←The mapResearch bank
Theses
Solid-state transformer: unrelieved constraint carrying semiconductor revenue as rent, not as scalePackage Perimeter as Hyperscaler Negotiating Leverage Against NVIDIASK Hynix: The Memory Supplier Suffering Its Own Customer's ConstraintHBM5: The Transmission-Line Bottleneck Transfers Margin to the Memory SuppliersWolfspeed: High-Voltage Moat Evaporates Under Thermal ScrutinyAccelink: The Valuation Anomaly in China's Optical Chip IndigenizationCoWoS Binds the Buyer, Not the LandlordArm: The Revenue-Mix Pivot Consensus Hasn't PricedInnolight: The Hyperscaler Pricing Ratchet Disguised as GrowthLong-Context Inference: Infrastructure Debt Becomes the LeverageInterconnection queue converts datacenter operator debt into utility equity premiumAmazon: The Capex-Margin InversionAixtron: Customer Capex Discipline Masks Downstream GrowthShin-Etsu: Hidden Exposure to Datacenter Grid FailuresTransformer scarcity is pricing power for industrial conglomerates, not a hyperscaler cost problemElectrical Steel Asymmetry: GE Vernova Captures Transformer Scarcity Without Steel ExposurePower Caps the AI Buildout — the bottleneck moved from chips to megawattsSemicap service is an installed-base annuity, and the market still prices these names as capex cyclicalsThe ABF Chokepoint Does Not Pay Its Owners — Buy the Bottleneck Fails HereTesla: AI-Buyer Margin Squeeze Hidden by Automotive NoiseABF substrate dual-gate arbitrage: substrate converters capture the spread the market assigns upstreamASML / China-DUV displacement is overpriced — the sell-off prices near-full substitution off a 5-machine data pointNAND Flash: Consumer Demand Collapse Disguises Enterprise Pricing PowerUALink: Consortium Activity Masks Structural IrrelevanceCo-packaged optics is a packaging play being misfiled as an optics transitionApplied Optoelectronics: Loss-Funded Market Access Captures Transceiver Supply ScarcityMicrosoft: Free Cash Flow Collapse Telegraphs Margin Compression Before It Hits the P&LRack Power Density: The Liquid Cooling Adoption MirageGPU collateral decay transmits to NVIDIA demand before physical constraints clearHBM moat vs. DRAM commoditization — is SK Hynix's HBM-mix 'miss' actually bullish?Advanced Packaging: The Scarcity Premium Flows Upstream, Not DownSilicon Wafer: Duopoly Ships into Oligopoly Margin Explosion, Captures NoneEaton: Margin expansion telegraphs demand scarcity more than growth can showMoE inference: stranded-power miners own the scarcest input, hyperscalers rent itMurata: The Margin the Bottleneck Hasn't ReachedAlphabet: The Capex Ratchet Liquidity Mispriced as Platform OptionalitySilicon cannot emit light — indium phosphide is the chokepoint CPO relocates but cannot removeOptical attach is set by topology, not by shipments — linear optical TAM models are mis-specifiedConstellation: Hyperscaler Monopsony Masks Nuclear Fuel TransmissionOracle: Monopsony Rent Capture Masked by Consolidated AccountingFoxconn: Customer Concentration Absorbs AI Margin Before It Hits the P&LThe China chokepoint moves from lithography to bonding — and the substitution trade is priced in the wrong categoryNVIDIA's Rent Compresses Through Software, Not SiliconPalantir: The Margin-Protected Infrastructure PlayCustom Silicon Pays a Different MerchantThe Upstream Is Single-Sourced and UnpricedSMIC: Subsidy Converts to Capacity Under Obscured Margin PressureThe Neocloud Rent Is Consumed By The Asset Before It Reaches EquityQualcomm's low capex masks an IP-only future neither consensus nor bears have pricedTransceiver margin expansion is a mirage: buyer concentration at 1.6T resets pricing power the market prices as durableSilicon capacitors expose Intel's package productivity deficitEMIB Shifts the Packaging Bottleneck to SubstratesMemory-demand durability — does algorithmic efficiency cap the memory super-cycle?Neocloud margin re-rating — do open-source deployment + rising rental rates break the thin-margin cap?HBM4 Qualification Concentrates Share, Not Spreads ItCost Per Token Is Set Outside the ChipArista: Hyperscaler capex intensity conceals margin compression riskGlobalFoundries: Customer Concentration Masks Structural Insulation from AI Capex WhiplashThe Megawatt Is the Unit of AccountChipflation Is a Wafer Allocation, Not a Demand Shock
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 timesSamsung memory long-term agreementsBuyer concentration tighteningHBF consortiumRent converting into capacityRent migrating upstreamRent not being competed awayUALink ConsortiumUltra Ethernet Consortiumco-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

KV Reuse Hazard Exponent (alpha)

value
1.53 dimensionless  2 se 1.5–1.56

The decay exponent of the KV/prefix-cache reuse hazard, measured at 1.53 (2 se: 1.50-1.56) on a public serving trace. It is the single free factor deciding whether flash tiering buys hours or months of cache residency, and therefore whether a 252 EB KV-NAND TAM closes. Nobody on either side of that debate had named the quantity.

reading
1.53, against the 0.77 the disclosed TAM requires. On this workload flash buys about eight hours of residency, not the 2.6 days the disclosure needs.

Why it matters

Economic cache residency is set by a break-even, not by a product decision: hold the marginal byte while P(reuse) x recompute_saving > storage_cost. The RATIO of two tiers' windows collapses to (price ratio)^(1/alpha), with the hazard's scale and origin cancelling entirely. TLC is ~25x cheaper per GB-day than DRAM and QLC ~44x, so at alpha 2.0 a flash window is five hours and at alpha 0.4 it is months. Nothing else in the flash-tiering debate moves the answer as much, and every other factor in the break-even is public or an arithmetic bound.

Method

trace
Moonshot's open-sourced Kimi serving trace — 12,031 requests, 182,790 distinct 512-token prefix blocks, a 3,537 s window.
estimator
Life-table hazard with CENSORED intervals kept in the risk set. The interval from a block's last access to the end of the trace is not 'never reused', it is 'not yet reused when the recording stopped'. Treating it as a completed observation manufactures decay out of the recording window and biases alpha upward — which biases the flash window down, which would have made this desk more confident against the TAM for a purely methodological reason.
validation
The hazard PEAKS at 77 seconds, which is the human read-and-reply interval. The estimator recovers a physical timescale it was never told about.

What had to come out first

Robustness

The verdict survives a 20x swing in the assumed token denominator (1e17 to 2e18 tokens/yr). It flips only at 3.83e18 — about 100x the one disclosed single-operator figure — and at that volume the required residency is 8.2 hours, which IS reachable. So the TAM has a VOLUME route even where the residency route fails, and that is the honest form of the caveat.

What it does not settle

ONE OPERATOR, ONE HOUR, CONVERSATION TRAFFIC. Gaps beyond 3,537 s are unobservable, so carrying the power law to a 2.6-day residency is a ~60x extrapolation beyond the observation window. A chat session reuses its prefix on a human turn-taking clock; an agent re-reading a document corpus is a different process. Alpha should be held as a PER-WORKLOAD quantity, not a constant.

What would refute it

Where it converges

The TLC/QLC price tape. Economics say the QLC window only opens if the hazard is flat; the tape shows TLC firming and QLC contract pricing flat, i.e. the window has not opened. A derivation and a price agreeing from entirely unshared inputs.

Computed by

Sources