Gianfranco's November 2025 Reading List

What’s worth reading this month—and why does it matter?

Each day, I devote three hours to reading. Occasionally an author writes a piece that feels like a gift—crafted with intention for the reader. It tends to be a piece that encourages a reader to pause, invites them to feel how a strong idea, delivered with empathy, can resonate.

Few writers achieve this.

Each selection did one of three things for me: it sharpened a thesis, challenged a prior, or opened a new line of inquiry. That’s the bar for inclusion.

In this curation practice, I believe the hero of this story isn’t the author of any single essay, nor me as curator, but you—the reader.

If these pieces don’t move you, I haven’t honored the attention you’ve given me. Hold me to that.

Use this as a tasting menu: no set order, categories for context, and a brief note on what each means for builders and capital allocators.

AI Systems & Interfaces

• GPT-5 Set the Stage for Ad Monetization and the SuperApp - SemiAnalysis
  • OpenAI’s “router,” not raw model gains, is the release: a control plane that detects intent, meters reasoning, and steers cost. That enables monetizing 700m+ free users via agentic purchasing—high-value queries trigger deeper compute, tool use, and checkout, with take-rate economics rather than display ads. Hiring Fidji Simo and integrations (Stripe, Shopify, Instacart) telegraph a consumer superapp, while variable-cost answers erode search’s fixed-supply ad moat.
  • Read More

• AI Isn’t Replacing Radiologists - Works in Progress
  • Radiology was the obvious testbed for AI displacement—digital images, repeatable tasks, clear metrics—yet headcount and pay are at highs. Benchmark-beating models fail out-of-distribution, remain per-finding “islands,” and hit regulatory and malpractice walls against autonomy; hospitals keep humans in the loop. Radiologists also do far more than reads, and demand is elastic: faster tooling drives more imaging. Net effect: early AI diffusion raises throughput and triage, not substitution.
  • Read More

• OpenAI Is Building a Banker ($) - Bloomberg Opinion
  • OpenAI’s “Project Mercury” hires 100+ ex-bankers at $150/hour to train models to build Excel deal models with bank-grade formatting. The goal is trust, not cleverness: italicized percentages as a proxy for “someone checked this.” If AI internalizes these tacit quality norms, banks can offload the checkable grunt work — and the apprenticeship ladder that produced senior bankers thins. Alumni get paid to codify their own replacements.
  • Read More

• Scaling the Memory Wall: The Rise and Roadmap of HBM - SemiAnalysis
  • HBM—not compute—is the throttle for AI throughput, and models expand to fill any new bandwidth and capacity. SemiAnalysis maps the bottlenecks: TSV yield, shoreline real estate, and a fragile bonder oligopoly, with CXMT/Huawei racing under export limits. JEDEC height relief enables 12–16-high stacks now, while hybrid bonding slips to HBM4E. Next: custom base dies, memory-controller offload, repeater PHYs, and LPDDR/HBM combos extend “beachfront,” shifting value toward HBM suppliers as Nvidia-driven bit demand explodes.
  • Read More

• Where We Go From Here - Fabricated Knowledge
  • O’Laughlin maps a likely GPU-led bubble powered by rate cuts, surging productivity data, and datacenter capex that diffuses gains into the real economy. Founder-controlled giants (Meta, Google) and easy credit tilt incentives toward overspend, while Oracle shows how negative FCF can buy share. The meta-risk: a coalition around OpenAI/Nvidia aligns suppliers and financiers to keep fueling demand. Timing unknown; animal spirits have policy cover.
  • Read More

• Sora the App, Sonnet 4.5 and the Question of Models as Processors ($) - Stratechery
  • OpenAI’s Sora app targets the 9% who remix, Meta’s Vibes courts the 90% who consume, and YouTube arms the 1% who create—product bets mapped to the 90/9/1 rule. OpenAI touts a “ChatGPT moment” for video, but attention hinges on consumption, not easy creation. Anthropic’s Sonnet 4.5 boosts agents, yet Cognition had to rebuild Devin, hinting models evolve non-compatibly. That undermines the tidy “models-as-processors” assumption for app builders.
  • Read More

• OpenAI Instant Checkout, AI and Long Tail E-Commerce, Is AI Different? ($) - Stratechery
  • OpenAI’s Instant Checkout pushes ChatGPT from product discovery into conversion, skimming a fee when buyers pick among multiple merchants. AI flips long-tail commerce: instead of merchants finding customers (Meta), customers find niche sellers (Shopify/Etsy). Lean-forward chat makes an in-app buy tolerable, even if tooling trails native storefronts. The protocol quietly creates an affiliate-like ranking incentive: non-bespoke items likely favor merchants enabling Instant Checkout.
  • Read More

Operator Playbook

• Graham Duncan: Talent Whisperer - Join Colossus
  • Duncan’s edge is a talent-first investing system: “long/short people,” not assets. Exhaustive referencing, elite-network pattern recognition, and bespoke roles create environments where rare operators compound. He treats seeding like casting, waiting for clear “source” ownership and an optimal grip rather than forcing timelines. The payoff is durable outperformance and a repeatable way to turn information networks into advantage.
  • Read More

Market Structure & Moats

• Paradigm Shifts and the Winner’s Curse - Stratechery
  • Thompson argues Apple and AWS defined the smartphone era, but prior winners often misread new paradigms. Apple and Amazon frame AI as commodity extensions of their models—on-device private context and low-cost cloud inference—which conveniently preserves control. If AI is a true paradigm of agents and augmentation, performance and model leadership trump cost, tilting advantage toward Google. The subtext: success breeds blind spots; the winner’s curse looms for incumbents while Google’s amorphous culture adapts across shifts.
  • Read More

• Tokens Are Getting More Expensive - Substack
  • Every 10x drop in cost applies to yesterday’s models; demand follows the frontier, where token prices hold steady while token counts explode via agents and test-time compute. Flat-rate plans become a margin short squeeze—Claude Code’s unlimited tier showed loops devour budgets. The viable paths shift to usage-based billing, enterprise entrenchment, or vertical integration that monetizes infra. Prosumer “unlimited” is marketing, not economics.
  • Read More

• Letter to a Young Person Worrying About AI - Commoncog
  • Chin argues prediction is a trap; careers are won by fast adaptation under uncertainty. History shows job displacement unfolds slower than hype because technologies live inside sociotechnical systems, not in isolation. Calibrating with real diffusion stories (containers, horses, rickshaws) reframes AI fear from imminent replacement to observable change over years. The stance shifts from forecasting to monitoring field evidence and repositioning as reality, not rhetoric, evolves.
  • Read More

• KPop Demon Hunters, Sony’s Risk, The Netflix Aggregator ($) - Stratechery
  • Netflix’s first true animated megahit didn’t just crown a new IP—it exposed a strategic inversion in Hollywood. Sony’s decision to sell the film to Netflix traded infinite optionality for guaranteed profit, embodying a content-as-arms-dealer model. Netflix’s algorithmic distribution, not box office theatrics, created the phenomenon—proof that reach, not release, defines cultural leverage in streaming’s post-theatrical era.
  • Read More

Capital & Macro

• OpenAI, Nvidia Fuel $1 Trillion AI Market With Web of Circular Deals ($) - Bloomberg
  • OpenAI and Nvidia are knitting circular deals—equity, pre-purchases, and capacity backstops—that make AI’s trillion-dollar buildout look self-referential. Cash-burning developers secure chips and debt; clouds secure clients; chipmakers secure demand, binding balance sheets across the stack. The structure amplifies capex and hype while monetization lags, echoing dot-com cross-buy dynamics despite more tangible products. If demand normalizes, the unwind hits startups, power, real estate, and credit pulled into the flywheel.
  • Read More

• How Common Is Accidental Invention? - Brian Potter — Construction Physics
  • Potter examines 190 “major” inventions (1800–1970) and finds only 14 (~7%) were truly accidental, with chemistry dominating (8 of 14). Most serendipity occurred inside directed research; only three arose outside formal R&D. Frequency is steady—roughly once per decade. Compared to ~40% “multiple invention,” accident is the exception, with chemicals closer (~17% accidental vs ~33% multiple), which quietly recasts what discovery processes actually cultivate.
  • Read More

• Class of ’25 - Net Interest
  • Fintech’s IPO window has reopened, but the pattern from 2021 looms: big first-day pops, weak after-market performance, and returns concentrated in a few scaled names. The 2025 cohort skews older and toward trading/financialization (Circle, Bullish, eToro, MIH), reflecting a market monetizing velocity more than payments. Six of eleven new listings now trade below issue, with notable spikes masking fragility. With premium assets staying private longer, adverse selection risk rises even as the pipeline swells.
  • Read More