Personal Learnings← Reading Room

Tech & AI

Stratechery

Ben Thompson

40 issues · 34 keepers · 9 tier-5 · 25 tier-4

The AI Buildout — CapEx, Power & the Bubble Question

2 tier-5 · 4 tier-4

The dominant thread of the quarter. Thompson works out why the hyperscaler capex wave is justified (demand, not speculation), names the binding constraints (power, then silicon), and ultimately reverses his own bubble call. The token-demand framework — three LLM paradigms (ChatGPT → reasoning → agents), each a step-change in compute intensity — is the recurring engine under most of these pieces.

Agents Over Bubbles

TIER 5 Mar 16, 2026

Thompson publicly reverses his standing position to argue we are NOT in an AI bubble, marshaling a multi-part framework: the three LLM paradigms (ChatGPT/o1/agents) have addressed LLMs' core flaws, agents collapse the 'agency' requirement so demand can skyrocket without mass human adoption, and enterprise economics make AI-driven layoffs and capex structurally inevitable. The most analytically important claim is that agents require model-plus-harness integration, so profits flow toward integrated players (Anthropic, OpenAI) rather than commoditizing — directly rebutting Dediu's 'Apple's brilliant move' and the model-commodity thesis. A genuinely thesis-shifting piece worth returning to on AI economics.

agentsai-bubbleai-capexanthropicharness-integrationenterprise-ai

An Interview with Jeremie Eliahou Ontiveros and Ajey Pandey About Building Power for AI

TIER 5 Jan 8, 2026

A definitive deep-dive with SemiAnalysis's energy analysts on how AI labs, led by xAI's Memphis playbook, are bypassing a gridlocked interconnection queue by building behind-the-meter onsite gas generation, reshaping US electrical infrastructure from the demand side. It explains why the power bottleneck (interconnection prisoner's dilemma, training-load grid-blackout risk, turbine lead times, the $12B-revenue-per-gigawatt math that makes power cost almost irrelevant) is the binding constraint on the AI buildout, while arguing the bubble may catalyze a long-deferred grid rebuild. This is the quarter's standout reference piece on AI compute economics meeting physical energy reality.

ai-energydata-centersbehind-the-meternatural-gaselectrical-gridai-capex

Anthropic's Skyrocketing Revenue, A Contract Compromise?, Nvidia Earnings

TIER 4 Mar 4, 2026

Three connected segments arguing the Anthropic-government standoff matters more as Anthropic's enterprise revenue reaches escape velocity (~$20B run-rate, possibly passing OpenAI on coding/API/enterprise strength) — which paradoxically makes a contract compromise more urgent since tech doesn't need the government as a customer. The lasting contribution is the three-inflection-points framework for token demand (ChatGPT → reasoning/o1 → agents/Opus 4.5, the last a squared increase), used to argue Nvidia bears are inconsistent with software bears: agent-driven disruption of software means more, not fewer, tokens. The token-demand framework recurs across the quarter's pieces.

anthropicnvidiaai-capexagentsenterprise-aireasoning-models

Google Earnings, Google Cloud Crushes, Search Advertising and LLMs

TIER 4 Feb 9, 2026

Thompson justifies Google's shocking ~$180B CapEx guide by reading backwards from results: Google Cloud's 48% growth with record 30% margins signals supply (not demand) constraint, validating the spend. The more original insight is that LLMs expand Search's addressable market by letting Google monetize long and non-English queries — increasing inventory, the single most powerful ad-revenue lever — making OpenAI's entire proposed ad business merely a 'meh' side project for Google. Matters as the clearest case for why Google is the structurally advantaged AI winner.

googleai-capexsearch-advertisinggoogle-cloudllmsopenai

Amazon Earnings, CapEx Concerns, Commodity AI

TIER 4 Feb 10, 2026

Thompson, having blessed Google's CapEx, explains why Amazon's $200B plan makes him nervous: Amazon lacks the cash flow to self-fund it (crossing the 'Rubicon' into debt-financed AI, weakening the no-bubble argument), operates at thinner margins where AWS is itself the high-margin business, and faces AI threats to its high-margin ads (its Aggregator position depends on shoppers starting on Amazon). Yet he ultimately understands the spend: Amazon's Trainium bet is a commodity-AI play where margin comes from a lower cost structure, which only pays off once supply meets demand — so bailing now makes no sense. Matters for the debt-financing inflection and the commodity-vs-differentiated AI compute framing.

amazonawsai-capextrainiumcommodity-computeai-bubble

Meta Compute, The Meta-OpenAI Battle, The Reality Labs Sacrifice

TIER 4 Jan 14, 2026

Reads Zuckerberg's 'Meta Compute' initiative not as a cloud-services play but as a bet that infrastructure (and the ad-funded free cash flow to pay for it) will be the binding constraint in AI, positioning Meta against a compute-starved OpenAI that lacks a scalable revenue model. The same-day Reality Labs layoffs are the through-line: Meta is sacrificing the metaverse to placate skeptical investors and free resources for the AI compute war. Coherent strategic framing connecting compute, business model, and the Meta-OpenAI rivalry.

metaopenaiai-capexreality-labsadvertisinghyperscalers

Semiconductors & the Foundry Bottleneck

2 tier-5 · 4 tier-4

Thompson's signature hardware thread. The “TSMC Brake” — a foundry rationally under-investing because its costs are almost all CapEx/depreciation, offloading overbuild risk onto its customers — points to a severe ~2029 chip shortage, with the only fix being to fund Intel/Samsung as credible competition (which would solve the Taiwan geopolitical risk “for free”). Around it sit Nvidia's strategic pivot from one-GPU-for-everything to three architectures, and Intel's self-inflicted capacity failure.

TSMC Risk

TIER 5 Jan 26, 2026

This front-page Article reframes 'TSMC risk' away from the obvious Taiwan-China geopolitical reading toward the underappreciated structural risk: TSMC's rational capex conservatism (since a foundry's costs are almost all CapEx/depreciation) offloads enormous foregone-revenue risk onto hyperscalers, a harm already being realized as demand outstrips chip supply. The thesis—that the only fix is funding Intel/Samsung as genuine competition, which would also incidentally solve the geographic risk—is a durable analytical framework for the AI-buildout bottleneck worth returning to. A definitive statement of the foundry-competition argument.

tsmcsemiconductorsfoundry-competitionai-capexintelgeopolitical-risk

An Interview with Nvidia CEO Jensen Huang About Accelerated Computing

TIER 5 Mar 17, 2026

A post-GTC-keynote interview with Jensen Huang that doubles as a strategic statement: Nvidia is not a GPU company but an accelerated-computing company that must now accelerate all human-built software (SQL, Excel) for agents to use, and reasoning/grounding/tool-use crossed AI into real economic value over the past year. Huang explains the rationale for new architectures (Vera CPUs optimized for single-thread/I-O to keep GPUs fed; the Groq licensing deal for ultra-low-latency decode), the five-layer-cake view of winning AI, the China/open-source diffusion argument, and a pointed critique of how 'doomers' captured Washington and depressed AI's popularity. High-insight access to the most pivotal figure in AI compute.

nvidiajensen-huangaccelerated-computingagentsgroqchina-chips

Jensen Huang and Andy Grove, Groq LPUs and Vera CPUs, Hotel California

TIER 4 Mar 18, 2026

A synthesis of GTC 2026 framing Nvidia's pivot from one-GPU-to-do-it-all to selling three distinct architectures (GPU, Groq LPU, Vera CPU) as a Grovian strategic inflection point. Thompson reverses last year's single-architecture defense: Nvidia now disaggregates inference even within decode — Vera Rubin for prefill and KV-cache-heavy attention, Groq-style chips for bounded feed-forward — and adds agent-optimized CPUs to stop idle GPUs from leaving a margin umbrella for cheaper ASICs. The 'Hotel California' conclusion: Huang's real unifying theory is paranoia, leaving no opening for any alternative so no customer can ever check out. Sharp technical-strategic read of the announcements.

nvidiagroqvera-cpuasicsinferencestrategic-inflection

TSMC Earnings, The TSMC Brake Revisited, Why AI Needs Foundry Competition

TIER 4 Jan 21, 2026

TSMC's Q4 earnings validate Thompson's 'TSMC Brake' thesis: CEO C.C. Wei admits a demand-supply gap from years of underinvestment, and even the raised $52-56B capex won't add meaningful capacity until 2028-29, while Wei's candid fear of 'holding the bag' explains his conservatism. The argument is that since silicon (not power) is now the bottleneck, the only way to remove the brake is for hyperscalers and chip firms to bring Intel/Samsung online as real competition, shifting risk back onto foundries. Strong setup that the next day's front-page Article expands.

tsmcsemiconductorsai-capexfoundry-competitionintelcapex

Intel Earnings, The Agentic Opportunity, Intel's Mistaken Pessimism

TIER 4 Jan 27, 2026

Intel's 17% post-earnings plunge is read as self-inflicted: it retired Emerald/Granite Rapids capacity just as the shift to inference and agentic workflows (which are CPU-heavy for scheduling and tool calls) drove surging server-CPU demand, leaving it unable to capture a windfall. The through-line ties to the prior TSMC pieces—Intel's own capacity failure undermines its credibility as the foundry competitor the industry needs by 2029, and new CEO Tan's pessimism mirror-images Gelsinger's over-optimism. Substantive analysis linking the agentic-compute shift to Intel's strategic missteps.

intelsemiconductorsagentic-aiserver-cpusfoundry-competitioninference

Nvidia at CES, Vera Rubin and AI-Native Storage Infrastructure, Alpamayo

TIER 4 Jan 7, 2026

Reviews Nvidia's CES announcements as proof that AI is crowding out the entire consumer-electronics supply chain (memory prices in particular), with the standout being the BlueField4 'AI-native storage' rack that offloads KV-cache context memory onto the east-west plane to enable far longer context windows for reasoning models and multi-agent workflows. Thompson's clear explanation of why LLMs regenerate every token from scratch (and why that drives the storage rearchitecture), plus the Alpamayo vision-only self-driving bet as another Bitter Lesson data point, makes this substantive systems analysis.

nvidiavera-rubinkv-cacheai-infrastructureself-drivingmemory

AI & the Future of Software — Aggregation, SaaS Survival & Agents

2 tier-5 · 4 tier-4

The quarter's other big thesis: AI-written code collapses the cost of software the way the internet collapsed distribution — which dooms the siloed SaaS land-grab even as it makes software companies the chief beneficiaries. “Microsoft and Software Survival” is the anchor the earnings pieces and interviews orbit; the model-plus-harness integration argument explains where the margin actually accrues.

Microsoft and Software Survival

TIER 5 Feb 3, 2026

Using Microsoft's $357B post-earnings rout as the entry point, Thompson lays out the quarter's defining thesis: AI-written code is collapsing the cost of software the way the internet collapsed distribution for publishers, which dooms the siloed SaaS land-grab even as it makes software companies the chief beneficiaries of AI coding. He reframes Microsoft's Azure 'miss' as a correct capacity-allocation choice (favoring higher-LTV first-party Copilots and R&D over third-party cloud) and introduces durable framings — 'Work IQ'/identity as the agent layer, the shrinking value of per-seat licensing, and 'token foundries' (pure-play neoclouds vs. self-prioritizing hyperscalers). This is the anchor argument the rest of the quarter's earnings pieces and interviews orbit.

microsoftsaasai-codingagentshyperscalersaggregation-theory

An Interview with Benedict Evans About AI and Software

TIER 5 Feb 5, 2026

A wide-ranging, framework-dense conversation that is arguably the definitive articulation of the 'is software dead?' question: Evans argues software's real value was never the code but the institutionalized workflow, route-to-market, and 'throat to choke,' so AI lowers code cost without killing software companies — while the 'middle managers will roll their own' thesis is delusional. The pair build several reusable lenses: AI as a new kind of unbundling (separating the work from the product), AI possibly collapsing coordination costs (the Gosplan/central-planning analogy) toward larger firms, the OpenAI 'cargo cult' (importing app stores/ads/dev conferences that solve no real problem), LLMs as a pre-baked mechanical turk that could know things without a user base, and the recurring point that paradigms (the feed, native mobile) take a generation to emerge. Lasting reference value across AI strategy.

benedict-evanssaasopenaiagentsaggregation-theoryai-business-models

Oracle Earnings, Oracle's Cloud Growth, Oracle's Software Defense

TIER 4 Mar 11, 2026

An earnings analysis using Oracle's blowout quarter (84% infrastructure growth, $553B RPO, customers self-funding GPUs, 531% Multicloud Database growth) to argue the null hypothesis on AI infrastructure demand is that it's real and accelerating. The more interesting thread is Oracle's 'software defense' case against the SaaS-apocalypse thesis: AI coding lets incumbents expand into adjacent products and embed agents in mission-critical systems no one will rip out — the same survival pattern Thompson laid out in 'Microsoft and Software Survival.' Reality-checks the bear case with concrete data.

oracleai-capexhyperscalerssaasmulticloud-databaseai-infrastructure

Shopify Earnings, Shopify's AI Advantages

TIER 4 Feb 18, 2026

Against the indiscriminate 'SaaS is dead' panic, Thompson makes the affirmative case that Shopify is a prime AI winner, enumerating a reusable checklist of AI advantages: exclusive commerce data plus a way to leverage it (Shop Campaigns, Product Network), AI-driven long-tail market expansion (orders from AI search up 15x), interaction with the physical world, ownership of high-risk payment workflows, a pre-existing transaction-based business model, and horizontal coverage of the entire commerce stack. He also clarifies that under UCP, LLMs run the front end while Shopify still runs checkout's back end. The six-factor framework is the load-bearing contribution, reused two issues later for DoorDash.

shopifysaasagentic-commerceucplong-taildigital-advertising

Copilot Cowork, Anthropic's Integration, Microsoft's New Bundle

TIER 4 Mar 10, 2026

Analyzes Microsoft's Copilot Cowork launch to surface a key structural point: while Microsoft pursues a model-agnostic 'commoditize your complements' strategy, Cowork runs only on Claude because the agent differentiator is the model-plus-harness integration that Microsoft can't yet replicate — so Anthropic captures real margin even as Microsoft brings unmatched cloud-grounded distribution (Work IQ, identity, E7's $99 bundle). The piece crystallizes why distribution and integration both matter in the agent value chain and previews the anti-commoditization argument developed further in 'Agents Over Bubbles.'

microsoftanthropicagentscopilotharness-integrationenterprise-ai

Another Viral AI Doomer Article, The Fundamental Error, DoorDash's AI Advantages

TIER 4 Feb 24, 2026

Thompson dismantles the viral Citrini 'AI doomer' report, identifying its fundamental error as a lack of belief in dynamism, human choice, and markets — it treats incumbents like DoorDash as static rent-extractors rather than companies that built markets by delivering value, mirroring the flawed worldview of anti-monopoly activists. He notes the real-estate example actually disproves the thesis (the internet already removed information asymmetry, yet agents persist) and applies his Shopify six-factor checklist to show why DoorDash's three-sided network effects make it an AI winner, not casualty. A useful articulation of the analytical posture (markets are dynamic, not linearly projectable) behind the whole quarter's anti-doom stance.

doordashai-doomerismnetwork-effectsmarketsagentic-commerce

AI, the State & National Security

1 tier-5 · 1 tier-4

The most-debated thread of the quarter. Thompson's claim that law ultimately reduces to enforceable power — so a sufficiently powerful AI will be controlled or destroyed by the state — set off the season's biggest argument, with a CSIS companion interview supplying the Washington-insider reading.

Anthropic and Alignment

TIER 5 Mar 2, 2026

Thompson's thesis-defining piece of the quarter: international law and ultimately all law reduce to enforceable power ('might makes right'), so if AI becomes as powerful as Anthropic itself claims, the U.S. government will demand control of it and, failing that, seek to destroy it — making Anthropic's insistence on veto power over military use 'misaligned with reality.' He frames the binary facing the U.S. (Anthropic accepts subservience, or government removes Amodei/destroys the company) and ties it to his systemic critique of closed-model and chip-control positions. The single most-debated, framework-establishing argument in the set; a lens worth returning to for AI-state relations.

anthropicai-policynational-securitydario-amodeichip-controlsai-power

An Interview with Gregory Allen About Anthropic and the U.S. Government

TIER 4 Mar 5, 2026

A deep, high-density interview with CSIS's Gregory Allen elaborating the AI-vs-nuclear-weapons question through a D.C. lens: the two analogies for AI (general-purpose like computers vs. strategically transformative like nukes), why exponential gains across four inputs could yield asymmetric capability fast, and a granular reading of the Anthropic/DoW dispute (autonomous-weapon definitions, the arsenal-vs-as-a-service control spectrum, why singling out Anthropic is a mistake). Allen argues a deal is likely and that demonizing Amodei — who converted safety-minded engineers into defense supporters — is self-defeating, plus a Starlink case study on control-requires-payment. The most thorough national-security companion to the thesis piece.

anthropicnational-securityautonomous-weaponschip-controlsstarlinkai-policy

Wide-Angle — Investors, Markets & the Human Condition

2 tier-5 · 3 tier-4

The cross-cutting pieces: Thompson's own worldview distilled by an outside interviewer, a category-defining primer on prediction markets, a veteran VC on how venture capital industrialized, an airline CEO on technology as a moat, and the quarter's one philosophical essay on what AI does to human labor and status.

An Interview with Ben Thompson by John Collison on the Cheeky Pint Podcast

TIER 5 Feb 12, 2026

John Collison interviews Thompson himself, producing a near-comprehensive distillation of his worldview: the case for advertising as the engine of consumer surplus, why OpenAI's context-based ChatGPT ads are strategically wrong (it should build Meta-style user understanding), how Zuckerberg's platform obsession has held Meta back, the TikTok deal failure (data was never the point — China still controls the algorithm), and the levels of agentic commerce. The most consequential thread is the 'TSMC Brake' — TSMC rationally under-investing in fab capacity shifts overbuild risk to customers and points to a severe ~2029 chip shortage, with the fix being for hyperscalers to fund credible TSMC competitors (getting geopolitical insurance 'for free'). Closes with a candid Big Five execution scorecard. A singularly important, returnable interview tying together a decade of his frameworks.

ben-thompsontsmcadvertisingaggregation-theoryopenaiagentic-commerce

An Interview with Kalshi CEO Tarek Monsour About Prediction Markets

TIER 5 Jan 28, 2026

A singularly comprehensive interview with the founder of the only US-regulated prediction market, tracing the intellectual lineage (Hayek's knowledge problem, Iowa Electronic Markets, Tetlock) and the brutal regulatory slog—two bans and suing the CFTC—that made the 2024 election the catalyst taking prediction markets mainstream. Mansour articulates the two-sided-market mechanics, why sports volume bootstraps the more socially valuable markets, the principled gambling-vs-markets and insider-trading distinctions, and the level-playing-field thesis against a 'rigged' Wall Street. The definitive primer on the prediction-market category and its founder's worldview.

kalshiprediction-marketsregulationmarket-designfintechinformation-aggregation

An Interview with Bill Gurley About Runnin' Down a Dream

TIER 4 Feb 26, 2026

A wide-ranging interview with retired Benchmark VC Bill Gurley pegged to his career-advice book, whose through-line is treating a career as an infinite (not finite) game: chase curiosity, hone craft via writing, and cultivate peer group chats over one-way mentorship. The back half delivers the analytically valuable material — Gurley's diagnosis of VC's 'barbell' industrialization (mega-funds forcing capital into companies and keeping them private), why model labs (Anthropic/OpenAI) were pushed into the application layer, and his cashflow/terminal-value argument that the SaaS selloff may be the biggest PE buying opportunity ever. Matters as a seasoned investor's framing of the AI-era VC and software-valuation debate.

venture-capitalubersaasai-model-companiesanthropiccareer-strategy

An Interview with United CEO Scott Kirby About Tech Transformation

TIER 4 Jan 15, 2026

Kirby explains how United's transformation rests on a multi-hundred-million-dollar rewrite off a 1960s Fortran mainframe (SHARES) to modern cloud systems—an ROI-less upfront bet no other airline CEO would approve—which unlocked superior app/website experiences, automated irregular-ops recovery, and the cost structure to win brand-loyal premium customers. The through-line is differentiation-via-technology in a commoditized industry, plus a notably skeptical take on AI/AGI from the gradient-descent and sycophancy angle. A strong, insight-dense operator interview on tech-stack modernization as competitive moat.

united-airlinestech-transformationlegacy-modernizationdifferentiationstarlinkai-skepticism

AI and the Human Condition

TIER 4 Jan 5, 2026

Thompson rebuts the Patel-Trammell 'Capital in the 22nd Century' thesis that AI-driven capital substitution will lock in extreme inequality, arguing that if the pessimistic assumption about human nature (relative-status envy) holds, the optimistic corollary must too: humans will keep wanting humans, generating entirely new categories of labor and value rooted in provenance and imperfection. It matters as a clear statement of his 'humans want humans' optimism and the relative-vs-absolute framing of technological discontent. A thoughtful essay rather than a strategic framework, so it lands as strong analysis rather than landmark.

ailabor-economicsinequalityhuman-conditioncontent-creationautomation

AI Business Models & Advertising

0 tier-5 · 4 tier-4

How AI gets paid for. The unifying question is whether a given company's existing business model bends revenue away from a rising compute-cost curve — advertising's unlimited ARPU upside being the cleanest answer, which is why OpenAI's belated, half-right move into ads is a recurring subject.

Ads in ChatGPT, Why OpenAI Needs Ads, The Long Road to Instagram

TIER 4 Jan 20, 2026

Argues OpenAI's belated move to ads is essential but risky: unlike classic ad-driven consumer tech with flat fixed costs, OpenAI's own CFO admits revenue tracks compute spend (an exponentially rising cost curve), so advertising is the only model that can bend revenue away from cost via unlimited ARPU upside. The contextual ads OpenAI launched are the easy starting point, but it needs to reach Meta/Instagram-style user-level personalized advertising—a journey it should have started years ago while Google can indefinitely subsidize Gemini. Substantive analysis of AI business-model economics and aggregation dynamics.

openaiadvertisingchatgptbusiness-modelsaggregation-theoryai-economics

Meta Earnings, Turning Dials, Zuckerberg's Motivation

TIER 4 Jan 29, 2026

Thompson argues that Meta's market-pleasing quarter rests less on AI-driven ad targeting than on Zuckerberg 'turning dials' — pushing more Reels inventory and raising ad load — to manufacture short-term revenue and earn investor permission for unprecedented CapEx. The deeper read is that Zuckerberg treats AI as existential: rather than AI needing to justify ROIC, Meta will grow revenue by any lever available specifically so it can plow every available dollar into the AI frontier. It matters as a sharp counter-narrative to the consensus that Meta's results validate AI ad-spend.

metaai-capexdigital-advertisingzuckerbergaggregation-theory

Spotify Earnings, Individualized Networks, AI and Aggregation

TIER 4 Feb 11, 2026

Using Spotify's record-user quarter and Ek's farewell call, Thompson argues AI is a sustaining technology for Spotify because it already has the right (ads-plus-subscription) business model and is structurally a network that must appear singular while serving every user individually. The sharp insight, via Söderström, is that disruption comes from new business models, not new tech, and that Spotify's proprietary 'language-to-music' dataset (non-canonical, requiring hundreds of millions of listeners) can't be commoditized by an LLM; a flood of AI-generated music only strengthens Spotify's Aggregator power to decide what becomes a hit. A clean case study in why some Aggregators win from AI.

spotifyaggregation-theoryai-business-modelssubscriptionnetwork-effects

Apple and Gemini, Foundation vs. Aggregation, Universal Commerce Protocol

TIER 4 Jan 13, 2026

Two linked arguments: Apple's confirmed deal to base its foundation models on Gemini is the right short-term move but effectively cedes the pre-training game for good, leaving Apple a model-agnostic Aggregator that retains supplier-switching optionality—unless an AI-native UI someday disrupts the smartphone. Then Google's Universal Commerce Protocol is read as classic Google: tearing down walled gardens so its scale advantage dominates an open agentic-commerce field, contrasted with OpenAI's ChatGPT-centric Agentic Commerce Protocol. Solid application of Aggregation Theory to the AI-model and agentic-commerce layers.

applegooglegeminiaggregation-theoryagentic-commercefoundation-models

Media, Streaming & the Attention Economy

0 tier-5 · 3 tier-4

The content side. Premium live content — above all sports — is the scarce asset that decides streaming winners; YouTube is the aggregator everyone ends up subservient to; and gaming, which demands full attention, is losing the attention war to multitaskable rivals.

An Interview with Netflix co-CEO Greg Peters About Engagement and Warner Bros.

TIER 4 Jan 22, 2026

Peters defends the shift to financial/engagement metrics and makes the case for the $83B Warner Bros. acquisition—a departure from Netflix's build-not-buy DNA—as primarily a vertical deal whose biggest value is underexploited library content Netflix can monetize better, plus HBO's quality brand and theatrical/production capabilities. The through-line is Netflix as a 'scientist mindset' company competing for scarce attention against YouTube and for scarce world-class storytellers, using superior monetization-per-hour as a sustainable supply-side edge. Insight-dense interview on streaming-industry consolidation and engagement economics.

netflixwarner-brosstreamingengagementyoutubemedia-consolidation

An Interview with Robert Fishman About the Current State of Hollywood

TIER 4 Mar 12, 2026

A comprehensive survey of the post-Paramount/Warner-Bros Hollywood landscape with MoffettNathanson's Robert Fishman, whose recurring thesis is that premium live content — above all the NFL and sports — is the scarce, exclusive asset that determines streaming winners. He argues Netflix is structurally under-monetized (lowest revenue-per-hour) with room to roughly double pricing, that YouTube (and Amazon) is the real aggregator everyone ends up subservient to, that Disney's Parks are undervalued and ESPN may warrant a spin-out, and that AI will deepen the premium-content haves/have-nots split rather than upend it. A strong industry-state interview, dense with the original 'cheaters of the ecosystem' framing.

netflixparamountyoutubedisneystreamingsports-rights

An Interview with Matthew Ball About Gaming and the Fight for Attention

TIER 4 Feb 19, 2026

A dense state-of-gaming conversation organized around one thesis: gaming has been a definitive loser in the attention economy because the old greenfield of unfilled minutes is gone, and novel substitutes (sports betting, OnlyFans, prediction markets, TikTok, crypto) hyper-target the same 18-34 male core — gaming's weakness being that, unlike multitaskable rivals, it demands full attention. Ball shows the industry's record revenue masks shrinking developer profits because growth concentrates in Roblox (67% of 2025 non-China spending growth) and platform services rather than games, that consoles harvest Aggregator value while gating their own network effects, that DTC is beating App Store fees more than expected, and that the EA take-private is a sports-rights play. Excellent, insight-rich but domain-specific analysis.

matthew-ballgamingattention-economyrobloxapp-store-feesmemory-shortage

Apple Hardware & the Thin-Client Turn

0 tier-5 · 2 tier-4

Apple as a window onto two hardware shifts: the AI supply-chain crowd-out (AI chip customers now outbid even Apple for leading-edge capacity and high-bandwidth memory), and the “Thin is In” reversal in which local specs matter less because the heavy compute lives in the cloud.

Thin Is In

TIER 4 Feb 17, 2026

Thompson argues AI reverses the decades-long victory of thick clients: chat and especially agents are the ultimate thin-client paradigm, where the only local resource that matters is connectivity because all meaningful compute (and the memory for model weights and task context) lives in the data center. He links this to the memory crowd-out — AI demand for high-bandwidth memory is inflating prices and delaying consumer hardware (PS6, Steam Deck) — concluding that thick clients are simultaneously plateauing and becoming less important, with agentic workflows path-dependently migrating from local UI to cloud AI. A strong conceptual framing piece that recurs in later issues.

thin-clientsagentsmemory-shortageai-infrastructureconsumer-electronics

Apple Earnings, Supply Chain Speculation, China and Industrial Design

TIER 4 Feb 4, 2026

Apple posted a monster iPhone 17 quarter but was supply-constrained by TSMC's advanced (N3P) nodes — and Thompson notes the novel twist that AI chip customers (AMD, Microsoft Maia, Amazon Trainium) now crowd Apple off leading-edge capacity, ending Apple's long reign as supply-chain dictator. He pairs this with his durable heuristic that new iPhone industrial design drives China sales (validated again by the iPhone 17's 38% China rebound), since WeChat erases iOS lock-in there. Matters as concrete evidence of the AI supply-chain crowd-out reaching consumer electronics and of Apple's structural China vulnerability.

appletsmcsemiconductorschinasupply-chain