SemiAnalysis produces the most rigorous semiconductor analysis on the internet — but it requires significant background to parse. McKinsey produces beautiful, accessible reports — but they rarely tell you what to buy. The financial press covers earnings beats and misses but rarely explains why the chip yield rate at TSMC’s N2 node matters to NVIDIA’s gross margin.
This site aims to sit in the middle: precise enough to be useful, accessible enough to be immediate, and always linked back to stocks and real money.
This is not financial advice. It is research and analysis. The goal is to give you a framework and the underlying data to make your own informed decisions. Always do your own research.
Over the coming weeks, this site will build out deep-dive articles on each sector, stock-specific pages with technical and fundamental analysis, and infographics that make complex technology roadmaps immediately legible.
The blog will eventually sync with a Substack publication for those who prefer email delivery.
Welcome. The future is already here — it’s just not evenly distributed.
]]>CoWoS (Chip-on-Wafer-on-Substrate) is TSMC’s advanced packaging technology that stacks the compute die and HBM memory on a silicon interposer. Every H100 and B200 requires it. TSMC’s CoWoS capacity is the single most important constraint on AI accelerator supply.
In 2024, TSMC could produce roughly 50,000–60,000 CoWoS substrates per month. Demand for H100/H200 exceeded this by a significant margin through most of the year. TSMC has been aggressively expanding CoWoS capacity through:
Estimated CoWoS capacity by end 2026: ~90,000–100,000 units/month. Still likely below peak demand if hyperscaler capex guidance holds.
Investment implication: TSMC’s packaging revenue is growing faster than wafer revenue. AMAT and LRCX supply key deposition and etch equipment for advanced packaging. Both are CoWoS beneficiaries beyond just the node story.
Each B200 GPU ships with 192GB of HBM3E (High Bandwidth Memory 3E) across 8 stacks. The NVL72 rack thus requires 72 × 8 = 576 HBM3E stacks. SK Hynix supplies the majority; Micron (MU) is qualifying to take meaningful share in 2025–2026.
HBM is a premium product with high ASPs and tight supply. MU’s HBM margin profile is significantly better than its DRAM commodity business. A successful HBM3E qualification at NVIDIA is a multi-year revenue and margin catalyst for Micron.
120kW per rack requires liquid cooling. Air cooling is not viable at this power density. This creates a demand signal for:
SMCI’s advantage is speed-to-market with custom cooling configurations. The company’s share price volatility has been extreme, but its positioning in the AI server ecosystem is structurally sound.
Disclaimer: This is analysis and commentary, not investment advice.
]]>Google Willow (December 2024) was the most significant quantum milestone since IBM’s 50-qubit system in 2017. Willow demonstrated that error rates decrease as qubit count increases — a necessary condition for fault-tolerant quantum computing. This was theoretically expected but experimentally unproven until Willow. It moved the timeline for fault-tolerant QC from “maybe never” to “2030s with high probability.”
IBM’s Quantum Roadmap targets 100,000 physical qubits by 2033. The company has been remarkably consistent in hitting its roadmap milestones. The challenge: qubit count is not the binding constraint — error correction overhead is. A single fault-tolerant logical qubit may require 1,000–10,000 physical qubits.
Trapped-Ion vs Superconducting — IonQ (trapped-ion) argues that native gate fidelity is higher and connectivity is better than superconducting qubits (Google, IBM, Rigetti). The disadvantage is gate speed: trapped-ion operations take microseconds vs nanoseconds for superconducting. For near-term NISQ applications, this matters less than long-term fault-tolerant scalability.
IonQ (IONQ) — The most mature revenue story in the public quantum space. Government contracts, cloud access via AWS/Azure/GCP, and growing bookings. Valuation is high relative to near-term revenue but investors are paying for 2030s potential. The stock trades as a binary bet on trapped-ion becoming the dominant architecture.
Rigetti Computing (RGTI) — Superconducting architecture, significant technology capability, but persistent execution and commercial challenges. Small market cap ($400M range) makes it highly volatile on any news.
D-Wave (QBTS) — Quantum annealing, not gate-based quantum computing. Different use cases (optimisation problems). Revenue generating but architecture debate is a persistent overhang.
Arqit Quantum (ARQQ) — Quantum key distribution and encryption software. Earlier commercial revenue potential than gate-based systems, but smaller TAM.
The combined market cap of all listed quantum pure-plays is under $10 billion — less than one percentage point of NVIDIA. The market is not pricing in quantum success as a near-term reality. It is pricing in optionality: small positions that pay off enormously if one or more companies wins the fault-tolerant race.
The sensible portfolio approach: treat quantum stocks as venture-style positions, size accordingly, and monitor technical milestones rather than quarterly earnings.
Key milestones to watch:
Disclaimer: This is analysis and commentary, not investment advice.
]]>