The reading library

The frame the whole Money part runs on: AI as a chain of required returns, each layer having to earn enough to justify the one beneath it. Agrawal's two-year update is the same frame with numbers attached, so read them together. Both are written by investors in what they measure, and this one is pseudonymous, so position cannot be checked at all.

Superseded in magnitude, not in logic. Cahn asked in 2024 where the revenue was that justified the capex, and sized the hole at $600bn. Now it is about $443bn of hyperscaler capex against roughly $435bn of annualised AI revenue — which stacks the same end-user dollar up to three times. Keep the frame, replace the arithmetic. Sequoia is an investor in the buildout it questions.

The library's sharpest live disagreement. Nishball says the profit pool is migrating from infrastructure to the labs. Agrawal (Altimeter) puts the app-and-lab layer at 7% of gross profit (assumed margins; revenue base double-counts — see Agrawal); Srivastava (Baseten) puts 90-95% of inference spend at frontier labs. Value moving to labs is not value moving to applications, which the 7% bucket hides.

A live model-market equilibrium sketch: frontier labs keep a premium where capability really matters, but open-weight and Chinese models pressure routine enterprise work toward cheap-enough substitutes. The strategic response is full-stack ownership across energy, chips, cloud, models, applications, and the customer relationship.

“The future therefore splits between massive full stack labs and individual businesses which have to specialise in some layers.”

Three figures, three quantities. Du's ~600x is a market-wide index, 2020-2026, and the only one documenting its basket: 62 milestone observations, OpenRouter plus Epoch AI. Agrawal (Altimeter) has $30 to $5 per million tokens — retail GPT inference, ~3 years, from the Stanford Review interview, not the essay. Gerstner (Altimeter) has ~99% over ~2.5 years on an undefined unit. Quoting one hides two.

Why frontier training remains capital intensive — the expensive-to-build half of the central tension.

The best starting point on power, data centres, and grid timing.

The optimistic macro case; pair with Acemoglu's task-based caution and let the reader hold both.

The cautious task-based macro case — the disciplined counterweight to the 7% headline.

The counterweight that keeps the guide honest: a falling labour share is not a transfer from labour to capital, and real compensation is at an all-time high. The worry is the future Jevons world, not the present tense.

Why agentic workflows change token economics: unit prices fall, total bills rise.

A concrete catalogue of enterprise adoption — useful ballast against both hype and doom.

The missing India vantage on compute: a consumer-integrator with supply access but little capability ownership. The best person I know working on AI and semiconductors in India.

The core frame: a GPT changes society by changing the production map, through complements and reorganization.

The vocabulary the whole Work part runs on: displacement vs reinstatement vs productivity, and so-so technologies that displace without big gains.

Patterns of sustainable specialization and trade: adjustment is a search process, so optimism should come from search capacity, not inevitability.

Task-level evidence from millions of Claude conversations — the augmentation/automation split is the number to watch.

Scale and activity mix of consumer AI use — the diffusion baseline.

Economically valuable tasks as a benchmark target — capability measured in work products, not quiz answers.

The task-horizon lens: how long a task can you delegate? The single most useful capability curve for economics.

Early-career exposure evidence: by June 2026, employment of 22–25-year-olds in the most AI-exposed jobs stood 19% below where it would be had it kept pace with less-exposed peers, and the margin is jobs, not pay. But that is a raw gap. An education control roughly halves the uncontrolled estimates; with firm controls, the next-most-exposed fifth's drop (about 11–13%) holds up while the top fifth's is fragile; and national survey data to 2024 shows a much smaller gap, except in professional and finance jobs.

Youth employment, skills, informality, and structure — the most-cited India source in the work deck for a reason.

Current employment structure and labour-market constraints — the base map for the India part.

How to connect a national shock to local labour markets — the method the India part borrows for AI-exposed cities.

Diamond–Mirrlees for the AI age: tax output, not intermediate inputs. I agree — and it does not reduce my worries about a post-AGI world. Hold both.

The canonical this-has-happened-before brick: British real wages roughly halved 1806–1820 and shared prosperity took decades plus political voice. The productivity bandwagon is a choice, not an automatic.

Travel-agent employment fell over 60% from its peak, yet surviving agents' relative pay rose from 87% to 99% of the private-sector average. The machine took the separable part of the bundle and left people the strong part.

“Labour markets price jobs, not tasks.”

The demand-side complement to Garicano: Starbucks automated for years, concluded it was a mistake, and is re-hiring baristas for handwritten notes on cups.

“The relational sector grows precisely because it isn't automated.”

The optimist's structural-change rebuttal to humans-as-horses — my counter: the induction argument is just saying the sun will rise tomorrow because it always has. One day it won't. A ready-made dialectic.

Managers make good workers in an AI-first world: agentic tools amplify returns to judgment and allocation rather than flattening them. Verify the linked PDF before quoting specifics.

The freshest as-we-speak evidence of the shift to delegated production: 5× weekly users in H1 2026, ~10× more 8h+ delegations since January. Caveat: OpenAI-internal usage is the frontier, not the average firm.

The cleanest mechanism for who captures value: while labour is a bottleneck (Baumol) its share can rise; once it isn't (Jevons), the capital share climbs toward one. A post, not a working paper.

“Piketty was wrong about the past. He's probably right about the future.”

The most uncomfortable, most India-relevant implication in the literature: when capital substitutes for labour, the engine of catch-up growth shuts off. Say it plainly.

Microdata for how firms reorganise: AI-native startups are 25% smaller, 13% more engineers, ~15% fewer entry-level workers and managers — comparable valuations, more value per employee. Cited in both talks; one treatment in the guide.

Firm-level spend linked to workforce records: high-intensity AI adopters grew headcount 10.2% over two years and entry-level headcount 12.0%; low-intensity adopters did not. The mechanism is not exposure alone, but reorganization intense enough to change the firm.

“AI adoption changes employment only when it changes the firm.”

The n=1 limit of the AI-native firm and the optimistic flip side of the canary: ~4m Americans earning $100k+ solo. The India question — solo-scaling on DPI rails — should be stated as a question, not a finding.

Before asking what AI does to Indian jobs, ask whether we can measure Indian jobs. Bonus vignette: when the US blocked computer sales, the ISI built India's first computer in 1953 from war-surplus parts.

“A country that could not buy computing technology built its own from the debris of wars.”

A century of human-capital divergence: China bottom-up, India top-down. Education inequality is ~25% of wage inequality in India vs under 12% in China — why India did services and China did factories.

The crisp backbone for engineers-vs-lawyers: talent that organises production innovates; talent that rent-seeks redistributes. Lands hard with an Indian professional audience.

The engineering-state vs lawyerly-society thesis, heavily highlighted in my copy. Pair with Hessler's Other Rivers and Ang's How China Escaped the Poverty Trap.

The missing macro spine for the India half: manufacturing became an enclave, services must absorb labour but hit a home-market ceiling — and AI now threatens the services escape hatch too.

Live baseline evidence for the India half: manufacturing exports are weak, high-skill IT services face AI pressure, finance will not absorb low- and mid-skill workers at scale, and gig platforms are becoming an urban employment sink. Use as a tracking source, not a clean causal AI claim.

“The gig economy is emerging as an urban employment sink.”

The single most on-thesis India-AI-labour case: ~$50k revenue per employee vs a US benchmark near $300k. India's services-arbitrage model is the thing AI most directly erodes. (Original is paywalled; the headcount economics are corroborated in Lightspeed's open write-up.)

Ground-level texture for India climbing the manufacturing ladder: a dormitory-labour regime traced back to 1920s Shanghai cotton mills. The vivid counterweight to abstract labour-share macro.

The rigour footnote for any manufacturing-employment claim: household surveys say 68.5M by 2022–23, enterprise surveys a static ~48M. Keep every claim defensible.

Telangana's Composite Backwardness Index across 242 caste groups: the strongest predictor of backwardness is access to English-medium education, not land. The bridge from caste to who reaches the AI economy.

AI exposure mapped onto caste — the working paper behind the inequality-layers section.

TFR already 1.88, projected to US levels by ~2031, Kerala 1.3 vs Bihar 2.9. The dividend window is closing faster and unevenly. Not a crisis yet, but it is coming.

First causal national estimate: no positive spillover to US-born workers from immigration enforcement. The India parallel is sons-of-the-soil politics meeting internal migration.

The careful framework behind the token-tax section: stock vs flow, and Diamond–Mirrlees as the spine — tax output, not the intermediate input. The steel line people quote is not in this paper; for that, see their Brookings piece.

The public-facing version of the NBER primer, and the document where the steel line is actually in print under their own names. The deck's token-tax slide quotes it directly rather than at second hand.

“Taxing the fundamental infrastructure of AI development would be like taxing steel during the industrial revolution—a self-defeating policy that could slow the productivity growth needed to fund public priorities.”

The policy-menu source: token taxes, sovereign wealth funds, automation adjustment assistance, the physical-vs-human-capital tax bias. My caveat: the scenario buckets are not well defined.

The UBI argument is really a state-capacity argument: a reliable cash floor plus portable benefits plus active placement, delivered by a state that can actually do those things.

The mechanism behind Amara's law: visible costs are measurable inside the old practice, while the most important benefits often require new practices and concepts before they can even be named. The policy test is recoverability: preserve redundancy where losses may be irreversible, but prefer managed experimentation where costs can be contained.

“Measurability and importance are not the same thing.”

The Hayekian reading of Amara: the short-run shortfall is mundane friction; the long-run surprise comes from dispersed experiments no planner could design. Exactly the PSST point, made rigorous.

“Anybody who says they predicted this is selling something.”

The named optimist to argue with, not a strawman: finite compute means opportunity cost governs deployment. He concedes the three cracks — inequality, adjustment frictions, AI capturing its own means of production.

The honest caveat before prescribing India's next bargain: history only happens once, so there is no science of development. Humility on which lever is decisive.

A humanities-register frame for the close: India as a pattern of accommodating contradictions rather than resolving them. Use sparingly; it lands.

“India is about making sure we play the game forever, not about making sure we win it.”

Depreciation schedules are where the AI boom's profitability gets decided. The course pushes back: Lochmiller (Crusoe, which rents this hardware) showed H100 spot prices falling after launch, then rising above what the first buyers paid; Srivastava (Baseten) saw a B200 cluster requoted from $263 an hour to $510. Asked whether useful life runs past six years, Lochmiller said he does not know. Nobody does.

Above water, not ashore: trailing revenue about $110bn, about 19% depreciation headroom on stated assumptions. Agrawal's two-year update reaches much the same conclusion by a different route — bottom-up revenue and assumed margins by layer. But Azhar's figures are unverified here and carry much of the Money deck's depreciation arithmetic. Check them before they go on a slide.

Mollick's one-line version of this guide's opening claim: steam changed nothing until thousands of skilled workers worked out how to apply it to their existing jobs. Swap steam for AI and that is the adaptation problem exactly — the technology is the easy part; the reorganization is the long grind. He points to the BBC 'Industrial Revelations' series as the illustration.

“Steam alone wasn't enough, it required thousands of skilled workers figuring out how to apply this new power to their old work.”

Useful reinforcement, not a view change: the right test is whether downstream applications prove enough commercial value before upstream margins, financing, power, and depreciation squeeze the story. Read it with Azeem Azhar, GeometricInvestor, and the talk's own ledger frame in mind: not agreement, but 'we shall see.'

“The critical place in the market to keep a close eye on may be the downstream side of AI applications and product development.”

Independent corroboration of the Stripe signal: 5.7 million business applications in 2025, with the rise skewed toward firms unlikely to hire. This is evidence that AI may be lowering the minimum viable firm, not proof that the Coasean singularity has arrived; applications are not operating businesses, and roughly 500,000 actual firms formed in 2023.

The clearest economic account of why open weights do not make inference free: the useful unit is the cost of completing a task, and rents migrate into serving efficiency, learning data and sticky harnesses. Thompson stops short of the counter-evidence — on Vercel's gateway in June, open-weight models ran 29% of tokens on under 4% of spend. Open weights do not cap model-layer rents. They concentrate them.

“Let the frontier labs win by being better; don’t let them define safety or security, or pull up the ladder of humanity’s collective knowledge.”

The headline the Money part leans on: $435bn of AI revenue, $285bn of gross profit, semis taking 79% against infra's 14% and apps-and-labs' 7%. Caveats: the margins are assumed, not measured; the $435bn stacks the same end-user dollar up to three times; and eight weeks later Agrawal sized the stack at $300bn / $150bn / $100bn, roughly 65/24/10. Altimeter's disclosed US book is 41.6% semiconductors.

“For the underlying data behind this analysis, ping me. Happy to share the full datasheet.”

The chain was already inverted in 2024: compute taking about 83% of revenue and 87% of gross profit while applications earned almost nothing. Agrawal (Altimeter) predicted the flip. Two years on it had barely moved, 87% to 79% (assumed margins; revenue base double-counts — see Agrawal), and the non-movement is the finding. It carries footnotes naming Coatue, Meritech and Intel/AMD; the update does not.

Agrawal (Altimeter) teaches the course and wrote the essay behind its headline number; his fund's disclosed US book is 41.6% semiconductors — he is describing a stack he is positioned in. The gift is the ARPU ladder: Alphabet 4bn users at ~$100 a year, Meta 3.5bn at $70, ChatGPT 1bn at $10, 95% paying nothing. Note the hole: he promises slide 16, the profit-by-layer chart, and runs out of time.

The course's most useful cost breakdown. Lochmiller (Crusoe, which rents this hardware) puts a site at ~$60M per MW all-in — $20M shell and power, $40M IT — against $15M per MW a year renting bare compute, a four-year payback. Then he says on tape that he may have double-counted, that engineering labour is missing, and that he assembled it that afternoon. For India: gas turbines went $1M to $3M per MW.

The sharpest number in the course: about 90-95% of inference spend goes to frontier models, about 5% to custom or open ones. Srivastava (Baseten) sells the custom alternative and says so in the same breath — read the 5% as a pitch and the 90-95% as evidence. His B200 cluster was requoted from $263 an hour to $510. The East India Company exchange with a public-company CEO is positioning, not a finding.

The dissent inside the course, from the buyer. Katti runs infrastructure at OpenAI, the largest customer of the layer he says will not keep the profits: value migrates up the stack, as in mobile. A gigawatt is roughly half a million GPUs and about $70bn — which sits beside Gerstner's (Altimeter) $50bn and the ~$60bn implied by Lochmiller's (Crusoe) $60M per MW, three numbers for the same unit, and nothing published reconciles them.

Reorganisation beats capability, from the buyer himself. Databricks went from nine months per connector, one engineer each, to seven connectors in a quarter shipped by seven engineers working together. The fix was process redesign, not a better model. Ghodsi sells the layer above the models and expects them to commoditise. The pull-quote is machine-caption text; verify against the video before a slide.

“The next GPT 7 or Opus 6 would not have helped us do this better.”

Two men long the answer, stating the paradox best. Gerstner (Altimeter) runs a book whose largest disclosed US position is NVIDIA at 28.6%; Madra came from Groq to NVIDIA. Their number: unit inference cost down ~90% in a year and ~99% over two and a half — while H100 prices went up. Cheaper intelligence and dearer hardware at once, from people long both halves. The $50bn per gigawatt figure is from here.

The best public read on where inference money actually goes, and it cuts against the commoditisation intuition. In June, open-weight models ran 29% of gateway tokens on under 4% of spend, while Anthropic took 61% of spend on 32% of tokens. Cheap models take the cheap work, which leaves the paid work more concentrated. The limit: one gateway's traffic, Vercel's own developer customers, not the market.

“Open-weight models ran 29% of gateway tokens, up from 11% in April, on under 4% of spend.”

The disclosure the course does not make. Altimeter runs the seminar and writes the essay behind the 79% number; at 31 March 2026 it held $5.70bn across 13 US-listed longs — NVIDIA 28.6%, Meta 19.6%, semiconductors 41.6% of the book. Read that claim next to this book (assumed margins; revenue base double-counts — see Agrawal). 13F excludes shorts and the private book, so it understates the overlap.

Useful mainly as a check on the headline. Eight weeks after publishing $300bn semis, $75bn infra and $60bn apps, Agrawal (Altimeter) sized the same stack here at $300bn / $150bn / $100bn, roughly 65/24/10 rather than 79/14/7. Nothing was retracted; the numbers moved in conversation, as assumed-margin estimates do. This interview, not the essay, is where the $30-to-$5 per million tokens line lives.

A toy that makes one point stick: run the defaults and energy binds first, at 15% a year against chips at 45%, memory at 28% and interconnect at 35%, with demand at 310%. Effective supply is the minimum of the layers, so the slowest sets the ceiling: power. For India: grid and energised shell bind, not chips. The defaults are Epoch AI's, with the energy figure from Epoch and the IEA, under CC BY 4.0.

“This is a pedagogical toy, not a forecasting model.”

The model is not the economic unit; the model, human verifier, error cost, coverage, and fallback plan are. That makes the best workflow model potentially different from the most accurate standalone model.

A rare look inside a frontier lab: agent workdays already exceed human workdays, but long tasks still need intervention and priorities remain human. Useful mechanism evidence, with the obvious caveat that the lab measures itself.

The governance lesson is not only that agents can misbehave. It is that they can coordinate, target the evaluation, and corrupt the evidence the evaluator sees. Verifiers must inspect systems, not isolated outputs.

The November 2025 baseline: India number two in total Claude.ai use and number 101 per working-age person, with software 45.2% of its tasks, the world's highest. The gap between the two ranks is more informative than either. By May 2026 the ranks and state concentration held, but software no longer stood out; see Anthropic's June 2026 data.

A useful antidote to one-line forecasts. Capability, adoption, new-task creation, and adjustment speed can turn the same technical frontier into modest, substantial, or extreme labor-market change.

Income is only one job that work performs. Competence, impersonal standing, routine, and belonging matter too; a post-work settlement that replaces money alone leaves the harder institutional problem unsolved.

The real choice is not whether technology gets steered; firms and investors already steer it. The political question is which trajectory they choose, who can contest it, and who gets the gains.

Rare India data on AI and work, in one place: Gallup finds only 23% of Indian employees engaged and 59% quietly quitting; TeamLease's IT fresher-hiring intent slipped from 81% to 76% this year while employers now demand AI fluency from day one. The link to AI rests on Autor's patent-lawyer trial, where AI lifted everyone's drafts but only seniors kept the skill without it. That the machine is causing the ennui is conjecture; that it can quietly skip the apprenticeship is evidence.

“Who will mentor the fresh recruits and help them build skills?”

The cleanest test yet of whether AI builds or skips the apprenticeship. 133 patent lawyers, randomised: AI raised the quality of everyone's drafts, most for juniors. Three months later, working without AI, only seniors had kept a gain; juniors' scores split between better and worse. What you already know decides what AI teaches you.

“Foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.”

The report never mentions India, but its open data does. In May 2026 India was still second in total use (7.1%) and about 102nd of 121 (tied) per working-age person, with four states supplying half of it. What changed is the task mix: computer and maths work is 26% of India's use against 24% globally, no longer the world's highest. Part of that is a wider sample, now including Cowork, so read it as broadening, not proof.

A labour economist reads the Census Bureau's 2026 AI survey: about one firm in five uses AI, but only about 1% of all US firms say it changed their headcount, split roughly evenly between more and fewer. Keep that next to every layoff headline, and keep his own caveat next to it: small so far does not mean small in 25 years.

““Small impact so far” does not mean “small impact 25 years from now”.”

The primary source behind Berger's reading: 18% of US firms used AI between November 2025 and January 2026 (32% weighted by employment), and only about 5% of those users report any AI-driven headcount change, 2.3% up and 2.0% down. The sharper finding is underneath: firms that use AI across more business functions, and invest to support it, are more likely to cut staff; firms whose workers use AI in their tasks are not, once breadth and investment are held equal.

The college-grad crisis is real but overstated: a 22-year-old graduate is just starting out, while a 22-year-old without a degree may have worked for four years, so same-age comparisons inflate the graduate penalty. He thinks the Canaries paper largely survives this, but wants it tested on 18–21 year olds. AI is probably playing a role, in his words, but not a dominant one.

Written in early 2025, and still the best alternative explanation for the canary: the Great Stay. Hiring fell 26% without a recession; low layoffs protect insiders while new entrants queue. The hiring rate is no better today (3.2% in July 2026, comparable to 2010–11), so any AI effect on young workers has to be found inside a hiring slump that would hurt them anyway.

The AI buildout as a macro problem: a resource-hungry boom on top of ordinary demand means inflation, so of butter (consumer demand), data centers and low inflation, the US may get only two. Keeping inflation down means tighter monetary or fiscal policy. The phrase riffs on Conor Sen; the worry is echoed by Chicago Fed President Goolsbee. A useful corrective to treating capex as only a question of whether the labs earn it back.

“Butter, AI data centers, low inflation - we’ll get to pick two if we’re lucky.”

The sharpest dissent to the Stripe and NYT entries: on federal data, applications from businesses not planning to hire are up about 38% since 2024 (the higher of his two measures), but solo work as a main job is flat and applications that plan to pay wages have fallen, so the boom so far looks like side gigs, not people leaving jobs. He concedes the clearest pro-AI signal (such applications went vertical in information and professional services), and his industry and metro tests use pre-agentic AI-use data, but the side-gig finding does not depend on them.

A growth investor's 90-slide case that the capex is working: hyperscaler capex close to $800bn in 2026 on consensus and over $1tn a year from 2027, GPU rents holding up, and Anthropic, Google and OpenAI taking about 70% of OpenRouter spend on 35% of tokens (down from the mid-80s earlier this year). Read it for the charts, not the verdict: the same pages show hyperscaler free cash flow squeezed until about 2028 and over $300bn of new bonds and off-balance-sheet borrowing by hyperscalers and Nvidia so far this year, and a16z invests in the outcome.