spreads ItNeocloud margin re-rating — do open-source deployment + rising rental rates break the thin-margin cap?The Megawatt Is the Unit of Account
Mechanisms
ABF substrate and build-up film supplyCoWoS advanced-packaging capacityConventional DRAM and NAND supply (HBM crowding-out)GPU residual value as loan collateralGrid interconnection queue positionHeavy-duty gas turbine delivery slotsLarge power transformer lead timesBuyer concentration tighteningHBF consortiumRent converting into capacityRent migrating upstreamRent not being competed awayco-packaged-optics displaces copperco-packaged-optics displaces optical-transceiveremib displaces cowosglass-substrate displaces abf-substratehybrid-bonding displaces euvsilicon-capacitor displaces mlccChip designDatacenter mathEfficiency arrives in steps, not trendsInference shapeMemory economicsPhotonicsTau scalingToken mathTokenomicsCapacity arriving — CoWoS (Chip-on-Wafer-on-Substrate)Capacity arriving — HBM4Capacity arriving — Silicon Wafer

Token math

driver  a named mechanism, not a conclusion

KV cache per token times context times concurrency sets whether a fleet is compute-bound or memory-bound. Compute binds by ~30x today; memory begins to bind at ~1M average context.

The path

nodeeffect
KV cache bytes per token, by attention architecture+0.35
Memory (KV/state) per token+0.35
Long-context inference (KV-cache growth)+0.35
Memory capacity (per accelerator and per node)+0.35
coverage

4 of 4 declared anchors resolve in the model. A driver is a named mechanism rather than a graded severity or a priced claim, so its effects are deliberately smaller than either.

Arguments about this mechanism

These name its subject — they are claims about this mechanism, and they are what the eligibility gates grade.

Arguments that run through it

These name a node on the path rather than the subject, so they pass through this mechanism without being about it. Shown separately and never graded as claims about it — hubs above the graph’s own 90th-percentile degree are excluded, or every argument touching NVIDIA would attach to every mechanism NVIDIA touches.