Single leading-edge foundry (TSMC) offloads capacity risk to customers via conservative investment; geopolitical concentration in Taiwan adds tail risk; resulting shortages compel Google/Am…

Game theory: if US achieves military-significant AI superiority, China's optimal response is destroying TSMC. Deep China dependency across manufacturing (not just chips) makes decoupling an…

Meaningful US AI military superiority creates incentive for China to destroy TSMC; US supply chain decoupling is economically irrational absent conflict (insurance too expensive); magical t…

Deep multi-tier dependency on China (not just TSMC) makes decoupling economically irrational for individual firms — 'insurance policy astronomically expensive'; diversification only occurs…

Deep, multi-layered reliance on Chinese manufacturing (beyond TSMC) makes decoupling economically irrational absent conflict — 'insurance' investments (India iPhone assembly, US fabs) are m…

Open models (Llama, GLM, Kimi) carry high inference costs; closed models with advertising feedback loops (Meta, Google) have superior verification via human click/purchase signals at global…

The AI buildout is consuming capital faster than revenue arrives — free cash flow, then debt markets, now equity issuance — creating a timing mismatch where a financial crisis could occur e…

Railroad bubble created century-lasting assets (BNSF still funding Google via Berkshire); GPU clusters depreciate fast but power infrastructure endures; if AI capex bubble leaves energy abu…

Per-seat pricing hid extreme cost variance between casual and heavy AI users. Usage-based pricing forces monthly budget decisions, exposing product value and enabling churn — but Microsoft…

Consumers won't pay for software and don't want productivity — Dropbox proved this; OpenAI's subscription ceiling forces ad pivot; Google/Meta have verification machines (ad markets) for AI…

Consumers won't pay for productivity (Dropbox lesson) — advertising is the only scalable consumer model; OpenAI's late pivot validates this. Enterprise usage-based pricing (Microsoft E7) br…

Consumers won't pay for software or productivity; OpenAI's subscription-first approach repeated Dropbox's error. Advertising aligns incentives (advertisers bear cost increases, zero consume…

Capex rising to $800B-$1.3T/year while revenue lags; capital curve progressing from FCF to debt to equity to pension/insurance money; risk of financial blowup if ROI timing mismatch persist…

Samsung/SK Hynix/Micron discipline created a target for memory reduction efforts (Apple lobbying for Chinese memory, algorithmic memory efficiency) — similar to Iran's Hormuz closure spurri…

Three-player memory market (Samsung, SK Hynix, Micron) avoids past over-investment mistakes but initially under-invested in HBM. Apple and model optimizers now target memory reduction; Chin…

TSMC's 30-year fab horizon bias toward underinvestment (2023-2025 growth deceleration) transferred overcapacity risk to big tech as foregone revenue; acute scarcity now makes previously irr…

Three-player memory oligopoly (Micron, Samsung, SK Hynix) learned boom/bust discipline but was slow to recognize AI/HBM secular shift; Samsung's counter-cyclical investment playbook won las…

TSMC's cultural bias against overcapacity (30-year fab horizons) makes them underinvest relative to AI demand, shifting shortage risk to customers who face foregone revenue — this dynamic e…

TSMC optimizes for 30-year fab utilization, deliberately under-investing vs. peak demand. Risk transfers to customers as foregone revenue. This scarcity economically justifies the pain of q…

OpenAI (mainline) and Anthropic (evangelical) operate with 'power of belief' — creating God fuels execution; Google only needs search not to die quickly; Meta's frontier push is pure Zucker…

Microsoft builds IBM-style middleware/harness for enterprise AI adoption — stable, backward-compatible, manages model churn; but AI agents (Codex, Claude Code) automate the tedious migratio…

Enterprises prefer a stable, backward-compatible platform that manages model churn (IBM 1990s playbook) over best-in-class point solutions. However, AI's ability to automate tedious migrati…

Microsoft avoids frontier training, builds enterprise AI harness (IBM 1990s middleware playbook); E7 usage pricing untethers revenue from headcount, forces monthly budget friction, risks pr…

Google (TPU), Amazon (Trainium), Microsoft (Maia) have lower cost of capital than neo-clouds; they sell chips externally as commodities (not differentiated), gaining R&D leverage without ca…

Natural gas (West Texas), behind-the-meter generation (xAI), and nuclear restarts have brought US power online faster than anticipated; energy abundance gives hyperscalers time to improve c…

Google and Amazon's custom chips benefit from lower cost of capital and don't need to maintain Nvidia's margins — they sell chips as commodities to amortize R&D, while Nvidia's circular fin…

Google selling TPUs to Anthropic, Amazon selling Trainium externally — both have lower cost of capital than neo-clouds, sell chips as commodities not differentiated products, don't cannibal…

Google/Amazon sell chips externally not for differentiation but as commodities to amortize R&D. They have lower cost of capital than neo-clouds, no CUDA lock-in need at scale, and don't can…

Behind-the-meter gas (Elon/Texas), nuclear restarts, and rapid US energy response surprised even Jensen Huang; abundant power removes token-efficiency constraint that would favor Nvidia, gi…

Behind-the-meter gas, nuclear restarts, and West Texas buildout brought power online faster than anticipated, extending the window where hyperscalers can iterate custom silicon (Trainium, T…

Faster-than-expected US power additions (gas, nuclear restarts, behind-the-meter) give hyperscalers more time to mature custom silicon (Trainium, TPU), eroding Nvidia's token-efficiency adv…