Markets · Call
TSMC: own the bottleneck, not the brand
Own the bottleneck, not the brand: the chokepoint every AI chip depends on, at a discount to all of them.
Stories / 13 pieces · 2 weeks
Everything we worked on, week by week.
The research behind each day: the markets calls, the AI tutorials and reviews, the startups and the code, on one page as it shipped.
The rollShipped daily
The topics
Each one filters the list below.
News, tutorials, reviews and model notes.
First looks at the startups we work with.
What got built, and how to build your own.
The record
Every call, graded when it resolves.
This week
w/c 8 Jun, newest first.Broadcom: the right custom-silicon thesis at the wrong price
The right custom-silicon thesis at the wrong price — watching the most expensive, lowest-quality way to play it.
Right now / the work so farcounted off this page
Published13pieces
Calls on file13graded in public
Weeks2of work
Everything
Newest first, week by week.w/c 8 Jun5 pieces
- Markets · CallTSMC: own the bottleneck, not the brandOwn the bottleneck, not the brand: the chokepoint every AI chip depends on, at a discount to all of them.
- Markets · CallBroadcom: the right custom-silicon thesis at the wrong priceThe right custom-silicon thesis at the wrong price — watching the most expensive, lowest-quality way to play it.
- Markets · CallSK Hynix: the largest HBM rent, valued like a Korean cyclicalThe largest slice of the HBM scarcity rent, valued like a Korean cyclical — with a Nasdaq listing as the re-rating catalyst.
- Markets · CallMicron: priced like a cyclical at the top, contracted like a utilityPriced like a peak-cycle memory-cyclical, contracted like a utility — a re-rating bet, not an earnings bet.
- Markets · CallNVIDIA: paying a market multiple for a franchise financing its own demandStill constructive on the best franchise in AI infrastructure — but the risk has migrated from the income statement to the balance sheet.
w/c 13 Apr8 pieces
- Markets · CallParameter Scaling Is Dead. Multi-Dimensional Scaling Is the New Paradigm.AI scaling has moved from raw parameter count to four memory-hungry dimensions — inference-time compute, MoE, data curation, architecture — making memory bandwidth the universal bottleneck.
- Markets · CallInference Costs Will Fall 90% by 2028 — The Memory Stack Is the MechanismFour independent memory-stack improvements — HBM4, KV-cache compression, optical interconnect, and architecture — could compound to a ~10x fall in per-token inference cost by 2028.
- Markets · CallBTC Miners Pivoting to AI Infrastructure: The Power ArbitrageBitcoin miners already hold the power, cooling, and space AI needs — mining revenue is expected to fall from 85% to under 20% of total as they convert to AI hosting.
- Markets · CallEnergy Ownership Is the New AI MoatPower, not GPUs or algorithms, is AI's real moat — a location-specific power contract is scarce and non-fungible in a way chip access never is.
- Markets · CallNuclear Is the Only Power Source That Scales for AI Data CentersAI data centers need reliable, carbon-free baseload at massive scale — nuclear is the only source that delivers all three, and owners can charge a premium for it.
- Markets · CallTransformer Successors: State-Space Models and Linear Attention Are Worth WatchingTransformers' quadratic memory scaling is the KV-cache problem. State-space and linear-attention models scale linearly — hybrids blending both look like the likely path.
- Markets · CallLumentum: The Optics Play That Bridges Power and MemoryBullish · LITE
- Markets · CallKV Cache Compression Will Cut Inference Costs 50% Before HBM4 Ships at ScaleKV caches are AI inference's biggest bottleneck. Software compression (6x cited) is shipping faster than HBM4 hardware — and will drive the cost collapse of 2026-2027.