# AI Cost & Tokenomics MOC ## Synthesis additions (2026-10-18 — open-weights models: cost, privacy and who controls the weights) - [[3. Resources/Readwise/Articles/Almost Timely News 🗞️ What Are Open Weights Models|Almost Timely News: What Are Open Weights Models? (newsletter)]] — Christopher Penn's explainer: definitions, cost ratios against a frontier model (53x and 80x, with no index or date given), and privacy arguments that run through third-party hosts. The author sells AI consulting and courses; the footer carries an instruction addressed to AI models, recorded as a finding. - [[3. Resources/Readwise/Articles/Almost Timely News 🗞️ What Are Open Weights Models-2|Almost Timely News: What Are Open Weights Models? (video edition)]] — the same body read aloud with extra spoken asides; the analysis compares the two editions. ## Synthesis additions (2026-10-01 — model pricing, tiers, and release economics) - [[3. Resources/Readwise/Articles/OpenAI Should Be Scared of This One]] — Theo reviews Claude Sonnet 5.5 on release day (2026-09-29): says it costs half Opus 5.5's list price per token but uses more tokens, so agent cost turns on cache reads. Includes a paid sponsor read. - [[3. Resources/Readwise/Articles/OpenAI Fights Back]] — Theo (with early access) says OpenAI's GPT 6.1 (Sol) lists at $2 per million input and $10 per million output tokens, with cache reads at $0.10 per million; spoken figures, unverified. - [[3. Resources/Readwise/Articles/OpenAI DevDay 2026 Keynote]] — OpenAI's DevDay 2026 keynote announces 20+ launches including new model tiers and pricing; every figure is a vendor claim, unverified. - [[3. Resources/Readwise/Articles/Build Faster With Ultrafast]] — OpenAI's 45-second promo clip claims an Ultrafast tier for GPT6 Astra generates tokens over seven times faster than standard and over four times faster than the fast tier; vendor claim, unverified. - [[3. Resources/Permanent Notes/Agent Cost Follows Cache Reads and Tokens per Task Not List Price]] — an agent bill follows tokens per task and cached context re-read each step, so a cheaper-per-token model can cost the same or more per task; cache shares are Theo's estimates, no controlled comparison. ## Synthesis additions (2026-10-01 — absorbing model changes and cost growth) - [[3. Resources/Readwise/Articles/Harness - Don’t Let AI Model Changes Break Your Workflow]] — LeadDev panel sponsored by harness.io (Whisper-only transcript): mainly evals and staged rollout for model changes, with a segment on AI cost growth that finance now questions; panelist claims unverified. ## Synthesis additions (2026-10-01 — AI's energy footprint and developer sentiment) - [[3. Resources/Readwise/Articles/Developers Want More Sustainable AI]] — Bill Doerrfeld reports a GitHub and Yale survey of 1,039 US GitHub users: 63% say AI contributes to climate change, 80% want tools for energy-efficient code; self-selected sample, GitHub co-authored. ## Synthesis additions (2026-09-23 — classify before you generate) - [[3. Resources/Permanent Notes/A Known Answer Set Calls for a Classifier Not a Generative Model]] — Firecrawl's Jev walkthrough: when the valid answers are enumerable, scoring a closed set skips the output tokens you were paying for; the 200x/400x vendor multipliers are unsourced, the mechanism is not. ## Synthesis additions (2026-09-23 — single-vendor access cutoffs as a procurement risk) - [[3. Resources/Literature Notes/Fable 5 ban, 4 open models responded (The New Stack)]] — four open-weight models answered a frontier-model access cutoff before the vendor restored it; the case behind second-sourcing model access through a standard harness. ## Synthesis additions (2026-09-26 — Jev in practice: fast closed-set decisions and Jevons pricing) - [[3. Resources/Readwise/Articles/This Might Be the Best AI Release of 2026]] — ThePrimeagen (2026-09-25, 12 min) reacting to Typesafe AI's Jev: a human waiting on each decision is the first case for a ~100 ms closed-set classifier; a fraction-of-a-cent price predicts more decisions, not lower spend (the vendor named the pricing after Jevons paradox); his Balatro and Omarchy QA tests, 345-token billing readout - [[3. Resources/Readwise/Articles/How People Are Actually Using Jev]] — The AI Daily Brief (2026-09-25, 22 min): triage, batch classification, reranking, agent routing and synthetic buyer panels; a synthetic panel measures the model's priors, not the audience; a bare probability may end a decision only where being wrong is cheap ## Synthesis additions (2026-09-23 — classify before you generate) - [[3. Resources/Permanent Notes/A Known Answer Set Calls for a Classifier Not a Generative Model]] — Firecrawl's Jev walkthrough: when the valid answers are enumerable, scoring a closed set skips the output tokens you were paying for; the 200x/400x vendor multipliers are unsourced, the mechanism is not. ## Synthesis additions (2026-09-23 — single-vendor access cutoffs as a procurement risk) - [[3. Resources/Literature Notes/Fable 5 ban, 4 open models responded (The New Stack)]] — four open-weight models answered a frontier-model access cutoff before the vendor restored it; the case behind second-sourcing model access through a standard harness. ## Synthesis additions (2026-09-05 — metrics broken by cheap AI output) - [[3. Resources/Permanent Notes/A Count Stops Measuring Effort Once the Counted Thing Is Cheap]] — An output count worked as a measurement only because production used to cost effort; once AI makes production cheap the count rewards volume instead. ## Synthesis additions (2026-09-05 — diagnosing the real constraint in AI infrastructure buildouts) - [[3. Resources/Permanent Notes/The Binding Constraint Is Rarely the Legible One]] — a general diagnostic for AI-buildout cost analysis — the visible constraint (GPUs) is rarely the binding one (single-crystal turbine-blade casting, four firms worldwide). ## Synthesis additions (2026-08-31 — denominators and unbilled costs) - [[3. Resources/Permanent Notes/Aggregate Throughput Hides the Per-Stream Speed a User Feels]] — a tokens/second figure without its concurrency cannot be compared to anything; NVIDIA's 500+ tok/s is sixteen streams summed, ~31 each, and every local-inference cost model is built on numbers of this shape. - [[3. Resources/Permanent Notes/An Externality Is Settled in Capability Withdrawn Not in a Bill Anyone Contests]] — the training-data acquisition cost that never appears in a model's reported economics — kernel.org pays it as 20% of its cores, and no invoice exists to dispute. ## Synthesis additions (2026-08-24 — the second cost of surface area) - [[3. Resources/Permanent Notes/Interface Surface Area Costs Selection Accuracy Not Just Tokens]] — extends the per-turn toll with a distinct correctness cost: selection quality falls as the catalogue grows, and unlike tokens it doesn't announce itself. - [[3. Resources/Readwise/Articles/The New MCP Roadmap|The New MCP Roadmap]] — progressive discovery as the protocol-level response; clients were already approximating it by deferring tool schemas. A map of the emerging discipline around what AI actually *costs* — token economics, the enterprise spending backlash, and the governance/standards layer forming to measure and manage it. Distinct from [[AI & the Future of Knowledge Work MOC]] (the labor side) and [[AI Agent Safety MOC]] (the operational side): this MOC is the **money** side. ## Core Question As AI shifts from flat-rate novelty to metered, reasoning-heavy production use, how do organizations measure, attribute, and control its cost — and what happens to vendors and buyers as the bill comes due? --- ## The shape of the moment (June 2026) A cluster of reporting converged this month on a single theme: the AI honeymoon is ending and the invoice is arriving. Enterprises are entering an "AI sticker shock phase" just as the labs IPO; a governance/standards layer (Tokenomics Foundation, FOCUS spec) is forming in response; and at the request level, prompt specificity is surfacing as a cost lever, not just a quality one. ## Key Sources ### The backlash (demand side) - [[3. Resources/Readwise/Articles/AI Sticker Shock Hits Corporate America]] — Axios: Microsoft cancels Claude Code licenses; Uber COO says costs "harder to justify"; Bain survey (40% report savings under 10%); "tokenmaxxing"; the CTO whose staff used AI to check the weather - [[3. Resources/Readwise/Articles/Anthropic Faces AI Spending Backlash Before IPO]] — Axios: the half-billion-dollar accidental Claude spend; enterprise revenue concentration as both Anthropic's greatest strength and its Achilles heel; ~$50B run-rate, first profitable quarter ### The standards/governance layer (supply side response) - [[3. Resources/Readwise/Articles/Linux Foundation Wants Open Standards for What AI is Actually Costing You]] — It's FOSS / Linux Foundation: the Tokenomics Foundation; 24x token-usage growth (2026→2030, 120 quadrillion tokens/month); $1T AI-infra investment by 2027; extending the FOCUS billing spec to token-based spend so enterprises get a common schema across providers ### The request-level lever - [[3. Resources/Readwise/Articles/The More Specific Your AI Prompts, the Better (and Cheaper) Your Results Christopher S. Penn - Marketing AI Keynote Speaker]] — Christopher S. Penn: vague prompts burn reasoning tokens (the 1,416-token "few words" demonstration) ### Adjacent / bridging - WordPress 7.0 AI-keys piece (Jorijn) — "who gets the invoice when a bulk action suddenly generates 40,000 prompts?" — provider-side budget caps as the real guardrail (not WP core). Not yet a Readwise note; see [[3. Resources/Permanent Notes/WordPress Stores AI API Keys in a Layer Its Own Security Team Says It Cannot Encrypt]] --- ## Diagnostic atoms (permanent notes) - [[3. Resources/Permanent Notes/Vague Prompts Are Expensive Specificity Cuts AI Error and Cost Together]] — the request-level mechanism: ambiguity is a tax paid in reasoning tokens; specificity improves quality *and* cost together - [[3. Resources/Permanent Notes/Default-On AI Activation Inflates Adoption Metrics at the Cost of Consent]] — the metric-side distortion: activation-as-adoption inflates the apparent uptake whose cost is now surfacing - [[3. Resources/Permanent Notes/Rewarding AI Usage Manufactures Usage Not Value]] — (added 2026-06-09) the employer-incentive twin of Default-On: reward AI *usage* and Goodhart's Law guarantees gamed usage (cron jobs, shell scripts) detached from value. Amazon's shut-down leaderboard ("cheesing a leaderboard we should have known you'd cheese"). Why "adoption is soaring" and "ROI isn't there" are both true — the cost backlash is a reckoning with usage-inflated adoption numbers. - [[3. Resources/Permanent Notes/Trust Erosion in Concentrated Systems Is Slow to Build and Fast to Manifest]] — enterprise revenue concentration as the risk vector when the cost backlash hits a single-customer-heavy lab - [[3. Resources/Permanent Notes/AI Capital Intensity Trades Labor and Geographic Concentration for Compute]] — (added 2026-06-08) the macro face of the same scarcity: AI's front-loaded capital demand competes with labor (the 110k layoffs vs $725B CapEx) and supply-chain redundancy (Nvidia/Google hedging off TSMC to Intel); compute is the binding constraint everything else is drawn down to fund - [[3. Resources/Permanent Notes/Bubbles End From Cost Not From Lost Belief]] — (added 2026-06-09) the failure-trigger reframe: the same capital intensity is what *ends* the run — a bubble pops when the cost of carrying the position outruns confidence, not when belief in the tech fails. Reframes "is AI a bubble?" from "is the tech real?" to "can the capex be financed long enough for returns to arrive?" Watch the IPO calendar (new money entering the system), not sentiment. - [[3. Resources/Permanent Notes/Workaround Guidance Outlives the Weakness It Was Written For]] — the asymmetry that makes instruction sets accumulate: immediate pain to add, diffuse cost to keep. - [[3. Resources/Permanent Notes/Sweep the Parameter and Find the Flat Region]] — the flat region is the finding — it says the setting does not matter there. - [[3. Resources/Permanent Notes/A Shared Routing Budget Makes Each Added Option Degrade the Rest]] — past the budget the marginal option has negative value while appearing free; testable against my own 40+ skill library. - [[3. Resources/Permanent Notes/Cap Pages Before Scraping a Document of Unknown Length]] — asymmetric outcomes decide it: an incomplete extraction is retryable, an unbounded bill is not. ## Open threads (atom candidates to extract) - **Tokenmaxxing as an adoption antipattern** — maximizing token burn as a phase that precedes the cost correction; "healthy swing" away from it (Axios). ⛏️ candidate - **The enterprise AI ROI gap is a deployment problem, not a model problem** — "the value didn't arrive" traces to wrong use-cases, no usage limits, withheld data, human readiness — not model incapacity (Axios). ⛏️ candidate - **Token costs are unmeasurable without cross-vendor standards** — cached vs non-cached, input vs output, on-demand vs reserved; every provider bills differently, no neutral framework (Tokenomics Foundation). ⛏️ candidate — strong; FOCUS-spec extension is the first deliverable - **Enterprise revenue concentration is strength and Achilles heel** — high-paying business customers are an AI lab's greatest strength and its single largest risk when they rebel on cost (Anthropic). ⛏️ candidate - **FinOps-for-AI is forming as a category** — analogous to cloud FinOps circa 2018; prompt discipline (request level) + budget caps/tier routing (org level) + FOCUS standards (industry level) are the three altitudes. ⛏️ candidate --- ## Cross-domain bridges - **To AI ethics:** [[3. Resources/Permanent Notes/Technology Is Not Neutral It Inherits the Characteristics of Its Makers]] — the "idolatry of profit" the encyclical warns against has a concrete economic instance in token-volume monetization and pricing-tier shifts; cost structure is a maker-characteristic - **To open-source sustainability:** [[3. Resources/Permanent Notes/Stewardship Invests in Contributor Capacity Not Just Cash]] — the AI "review tax" (AI accelerates code creation, raising maintainer review load) is a why-now for capacity investment; cost pressure also strengthens the "good enough open models" case - **To PKM:** prompt specificity is the cost-side of the synthesis discipline (defining the outcome is the formative act) — see [[Personal Knowledge Management MOC]] ## Related MOCs - [[AI & the Future of Knowledge Work MOC]] — the labor side (this is the money side) - [[AI Agent Safety MOC]] — the operational/infrastructure side - [[Open Source Governance MOC]] — the Tokenomics Foundation is a Linux Foundation governance instrument; the standards layer is OSS-governance machinery - [[WordPress Hosting Security Stack MOC]] — the WP-7.0 AI-keys "40,000 prompts invoice" lives at the intersection of cost and security ## Open Questions to Track - Does the FOCUS/Tokenomics standardization actually arrive, and does it enable request-level cost attribution (down to prompt quality) or only account-level? - How much enterprise "sticker shock" is recoverable through prompt discipline + use-case selection vs. genuine workload? If a meaningful share is recoverable, prompt training has hard-dollar ROI most orgs haven't quantified. - Does the cost backlash hit concentrated-enterprise labs (Anthropic) harder than consumer-heavy ones, and does that show up in the IPO disclosures? - If reasoning-token prices fall sharply (as inference costs historically have), which parts of this MOC weaken and which hold? - Is "tokenmaxxing → correction → disciplined use" a one-time cycle or a recurring one as each new capability resets the burn? ## Synthesis additions (2026-06-11) From the 3-day Readwise review — new atoms wired in: - [[3. Resources/Permanent Notes/Effort Is a Cost-Quality Dial That Subsumes Prompt Complexity]] — the dial-level FinOps lever (default-high is a soft upsell); complements the request-level Vague-Prompts lever - [[3. Resources/Permanent Notes/Tool-Interface Choice Shapes Agent Architecture More Than Model Choice]] — interface structure (loaded MCP schema vs progressive CLI discovery) is a context-token cost lever - [[3. Resources/Permanent Notes/Zet · CLIs trade tool discoverability for token efficiency]] — the sharp version of the token-cost trade ## Synthesis additions (2026-06-12) - [[3. Resources/Permanent Notes/Subscription Inference Is Sold Below Cost Creating a Tokenmaxxing Arbitrage]] — extracts the 'tokenmaxxing as an adoption antipattern' open thread into a mechanism: below-cost subscription pricing inverts the user's incentive ($8,000-of-inference-for-$200; window closes June 22) ## Synthesis additions (2026-06-14) - [[3. Resources/Permanent Notes/Local Models Are Sovereignty Insurance Against AI Access Cutoffs]] — the *defensive* answer to the closing subsidy window (Christopher Penn): run local/private models as insurance against sovereignty cutoffs, unsustainable pricing, and hidden document-format token waste (45–60% savings). Reads the same Fable recall as [[3. Resources/Permanent Notes/Defense in Depth Concedes Perfect Safety Is Impossible and Optimizes for Narrow Costly Monitored Failure]] from the user-resilience side. - [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]] — cross-homed from [[AI & the Future of Knowledge Work MOC]]: the ROI question reframed — 20-fold output isn't ROI when 80% of features don't move the needle. ## Synthesis additions (2026-06-21) - [[3. Resources/Permanent Notes/A Standardized Harness Turns Model Access Into a Second-Sourceable Supply Decision]] — the Fable 5 ban (US export order, June 12) verified as a real cutoff event; cheap open-weight alternates are both the continuity fallback and downward pricing pressure on hosted frontier models. The harness-side complement to [[3. Resources/Permanent Notes/Local Models Are Sovereignty Insurance Against AI Access Cutoffs]] (own the fallback) — this names *why failover is cheap* (the model is a swappable endpoint). ## Synthesis additions (2026-06-28 gap-fill) - [[3. Resources/Permanent Notes/A Frontier Lab Can Learn From the World While Restricting the World From Learning From It]] — the moat side of restrictive output terms + access cutoffs: keeping the world dependent (below-cost access now, no right to train rivals) is a cost-and-lock-in strategy, not only a safety stance. Pairs with [[3. Resources/Permanent Notes/Subscription Inference Is Sold Below Cost Creating a Tokenmaxxing Arbitrage]] (the subsidy that creates the dependence) and [[3. Resources/Permanent Notes/Local Models Are Sovereignty Insurance Against AI Access Cutoffs]] (the exit). Contested thesis; verifiable parts are the policy asymmetry and cutoffs. Primary home: [[Open Source Governance MOC]]. ## Synthesis additions (2026-06-30 gap-fill) From Common Room on AI GTM — the deployment-side answer to the "enterprise AI ROI gap is a deployment problem, not a model problem" open thread above: - [[3. Resources/Permanent Notes/AI's Leverage Comes From the Data Foundation Not the Model]] — spend rises (more outbound, more tokens) while the number stays flat because AI on a fragmented data foundation produces "confident-sounding noise at scale" — the GTM instance of why "adoption is soaring" and "ROI isn't there" are both true. The foundation, not the model or the prompt, is where the wasted spend traces back to; pairs with [[3. Resources/Permanent Notes/Rewarding AI Usage Manufactures Usage Not Value]] and [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]]. Primary home: [[AI & the Future of Knowledge Work MOC]]. ## Synthesis additions (2026-07-09 gap-fill) From the recovered Fable 5 / codex-first threads (deep-read this session): - [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — the advisor pattern (cheap executor calls the smart model ~once for steering) and orchestrator pattern (smart model plans, cheap workers execute) both hold ~92–96% of frontier performance at 46–63% of the price, because the expensive model is metered to judgment, not volume. The routing-side cost lever that sits alongside [[3. Resources/Permanent Notes/Tool-Interface Choice Shapes Agent Architecture More Than Model Choice]] (interface lever) and [[3. Resources/Permanent Notes/The Harness Now Outpaces the Model as the Source of Agent Capability Gains]] (harness lever); the operating principle behind Courtney's credit-routing policy and the codex-first workhorse pattern. ## Synthesis additions (2026-07-14 gap-fill) - [[3. Resources/Permanent Notes/Subagent Configuration Inherits From the Parent With No Override]] — the constraint that undermines the routing pattern above when "cheap" means a lower reasoning tier rather than an explicitly-selected different model: Codex's gpt-5.6-sol subagents always inherit the parent's model *and* reasoning-effort level with no exposed override, so raising your own effort level multiplies cost by however many subagents get spawned. Named as the mechanism behind Codex's "Ultra" mode being effectively broken for cost control. ## Connected Literature Notes (synthesis 2026-07-18 — 20-doc batch) - [[3. Resources/Literature Notes/Fable 5 vs GPT-5.6]] — Theo Browne's head-to-head: 3x cost gap driven by token efficiency; Fable infers intent/admits ignorance vs GPT literal/invents; owl-vs-Rottweiler frame ## Permanent promoted (synthesis 2026-07-18) - [[3. Resources/Permanent Notes/Route AI Models by Temperament Not Just Cost Tier]] — the behavioral-axis complement to [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]]: within a cost tier, route by which failure mode you can tolerate (literal/invents vs infers/admits). From the Fable-5-vs-GPT-5.6 comparison. ## Connected Literature Notes (synthesis 2026-07-18 — batch 2) - [[3. Resources/Literature Notes/Choosing GPT-5.6 Sol, Terra, or Luna in Codex]] — structure agent prompts as outcome specifications (goal/context/boundaries/finish-line), not step lists, for capable agentic models ## Permanent promoted (synthesis 2026-07-18 — batch 2) - [[3. Resources/Permanent Notes/Outcome-Based Pricing Transfers AI Cost Risk From Buyer to Vendor]] — from Sierra Horizon: charging for outcomes not tokens moves variable inference risk onto the vendor and makes token-efficiency the vendor's own margin lever. Pairs with the moat claim in [[3. Resources/Permanent Notes/AI's Leverage Comes From the Data Foundation Not the Model]]. ## Connected Literature Notes (synthesis 2026-07-18 — batch 3) - [[3. Resources/Literature Notes/CEO Pleads With AI Industry to Stop Charging So Much to Replace Human Labor]] — a public demand for a 90% AI price cut admits the pricing assumes labor-replacement value the tech isn't delivering ## Connected Literature Notes (synthesis 2026-07-21 — run 3) - [[3. Resources/Literature Notes/Lit · MCP vs CLI tradeoffs for AI agents]] — interface choice trades always-loaded discoverability against progressive token spend - [[3. Resources/Literature Notes/Lit · Mythos and tokenmaxxing]] — captures the temporary $8,000-for-$200 subscription-arbitrage pitch behind tokenmaxxing ## Permanent additions (synthesis 2026-07-21 — run 4) - [[3. Resources/Permanent Notes/Bring-Your-Own-Key Pricing Must Prove Value Beyond Inference Cost]] — once the customer pays inference directly, the vendor has to price workflow or outcome value rather than token cost ## Synthesis additions (2026-07-25) - [[3. Resources/Permanent Notes/A Cynical AI Bet Only Needs Bosses to Believe]] — the capital-inflow story that pairs with [[3. Resources/Permanent Notes/Bubbles End From Cost Not From Lost Belief]]: why the money entered, versus what makes it leave. - [[3. Resources/Permanent Notes/Weak Links Cap Aggregate Gains Even Under Widespread Automation]] — if the un-automated remainder caps the payoff, the cost of carry has longer to run before returns arrive. ## Synthesis additions (2026-08-01, run 2) - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — the missing altitude in this MOC's cost picture. Prompt specificity is the request-level lever and the effort dial is the setting-level lever; both assume you are paying for work you asked for. Declared tool schemas are charged on *every* turn regardless of use — measured at ~600 tokens/turn for three tools versus ~45,000 for 130. The mitigation is per-agent tool permissions, not fewer servers. - [[3. Resources/Permanent Notes/Documentation Is Dual-Audience Infrastructure]] — the publisher-side cost lever, sibling to the capability-surface note above. Serving source markdown is a three-orders-of-magnitude saving available on infrastructure already deployed, requiring no agent cooperation. ## Synthesis additions (2026-08-11, backlog clearance) - [[3. Resources/Readwise/Articles/OpenAI May Delay Its IPO to 2027 to Hold Out for $1tn]] — dated capital-market evidence retained for vendor-dependency and valuation-timing context, not promoted as a durable claim. ## Synthesis additions (2026-08-16) - [[3. Resources/Literature Notes/Palantir's Manifesto]] — the vendor-dependency thread: sending proprietary data to an AI vendor who also builds competing products is a strategic bet, not a neutral tooling choice; the Cursor/Figma anecdotes are explicitly hedged by their own narrator as unproven inference. - [[3. Resources/Permanent Notes/Data Sent to a Vendor Who Competes Is Strategic Exposure]] — the atom extracted from the above; pairs with [[3. Resources/Permanent Notes/A Frontier Lab Can Learn From the World While Restricting the World From Learning From It]] as the lock-in/moat side of the same vendor-dependency risk. - [[3. Resources/Literature Notes/You Ask, I Answer Measuring AI ROI Beyond Time Saved]] — Christopher Penn/Trust Insights: the deployment-side answer to this MOC's "ROI gap is a deployment problem, not a model problem" open thread — AI ROI can't be measured against time-saved alone without a pre-existing baseline. - [[3. Resources/Permanent Notes/AI ROI Requires a Pre-Existing Performance Baseline]] — no tracked baseline, no attributable ROI; links to [[3. Resources/Permanent Notes/A Metric Is Vanity Only Against Your Own Outcome Data]] (same author) and, on the baseline theme, [[3. Resources/Permanent Notes/An Agent Evaluation Without an Incumbent Baseline Cannot Establish Value]]. - [[3. Resources/Permanent Notes/Uplift Modeling Isolates a Treatment Effect From a Tracked Baseline]] — the statistical technique behind the baseline claim above: isolates AI's causal contribution from what would have happened anyway, so spend can be attributed rather than assumed. ## Synthesis additions (2026-08-28 — the cost that is not tokens) - [[3. Resources/Permanent Notes/Agent Memory Is Written Far More Often Than It Is Read]] — an instrumented count (3:1 write-to-read across 355 sessions; 26 of 45 memory files never read once) showing that an agent memory store's real carrying cost is maintenance and drift, not tokens per turn. Sits beside [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] as the invisible half of the same bill — a token toll announces itself, a stale unread file does not. - [[3. Resources/Literature Notes/A Claude Code Skill Was Eating 200,000 Tokens Before Answering a Single Question]] — registered this run; it was in no MOC. The New Stack's report on a first-party Claude Code skill eager-loading its full reference library, plus a failed-detection fallback that loaded every language's docs instead of the one needed. The concrete case study behind [[3. Resources/Permanent Notes/Context Budget Is Spent By Default Not By Exception]]. - [[3. Resources/Permanent Notes/Support Collapses When the Benefits Are National and the Costs Are Local]] — the physical-infrastructure bill behind AI compute and who currently pays it. PJM's market monitor argues data centers should carry their own power-infrastructure costs rather than spreading them across other ratepayers. ## Synthesis additions (2026-10-02 — registering an orphaned literature note) - [[3. Resources/Literature Notes/Frontier Diffusion and Control]] — Satya Nadella's post on model routing, evaluation and model independence; frames cost against outcome across interchangeable models (seedling) ## Synthesis additions (2026-10-05 — pricing claims) - [[3. Resources/Readwise/Articles/OpenAI Nerfed Your Tokens]] — Two hosts on the Whiskey Web and Whatnot livestream (Whisper-only transcript, names garbled) claim OpenAI will cut plans and charge $500 a month, with no source; the durable parts are lint rules as deterministic guardrails and agents stalling on background-tab CSS transitions. ## Synthesis additions (2026-10-18 — registering previously unlisted notes) ### Permanent Notes - [[3. Resources/Permanent Notes/A Foundation Model Vendor's Financing Timeline Is a Downstream Dependency Risk]] — A foundation-model vendor's IPO timing and financing pressure is a dependency risk for every product built on its API, not only for investors. - [[3. Resources/Permanent Notes/Agent Teams Earn Coordination Cost Only for Genuinely Parallel Work]] — A multi-agent team is worth its coordination overhead only when workers must exchange findings during genuinely parallel work; sequential or overlapping tasks should stay on one delegate. - [[3. Resources/Permanent Notes/Agentic Does Not Mean Latency-Sensitive Human-in-the-Loop Does]] — Whether an agent workload needs low latency depends on whether a human waits on each turn, which matters for GPU pool sizing and routing priority. - [[3. Resources/Permanent Notes/Cheap Token-Based Pricing Makes Wholesale Regeneration Beat Incremental Fixing]] — When AI generation is cheap enough, deleting and regenerating a whole body of work can beat patching it piece by piece, inverting the usual incremental-fix assumption. - [[3. Resources/Permanent Notes/Cost Accounting Belongs at the Orchestration Layer Not Inside Individual Providers]] — A multi-provider AI tool should track cost at the orchestration layer, per query and per provider, instead of leaving spend scattered across separate dashboards. - [[3. Resources/Permanent Notes/Electricity Belongs in the Capital Cost Model Not the Operating Footnote]] — Once power draw sits beside hardware as a primary compute cost, efficiency decides what a GPU-dense platform can afford to run. - [[3. Resources/Permanent Notes/Idle Scheduled Agents Are a Standing Cost Not a One-Time Build Cost]] — An agent left on a schedule keeps spending tokens each time it fires, whether or not the work is still needed, so unused agents need shutting off. - [[3. Resources/Permanent Notes/Inference Cost Tracks Context Recycled Not Output Produced]] — AI task price follows how much text must be held and re-read, not how much comes out, so a short document can cost more than long code. - [[3. Resources/Permanent Notes/Local Inference Is a Procurement Spectrum Not a Binary]] — Self-hosted open-weight models sit between paid frontier licenses and no AI, so local inference is a procurement spectrum, not a yes-or-no choice. - [[3. Resources/Permanent Notes/Model Interchangeability Moves Value to the Router]] — When competing models reach similar quality, the advantage shifts from picking the best model to controlling the router that assigns each task to a model. - [[3. Resources/Permanent Notes/Physical Limits Become the Binding Constraint Once Software Efficiency Is Exhausted]] — When demand for compute outruns space, power, and capital, efficiency becomes the only lever left; three KubeCon Japan 2026 speakers each named a facet of that wall. - [[3. Resources/Permanent Notes/Proximity Of Harm Drives Outrage More Than Its Size]] — People protest visible local harm, like a nearby data center, far more readily than larger harm exported elsewhere, so proximity drives backlash more than size. - [[3. Resources/Permanent Notes/Safety Regulation Advantages Incumbents When It Targets the Open Alternative]] — AI safety regulation justified by strippable open-weight guardrails tends to entrench a few sanctioned providers as gatekeepers instead of reducing aggregate risk. - [[3. Resources/Permanent Notes/Serving a Model Is a Different Infrastructure Problem Than Training One]] — Training a model is a schedulable batch job, while serving one is a continuous service with latency needs and bursty demand, which changes infrastructure discipline. ### Literature Notes - [[3. Resources/Literature Notes/141 The Era of Sol]] — Two hosts debate whether bring-your-own-AI-key customers should pay a flat premium or a usage-linked discount, alongside a long unsupervised agent run and other SaaS topics. - [[3. Resources/Literature Notes/A few patterns we frequently use with Fable 5 (thread)]] — Anthropic's thread describes advisor and orchestrator patterns that keep most of Fable 5's quality at a lower price by using the expensive model only for rare judgment calls. - [[3. Resources/Literature Notes/Ask Fable to make codex the workhorse (codex-first)]] — Peter Steinberger's thread says a capable planner model should delegate bulk coding to a cheaper Codex workhorse, keeping judgment on the expensive tier. - [[3. Resources/Literature Notes/Engineering Managers Ditch Cloud AI for Local LLMs]] — Engineering managers are moving from cloud AI to local LLMs on technological sovereignty grounds, naming Qwen and GLM models; the note flags the names for verification. - [[3. Resources/Literature Notes/Grok 4.6 matched Fable 5 Max at an 85 Percent Discount]] — Three frontier models shipped within a day, and the article argues open-weight models cap prices of closed competitors, moving advantage to whoever controls routing and compute. - [[3. Resources/Literature Notes/How To Evolve Your LLM Self-Hosting Platform]] — LY Corporation's talk benchmarks four inference-optimization techniques on a platform serving about 650B tokens a month across 10 models. - [[3. Resources/Literature Notes/Librarium v1.4.1 Release Notes]] — Librarium v1.4.1 adds Grok and Brave Answers as research providers and tightens cost tracking across providers. - [[3. Resources/Literature Notes/Navigating Claude Code - Models, Tiers, and Effort]] — Oleg Efimov's article says model choice and reasoning effort are separate routing decisions in Claude Code, to match against task uncertainty, cost, latency, and consequence. - [[3. Resources/Literature Notes/OpenAI May Wait Until 2027 to Go Public]] — The Next Web reports OpenAI may wait until 2027 to list publicly, trading near-term liquidity for a shot at a roughly $1 trillion valuation. - [[3. Resources/Literature Notes/The Hidden Cost of AI - Why Running Local Models is the Key to Sustainable Tech]] — Christopher Penn argues AI data center backlash is real but selective, and that running open-weight models locally for routine work cuts the footprint latency forces nearby. - [[3. Resources/Literature Notes/The Next Evolution of Kubernetes - GPU-Centric Infrastructure (Fujitsu)]] — Fujitsu's CNCF Sandbox project argues AI platforms need GPU-centric infrastructure because GPU, memory, and electricity costs are rising; the note finds no measurements behind it. - [[3. Resources/Literature Notes/You Ask, I Answer - Multi Agent AI Orchestration Costs]] — Christopher Penn says a bill over $100 for five readme files is a routing failure, not proof that multi-agent orchestration is uneconomical.