←The map Research bank

Theses

Transformer lead times gate hyperscaler returns, not start dates Package Perimeter Arbitrage: Optics Suppliers Capture Value NVIDIA Cannot Retain NAND Flash: Datacenter Crowding-Out Masked by Low Consumer Pull CoWoS Binds the Buyer, Not the Landlord Long-context inference: memory constraint compresses cloud margin before GPU-poor plays feel it Co-packaged optics transfers margin to substrate integrators, not optics specialists Silicon Wafer: Duopoly Capturing Nothing as Downstream Rents Compound Advanced Packaging: The Scarcity Premium Flows Upstream, Not Down Amazon: The Capex-Margin Inversion Microsoft: The High-Multiple Cloud Business Is Already Subscale Foxconn: Margin Squeeze Masked by AI Server Revenue Mix SK Hynix: The Packaging Bottleneck Inverts the Margin Narrative GPU collateral decay gates buildout faster than physical supply Oracle: Monopsony Rent Capture Masked by Consolidated Accounting Innolight: Negative Operating Leverage Hiding in Hyperscaler Concentration Interconnection queue converts hyperscaler capex into generator rent Eaton: Margin expansion telegraphs demand scarcity more than growth can show Electrical Steel Asymmetry: GE Vernova Captures Transformer Scarcity Without Steel Exposure Constellation: Hyperscaler Monopsony Masks Nuclear Fuel Transmission ABF substrate dual-gate arbitrage: substrate converters capture the spread the market assigns upstream Rack Power Density: The Liquid Cooling Adoption Mirage Arm: The Royalty Stream That Scales With Everyone's Margin but Its Own HBM5: The Signal Integrity Bottleneck Transfers Margin to Silicon PHY Designers Shin-Etsu: Margin Expansion Masked by Commodity PVC Consolidation SMIC: Subsidy Converts to Capacity Under Obscured Margin Pressure Semicap service is an installed-base annuity, and the market still prices these names as capex cyclicals Alphabet: The Capex Ratchet Liquidity Mispriced as Platform Optionality Palantir: The Margin-Protected Infrastructure Play Accelink: The Valuation Anomaly in China's Optical Chip Indigenization Transceiver margin expansion is a mirage: buyer concentration at 1.6T resets pricing power the market prices as durable Qualcomm's low capex masks an IP-only future neither consensus nor bears have priced Murata: The Margin That Doesn't Move Applied Optoelectronics: Hyperscaler Procurement Arbitrage Driving Negative Spreads GlobalFoundries: Customer Concentration Masks Structural Insulation from AI Capex Whiplash Silicon capacitors expose Intel's package productivity deficit Arista: Hyperscaler capex intensity conceals margin compression risk Silicon cannot emit light — indium phosphide is the chokepoint CPO relocates but cannot remove MoE inference: converted miners capture memory scarcity, hyperscalers leak it Tesla: AI Pivot Insulated from Datacenter Bottlenecks Custom Silicon Pays a Different Merchant The ABF Chokepoint Does Not Pay Its Owners — Buy the Bottleneck Fails Here NVIDIA's Rent Compresses Through Software, Not Silicon The China chokepoint moves from lithography to bonding — and the substitution trade is priced in the wrong category The Upstream Is Single-Sourced and Unpriced Optical attach is set by topology, not by shipments — linear optical TAM models are mis-specified HBM moat vs. DRAM commoditization — is SK Hynix's HBM-mix 'miss' actually bullish?Power Caps the AI Buildout — the bottleneck moved from chips to megawatts EMIB Shifts the Packaging Bottleneck to Substrates ASML / China-DUV displacement is overpriced — the sell-off prices near-full substitution off a 5-machine data point Cost Per Token Is Set Outside the Chip Chipflation Is a Wafer Allocation, Not a Demand Shock The Neocloud Rent Is Consumed By The Asset Before It Reaches Equity Memory-demand durability — does algorithmic efficiency cap the memory super-cycle?HBM4 Qualification Concentrates Share, Not Spreads It Neocloud 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 supply CoWoS advanced-packaging capacity Conventional DRAM and NAND supply (HBM crowding-out)GPU residual value as loan collateral Grid interconnection queue position Heavy-duty gas turbine delivery slots Large power transformer lead times Buyer concentration tightening HBF consortium Rent converting into capacity Rent migrating upstream Rent not being competed away co-packaged-optics displaces copper co-packaged-optics displaces optical-transceiver emib displaces cowos glass-substrate displaces abf-substrate hybrid-bonding displaces euv silicon-capacitor displaces mlcc Chip design Datacenter math Efficiency arrives in steps, not trends Inference shape Memory economics Photonics Tau scaling Token math Tokenomics Capacity arriving — CoWoS (Chip-on-Wafer-on-Substrate)Capacity arriving — HBM4 Capacity arriving — Silicon Wafer

Inference shape

driver a named mechanism, not a conclusion

Agentic serving pushes context and concurrency at once, so the workload grew faster than compression did and the bytes migrate down the hierarchy.

The path

node	effect
Long-context inference (KV-cache growth)	+0.35
Mixture-of-experts inference	+0.35
Inference serving stack (vLLM / TensorRT-LLM class)	+0.35
Gpu Accelerator	+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.

Cost Per Token Is Set Outside the Chip

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.