$75.2bn
NVIDIA data-center revenue in Q1 FY2027, up 92% year over year. Compute is $60.4bn of that; networking is the other $14.8bn.
Cash has already reached the bottom of the stack.
Saturday Seminar - June 2026 - working professionals
The Economics
of AI
How value is created, who captures it, and why depreciation suddenly matters
Ashish Kulkarni
an economic lens, an accounting lens, and a clearer way to read the news
Press P to switch into slide-show mode. Press P again, or Esc, to return to the single scrolling version. Press C to hide the six optional depth slides.
The puzzle
Four entries in the AI ledger all point in different directions.
The technology is getting cheaper to use. The infrastructure is getting more expensive to build. Suppliers are already earning. The final buyer return is still being proven.
Supplier ledger
Power ledger
415 → 945 TWh
IEA base-case global data-centre electricity use, 2024 to 2030.
Software is turning back into physical infrastructure.
Cost curve
600x
Estimated decline in LLM token prices across 2020-2026.
Cheaper intelligence may create more usage, not a smaller bill.
Return hurdle
$1tn+
Illustrative annual monetizable AI revenue needed if capex reaches the multi-trillion-dollar range.
This is an estimate, not a settled fact.
Sources: NVIDIA Q1 FY2027 results; IEA, Energy and AI; Du, token-price study; GeometricInvestor, The AI Capex Ledger.
How we'll do this
Two lenses, one story.
Lens one - economics
How value is created
Trace the value chain from chips and data to models, apps, firms, workers, and consumers. This lens asks: what gets cheaper, what stays scarce, and who benefits?
Lens two - accounting
Whether value is captured
Trace the required returns across the stack: suppliers sell chips and power gear to cloud operators; cloud operators sell compute to model labs; model labs sell tokens to application builders; apps sell AI capability to firms and consumers. Each of those five layers has to earn enough for the layer below it to stay funded.
The thesis is deliberately modest: AI can create enormous social value while still disappointing some of the investors who financed the infrastructure. That is not a contradiction. It is the normal tension between value created and value captured.
I
Lens one - economics
How value is created
Start with the simple value chain. AI is not magic floating in the cloud; it is an industry with upstream inputs, midstream producers, and downstream users.
Upstream
What does it cost to make AI?
Chips, data, training, inference, power, cooling, and talent.
Midstream
Why do so few firms sit at the frontier?
Fixed costs, scale economies, model know-how, distribution, and open-weight pressure.
Downstream
Who benefits from using it?
Consumers, firms, workers, capital owners, and states - rarely in equal measure.
Upstream - build vs run
The central tension: expensive to build, cheapening fast to use.
The economics of AI begins with a split: frontier training looks like a giant fixed cost; inference looks like a digital service whose unit price keeps falling.
Training - build the factory
Inference - run the factory
Model training costs have risen about 2.4x/year since 2016 in one cost-model estimate.
Token prices have fallen roughly 600x across 2020-2026 in one market-pricing study.
If the trend continues, the largest training runs cost $1bn+ by 2027.
Lower unit prices invite more tasks, longer contexts, and agent loops.
This is why both claims can be true: AI can become cheaper per unit and more expensive in total. That is the economic logic behind the capex boom.
Sources: Cottier et al., The rising costs of training frontier AI models; Du, Tiered Super-Moore's Law; IEA, data-centre demand.
Stylized cost curves
The picture is deliberately stylized: one curve is the cost of building frontier capability; the other is the price of running useful intelligence.
Upstream - scarcity
Three inputs make the "cloud" feel very physical.
Compute
$60.4bn
NVIDIA data-center compute revenue in Q1 FY2027 under its prior reporting framework. It is the compute part of the $75.2bn data-centre total quoted earlier; networking is the remaining $14.8bn.
Electricity
15%/yr
IEA base-case growth rate for global data-centre electricity demand from 2024 to 2030.
Web data rights
45%
Share of the C4 web-training corpus now restricted by website Terms of Service in one 2024 audit.
The common thread is not that AI will "run out" of everything. It is that each scarce input changes bargaining power. Chips, power, and high-quality permissioned data are all places where control can turn into surplus capture.
Sources: NVIDIA Q1 FY2027 results; IEA, Energy and AI; Longpre et al., Consent in Crisis; Zhang et al., Regurgitative Training.
Midstream - market structure
Why the market wants to concentrate - and why open weights keep pushing back.
The concentration force
- High fixed costs favor firms that can spread them across huge user bases.
- Distribution matters: cloud, search, office software, devices, and developer tools are demand channels.
- Know-how compounds because model training is partly engineering craft.
The price-ceiling force
- Open-weight models make "good enough" intelligence cheaper to adopt.
- They do not erase inference costs: the competitive unit is useful work, not raw tokens.
- They push advantage toward serving efficiency, learning data, and sticky workflow harnesses.
What open weights have won
What they have not won
Open-weight models ran 29% of the tokens on Vercel's AI Gateway in June 2026, up from 11% in April.
They took under 4% of the spend. Anthropic alone took 61% of spend on 32% of tokens.
Roughly one in eight of that gateway's enterprise customers now runs an open-weight model in production.
Baseten's Tuhin Srivastava puts 90-95% of all inference spend on frontier models, about 5% on custom or open ones.
Read that against this slide's own bottom line, because it cuts the other way. Open weights cap model-layer rents in principle, and they are winning volume fast. The money has not followed. Cheap tokens are going to the cheap models and the dear work is still going to the frontier. Two honest limits: one gateway is not the market, and Vercel sells the gateway. Scale still pulls toward concentration, but value migrates to whoever completes the task most cheaply, learns fastest from use, or owns the harness users are reluctant to leave.
Sources: Ben Thompson, Who’s Afraid of Chinese Models?; Vercel AI Gateway Production Index, 13 July 2026 (June data); Tuhin Srivastava (Baseten), Stanford MS&E 435, lecture 7; Meta, Llama 3.1 release; Du, token-price competition; Coles et al., Apertus engineering journey; Vake et al., open-source AI; SemiAnalysis, AI Value Capture. Srivastava sells custom-model serving, so he has an interest in the 5% growing.
Downstream - surplus
The buyer may capture value that the lab cannot bill.
If a professional pays a flat monthly subscription and saves hours of work, much of the value never appears as the model lab's revenue. Economists call that consumer surplus.
Value created
Value captured
Time saved, better drafts, faster code, cheaper analysis, more learning.
Subscription revenue, API fees, enterprise contracts, cloud margins.
Can be enormous and widely distributed.
Must be visible on a profit-and-loss statement.
This distinction is the spine of the talk: good technology and good investment are related, but they are not the same question.
Sources: OpenAI ChatGPT pricing; The Verge on Claude Max pricing; GeometricInvestor, value-created/value-captured frame.
Downstream - subscription wedge
$200 can buy more AI work than $200 of revenue suggests.
Public coverage of SemiAnalysis testing reports that maxed-out flat-rate plans can deliver thousands of dollars of API-priced usage, especially on long coding and agent tasks.
ChatGPT Pro
$14k
Reported API-priced monthly usage from a fully used $200 plan.
Claude Max
$8k
Reported API-priced monthly usage from Anthropic's comparable top tier.
The margin climb
0 → 40-70%
Where application companies start on gross margin, and where Baseten's Tuhin Srivastava says they have to get to. His route is moving token volume off the frontier models.
Value created
Value captured
Repo sweeps, debugging loops, research, implementation passes, and long agent runs.
A flat subscription, plus routing limits, usage caps, and pricing experiments.
User surplus can be enormous when the task is valuable.
The provider still has to turn that usage into sustainable margins.
Read the numbers carefully: they are API-equivalent retail values, not the lab's literal cost. But the economics lesson is exactly the one we need - a flat subscription can make value visible to the user and only weakly visible to the seller. The labs do not disclose their own gross margins, so the 0 to 40-70% climb is an operator's target rather than a measured figure.
Sources: SemiAnalysis Tokenomics Model; Tom's Guide on SemiAnalysis subscription testing; Let's Data Science summary; SemiAnalysis, AI Dark Output; Tuhin Srivastava (Baseten), Stanford MS&E 435, lecture 7. The first three sources carry the $14k and $8k figures; the margin climb is Srivastava's, and he sells the service that delivers it.
Downstream - labour and productivity
AI changes tasks first. Jobs and productivity statistics move later.
The task lens
A job is a bundle of tasks. AI automates some, complements others, and creates new ones. The practical question is not "Will my job vanish?" It is "Which tasks move, and where does my comparative advantage remain?"
The macro uncertainty
Goldman Sachs models large upside; Acemoglu's task-based model is much more cautious. The disagreement is not about whether AI helps somewhere; it is about whether enough firms reorganize fast enough to move aggregate productivity, and whether the benefits that matter most are even measurable before new practices exist.
Question for the room
If AI is better at many tasks, what should humans become relatively best at?
Sources: Acemoglu, The Simple Macroeconomics of AI; Goldman Sachs Research, GDP/productivity upside; Handa et al., Anthropic Economic Index task data; Carlo Cordasco, What we can’t measure about AI – yet.
Downstream - organization
Messrs Coase and Cheung: productivity appears when contracts change.
Coase asked why firms exist if markets can contract for everything. Cheung pushes the question down one level: which contract is cheapest when output, effort, quality, and risk are hard to measure?
Coase
Costly markets
Search, bargain, monitor, and enforce. Firms economize when internal coordination is cheaper than contracting task by task.
Cheung
Contract spectrum
Wages, piece rates, subcontracting, platforms, and franchises are alternative ways to price what can be measured.
AI implication
Redesign lag
AI makes tasks cheap; firms still need new rules for verification, responsibility, handoffs, and surplus sharing.
Before AI
After AI
Work was routed by job title, team, and software seat.
Work can be routed by task, model, reviewer, and liability.
The double thank-you was mostly human-to-human.
The contract has to keep both thank-yous alive when an agent does part of the work.
That is why productivity statistics move late. The tool arrives first; the organization has to discover the contract, vocabulary, and workflow that make the new division of labor legible.
Sources: Coase, The Nature of the Firm; Foss on Cheung's contractual view of the firm; Ashish Kulkarni, Contracts and the Double Thank You note, 2026; Carlo Cordasco, What we can’t measure about AI – yet.
II
Lens two - accounting
Whether value is captured
The accounting lens asks a colder question: after capex, depreciation, power, financing, and model costs, who earns a return?
The stack as a ledger
Each layer's revenue is the next layer's cost.
I
Infrastructure suppliers
Chips, HBM, networking, power gear, cooling.
already earning
II
Hyperscalers and neoclouds
Convert capex into rentable compute.
must cover depreciation
III
Model labs
Train and serve frontier and open models.
must hold its margin
IV
Applications
Wrap models into products people pay for.
must earn a gross margin
V
Buyers and the macro economy
Firms, consumers, productivity, prices, jobs.
must show real ROI
The proof lives at the top of the stack. The supplier can be paid in cash today while the buyer's productivity gain is still only a hope. One practitioner has tried to put money against these five layers, and that is the next slide.
Sources: GeometricInvestor, The Four Ledgers of AI; Apoorv Agrawal (Altimeter Capital), Stanford MS&E 435, lecture 1.
The stack as a ledger - the money
One investor has put numbers on the layers.
Apoorv Agrawal of Altimeter Capital estimates where the ecosystem's gross profit actually sits. It is the only public attempt at the number, and it comes with three health warnings.
Where the gross profit sits
What the split does not mean
Semiconductors ~$225bn, infrastructure ~$40bn, applications ~$20bn. That is 79 / 14 / 7.
The margins behind it are assumed, not measured: about 73%, 55% and 33%. Change one assumption and the split moves.
Two years earlier the same split was 87 / 10 / 3. The bottom of the stack is giving ground, slowly.
The $435bn revenue base counts the same end-user dollar up to three times, chips to cloud to app. It is not value added and is not a GDP-style number.
OpenAI and Anthropic together are about $45bn, roughly three-quarters of the whole application layer.
Eight weeks later the same author sized the stack at $300bn / $150bn / $100bn, which recomputes to roughly 65 / 24 / 10.
Interest: Agrawal is a partner at Altimeter Capital, whose Q1 2026 13F reports 41.6% of $5.70bn of US-listed longs in semiconductors and 28.6% in NVIDIA alone. He is sizing the profit pool of the layer his fund owns most of.
Sources: Apoorv Agrawal (Altimeter Capital), The Economics of Generative AI: Two Years Later, 1 April 2026; Agrawal's revised sizing, Stanford Review, 28 May 2026; Agrawal, Stanford MS&E 435, lecture 1; Altimeter Capital Q1 2026 13F. A 13F is long-only, US-listed, filed 45 days late, and omits the private book, so it understates the interest.
The missing accounting slide
Depreciation is why the useful life of a GPU matters.
Depreciation is the accounting rule that spreads the cost of a long-lived asset over the years it is expected to be useful. It is not just bookkeeping: it changes reported profit.
Illustrative cost
$1bn
spent on AI servers today
If useful life = 5 years
$200m/yr
depreciation expense
If useful life = 2 years
$500m/yr
depreciation expense
Same server, different assumed life
Shorter useful life means the same cash outlay hits profit faster. The machine is identical; the accounting assumption changes the yearly expense.
The other side of the argument
- H100 spot prices fell after launch, then rose back above what the first buyers paid. Crusoe's Chase Lochmiller showed the series in April 2026 and says Blackwell pricing is tracking the same shape.
- Baseten was quoted $510 an hour to renew a B200 cluster booked at $263. Scarcity can push the price of installed kit up, not down.
- Amazon books 5-6 year lives for chips, servers and networking gear and 30-plus years for the datacentre shell. Andy Jassy says returns on that capex are cumulatively attractive a couple of years after the kit is in service.
The cash went out on day one either way. The accounting question is how quickly that cash cost hits profit. If AI hardware becomes obsolete faster than the books assume, reported margins can look healthier than the economics really are. The counter-evidence does not settle it: a resale price tells you about this year's shortage, not about what a chip is worth in year five. It does mean the short-life case has to be argued rather than assumed.
Sources: WSJ, Big Tech Accounting Creates a Blind Spot in the AI Boom; GeometricInvestor on depreciation risk; Chase Lochmiller (Crusoe), Stanford MS&E 435, lecture 3, citing Bloomberg and SemiAnalysis price series; Tuhin Srivastava (Baseten), lecture 7; Andy Jassy, 2025 letter to shareholders. Arithmetic is illustrative. Lochmiller and Jassy both sell compute.
Accounting lensDemand lens
The Azeem report says AI is above water, not ashore.
The useful update is that revenue is real and growing fast. The unresolved question is whether it keeps outrunning depreciation, power, and efficiency assumptions.
What improves the ledger
What still has to hold
$110bn trailing revenue and a $175bn run rate; AI infrastructure revenue now clears quarterly depreciation with 19% hyperscaler/neocloud headroom.
That headroom assumes 6-year IT life and 14-year building life. Shorter chip lives can erase it quickly.
A 1GW fleet model gives 75-85q tokens/year and about $0.10 provider cost per million tokens before margin.
The denominator depends on GB200 NVL72 systems, FP4 inference, 65% utilization, workload mix, and speculative decoding.
Demand appears elastic: a 10% price cut is associated with 12-18% more tokens, so spend can still rise.
OpEx remains physical: the 1GW model carries $900m/yr OpEx, including $594m for energy.
Verdict: promising, but not yet a blank-cheque bull case. The bull case needs revenue growth, utilization, tokens-per-watt, and useful lives to improve together.
Sources: Exponential View, State of the AI Economy 2026; Azeem Azhar, accompanying essay.
The size of the bet
The capex cycle needs revenue, not just enthusiasm.
Buildout
$3tn
One reported estimate for big-tech property-and-equipment spending over the next four years.
Revenue hurdle
$1-2tn/yr
Range of estimates for annual AI revenue needed to validate the buildout by decade-end.
Disclosure gap
79 / 14 / 7
The decisive metric is not token volume; it is gross profit after depreciation, power, and financing. No firm reports it. Apoorv Agrawal estimates the pool at ~$225bn for semiconductors, ~$40bn for infrastructure and ~$20bn for applications.
His margins are assumed, not measured. The $435bn revenue base counts the same end-user dollar up to three times. Eight weeks later he resized the stack to $300bn / $150bn / $100bn, roughly 65 / 24 / 10 on the same margins.
This is the accounting lens in one sentence: usage is necessary, but not sufficient. A trillion tokens are good news only if they become durable gross profit and durable buyer ROI. The honest position is not that the number is unknowable. It is that nobody is obliged to publish it, so the best figure we have is one investor's estimate built on three margin assumptions.
Sources: WSJ on $3tn capex estimate; WSJ on Bain's $2tn revenue estimate; Sequoia, AI's $600B Question; GeometricInvestor hurdle framework; Apoorv Agrawal (Altimeter Capital), The Economics of Generative AI: Two Years Later, 1 April 2026. Interest: Altimeter's Q1 2026 13F holds 41.6% in semiconductors, 28.6% in NVIDIA alone.
depth - optional
How to audit the bet
The useful scoreboard is boring, which is why it matters.
For compute owners
- Utilization-adjusted compute margin.
- Gross profit per watt.
- Depreciation-adjusted return on deployed compute.
- Customer concentration and contract durability.
For buyers
- Revenue per employee rising.
- SG&A intensity falling.
- Support, coding, analysis, and operations budgets moving from pilots to operating spend.
- Productivity data eventually confirming the firm-level stories.
Sources: GeometricInvestor, token and buyer ledgers; BLS Productivity and Costs, Q1 2026 revised.
III
The present tense
How to read today's AI headlines
Every headline belongs to one or more ledgers. Ask: does it prove value creation, value capture, or just another financing round?
Economic lensAccounting lens
Cheaper tokens do not automatically mean a smaller AI bill.
A falling unit price can trigger more usage, longer contexts, and agentic loops that spend far more tokens per task.
Good news
Accounting trap
Blackwell-era systems may produce far more tokens per watt and at lower cost per million tokens.
Agentic coding tasks can consume vastly more tokens than ordinary chat or code reasoning.
Capability becomes cheaper and more widely available.
Total spend can still rise if use expands faster than unit cost falls.
Unit inference cost fell about 90% in one year and about 99% in two and a half, on Brad Gerstner's account.
Over the same stretch H100 prices went up. Cheaper tokens and dearer machines are not a contradiction; they are the same demand surge seen from two ends.
The elasticity slide gave the number that decides this. A 10% cut in the token price is associated with 12-18% more tokens bought. That is an elasticity between 1.2 and 1.8, and anything above one means the bill goes up when the price comes down. Falling prices are not a saving. They are an invitation.
Sources: Business Insider reporting on SemiAnalysis Blackwell token economics; Bai et al., agentic coding token consumption; Du, token-price evolution; Exponential View, State of the AI Economy 2026 for the 12-18% elasticity; Brad Gerstner and Sunny Madra, Stanford MS&E 435, lecture 2 for the cost decline and the H100 price.
Economic lens
Sovereign AI is an insurance policy against dependency.
The reported Anthropic export-control episode made the dependency problem vivid: a strategically important model can become unavailable for reasons outside the user's control.
What the headline means
The issue is not just model quality. It is continuity of access. If critical workflows depend on a foreign-hosted frontier model, policy risk becomes business risk.
Why India cares
India's AI strategy is trying to widen access to compute and build domestic capacity. The hard question is whether buying infrastructure creates sovereignty if the chips, models, and clouds remain externally controlled.
Question for the room
If a hospital, bank, or startup can be unplugged from a frontier model overnight, what is that model worth on its balance sheet?
Sources: The Verge on Anthropic/export controls; IndiaAI Mission; Economic Times on AI Impact Summit commitments.
Economic lensAccounting lens
India's AI infrastructure moment is also a return-on-capital question.
The AI Impact Summit reportedly secured more than $250bn of infrastructure commitments. That is an opportunity, but it is not automatically sovereignty or profit.
The economic case
The accounting question
Cheaper local compute can widen access for startups, universities, firms, and public services.
Domestic deployment can create large user surplus even without a frontier lab.
Returns may leak abroad if the scarce inputs and distribution are controlled elsewhere.
Sources: Economic Times, $250bn infrastructure commitments; IndiaAI Mission; Takshashila, GPUs and India.
depth - optional
Accounting lens
The financial plumbing is getting strange because compute is the scarce asset.
Reported deals
- SpaceX agreed to acquire Cursor for $60bn after a prior option-style arrangement.
- MarketWatch reported Anthropic was paying SpaceX about $1.25bn/month under a compute lease, later clarified as short-term and cancellable.
Ledger reading
- The scarce thing is not just the model; it is energized compute at scale.
- Rivals may rent to rivals when the capital cost and utilization risk are large enough.
This is not a moral claim about any one company. It is a structural clue: in a capital-intensive stack, financing arrangements and customer contracts become part of the product.
Sources: The Verge on SpaceX/Cursor; MarketWatch on SpaceX/Anthropic compute lease.
Economic lensAccounting lens
Finance makes a unilateral pause unstable — not a coordinated pause impossible.
Committed capital raises the private cost of waiting. It does not prove that uninterrupted frontier training is socially optimal.
The private incentive
The social decision
Depreciation, financing costs, power contracts, and GPU obsolescence keep running while one lab waits.
Those costs measure how expensive delay is; they do not decide whether another training run is worth its wider risks.
A lab that pauses alone gives rivals the option to keep scaling and capture the upside.
Coordination can remove the first-mover penalty by making competitors and their financiers wait together.
The crucial distinction
The hurdle rate is the price of delay, not a veto over policy.
Source: A balance-sheet stress test of the frontier-AI pause (EconForEverybody, 29 July 2026).
Economic lensAccounting lens
A workable pause has to stand still contracts as well as frontier training.
The financially coherent version is narrow: inference keeps earning, while the frontier race and the obligations that propel it stop together.
Keep running
- Inference and customer deployment below the frontier threshold.
- Safety research, evaluations, and verification work.
- Redeployment of installed GPUs to useful inference.
Stand still together
- Training runs above a defined capability or compute threshold.
- Hardware deliveries, leases, and take-or-pay obligations tied to them.
- New guarantees or circular credit support that make exit harder.
Institutional test
Synchronize across firms and jurisdictions, verify at the compute bottleneck, and define the exit rule before the pause begins.
Source: A balance-sheet stress test of the frontier-AI pause — especially its comparison of pause regimes and coordinated financial-standstill architecture.
Accounting lensEconomic lens
So is anyone actually making money?
The honest answer is ledger by ledger, not yes or no. One verdict for each layer of the stack.
I - Suppliers
Yes
NVIDIA's data-center results show supplier revenue and margins today.
II - Compute operators
Maybe
Hyperscalers and neoclouds must prove utilization and margins after depreciation, power, and financing.
III - Model labs
Contested
Brad Gerstner says lab gross margins have gone from highly negative to very positive. No lab discloses them. Ali Ghodsi expects frontier models to end up a commodity with tiny gross margins.
IV - Applications
Thin
About 7% of ecosystem gross profit on Apoorv Agrawal's estimate. Tuhin Srivastava puts the survival test at climbing from roughly zero to 40-70% gross margin.
Agrawal's margins are assumed, not measured, and his $435bn base counts the same end-user dollar up to three times.
V - Buyers and economy
Emerging
Ramp/Revelio now shows high-intensity adopters expanding headcount; aggregate productivity evidence remains mixed.
The most sensible verdict is not "bubble" or "revolution." It is: the bottom ledger has cleared; the upper ledgers are still on trial, but buyer-side evidence is no longer blank. Note where the disagreement is sharpest. It is at the two layers nobody is required to report. Interest: Gerstner and Agrawal are both Altimeter partners, and that fund is 41.6% semiconductors; Ghodsi and Srivastava both sell the alternative to frontier models.
Sources: NVIDIA Q1 FY2027 results; BLS Productivity and Costs, Q1 2026 revised; GeometricInvestor ledger framework; Kharazian, Simon and Stevens, A New Look at AI’s Impact on Jobs (Ramp/Revelio, June 2026); Brad Gerstner (Altimeter) and Sunny Madra, Stanford MS&E 435, lecture 2; Ali Ghodsi (Databricks), lecture 4; Tuhin Srivastava (Baseten), lecture 7; Apoorv Agrawal (Altimeter Capital), The Economics of Generative AI: Two Years Later. Interest: Gerstner and Agrawal are both at Altimeter, whose Q1 2026 13F is 41.6% semiconductors. Ghodsi and Srivastava both sell the alternative to frontier models.
Firm-level evidence
Firm reorganization is what productivity looks like before it becomes macro data.
AI-native startups show one mechanism: smaller, denser, flatter teams. Ramp/Revelio adds another: high-intensity adopters can grow headcount when AI becomes a real workflow investment.
AI-native
25%
Smaller than non-AI startups in the same industry-cohort.
Talent mix
+13%
Engineer share is higher; entry-level and manager shares are each roughly 15% lower.
Adopters
+10.2%
Total headcount growth among high-intensity AI adopters over two years.
Entry rung
+12.0%
Entry-level headcount growth among high-intensity adopters; low-intensity adopters show no significant change.
Process channel
Product channel
AI changes how employees code, sell, serve, analyze, and coordinate inside the firm.
AI is embedded into what the firm sells, so some knowledge work moves into the product.
Tools can make a worker more productive.
Products can let a small team serve demand that once required a larger organization.
This is not yet aggregate proof. It is a clue about mechanism: AI value first appears as a changed firm boundary, changed team shape, or changed hiring path. The macro edge is beginning to move too: Americans filed 5.7 million business applications in 2025, while roughly 500,000 actual firms formed in 2023. Applications lead; firms and productivity follow.
Sources: Kim and Koning, AI-Native Firms; Marginal Revolution, AI-Native Firms; Rem Koning thread; Kharazian, Simon and Stevens, A New Look at AI’s Impact on Jobs (Ramp/Revelio, June 2026); Sydney Ember, America’s Enterprising Spirit Is Booming After Decades-Long Slump (New York Times, July 2026).
Where the two lenses meet
AI can enrich the world and disappoint its financiers.
That is the key idea. The economic lens can see a real general-purpose technology. The accounting lens can still ask whether the capital cycle earns its keep.
The economic lens sees...
The accounting lens sees...
Cheaper intelligence, wider access, consumer surplus, workflow redesign.
Capex, depreciation, power, financing, utilization, and required returns.
Value created.
Value captured.
The economy can get the railroads while the railway shareholders get the lesson. That historical analogy is imperfect, but the accounting warning is exactly right.
The question to leave with
When you hear an AI headline, which ledger changed - and what evidence would prove it?
Sources: GeometricInvestor, chain of returns; Sequoia, AI's $600B Question.
Further reading
A reading list from the clipping trail.
Use these to go deeper. I have grouped them by the job they do in the argument: capex ledgers, cost curves, productivity, labour, India, and the practical future of agents.
Value capture
AI Value Capture - The Shift to Model LabsSemiAnalysis on where the profit pool may migrate.
Training
The Rising Costs of Training Frontier AI ModelsWhy frontier training remains capital intensive.
India
Should the US Sell Advanced GPUs to China? An Indian PerspectiveA strategic view of chips, controls, and India's position.
Source note: This slide draws on the Obsidian clippings folder, especially "The AI Capex Ledger," "AI Value Capture," "The Labor Share Fell," "1,302 real-world gen AI use cases," and "Should the US Sell Advanced GPUs to China? An Indian Perspective."
Changelog appendix
What changed since the June version?
2026-08-04 - The stack became five layers, and the money got its own slide
What changed: "The stack as a ledger" was rebuilt from four rows to five, splitting token buyers into model labs, applications, and buyers-and-the-macro-economy. "How we'll do this" was edited at the same time so that its lens-two card names those same five layers explicitly; it had previously named four, and the two slides described different chains. Agrawal's quantification, which first sat on the ledger slide and overflowed it, now has a slide of its own, "One investor has put numbers on the layers": ~$225bn / ~$40bn / ~$20bn, or 79 / 14 / 7, against 87 / 10 / 3 two years earlier, with the assumed margins, the triple-counted revenue base, his own revised sizing eight weeks later, and Altimeter's semiconductor position all stated on the slide rather than buried in the source line. "So is anyone actually making money?" grew from three verdicts to five to match, adding Contested for model labs and Thin for applications. "The size of the bet" stopped saying the decisive number is unknown and now says nobody discloses it and one investor has estimated it.
Why: the opening slide and the ledger were describing different stacks, and a five-row chain plus a three-row profit ledger did not fit one screen in Present mode. Naming the two previously missing layers is also where the argument is most alive, because they are the two nobody is required to report.
2026-08-04 - Five repairs to the evidence
What changed: the two NVIDIA figures were reconciled - $60.4bn is the compute part of the same $75.2bn data-centre quarter, and networking is the $14.8bn difference. "Why the market wants to concentrate" gained its first concentration measure: open-weight models take 29% of Vercel gateway tokens on under 4% of spend, and Srivastava puts 90-95% of inference spend on frontier models. That cuts against the slide's own claim, and the slide now says so. "Depreciation" gained a counter-block: H100 prices recovered above launch, a B200 renewal was quoted at double, and Amazon books 5-6 year chip lives against a 30-plus year shell. The subscription slide retired a third-hand 5.7% margin-cliff figure and replaced it with Srivastava's on-the-record 0 to 40-70% climb. "Cheaper tokens" now carries the elasticity that proves it: 12-18% more tokens per 10% price cut, and unit inference cost down ~99% in two and a half years while H100 prices rose.
Why: each was a place where the deck asserted something it had not shown, or showed two numbers that a careful listener would read as a contradiction.
Sources: Apoorv Agrawal (Altimeter Capital), The Economics of Generative AI: Two Years Later, 1 April 2026, with his revised sizing in the Stanford Review, 28 May 2026; Stanford MS&E 435, Economics of the AI Supercycle, Spring 2026 - lectures by Agrawal, Gerstner and Madra, Lochmiller, Ghodsi and Srivastava; Vercel AI Gateway Production Index, July 2026; Andy Jassy, 2025 letter to shareholders; NVIDIA Q1 FY2027 results. Interest: Agrawal and Gerstner are both at Altimeter Capital, whose Q1 2026 13F reports $5.70bn of US-listed longs, 41.6% in semiconductors and 28.6% in NVIDIA. A 13F is long-only, US-listed, filed 45 days after quarter end, and excludes the entire private book, so it understates the interest rather than overstating it.
Changelog appendix
What changed since the June version?
2026-08-03 - The balance-sheet economics of a frontier pause
Added two Act III slides distinguishing the private instability of a unilateral pause from the economic possibility of a coordinated pause, then showing why a credible standstill must cover contracts and credit support as well as frontier training.
Why: the pause question is a direct application of the deck's two-lens method. Accounting explains the first-mover penalty; economics asks whether institutions can change it.
Source: A balance-sheet stress test of the frontier-AI pause (EconForEverybody, 29 July 2026).
Changelog appendix
What changed since the June version?
2026-07-21 - Open-weight market structure
Strengthened “Why the market wants to concentrate” to distinguish free weights from free inference and show where advantage migrates: cost per completed task, learning data, and sticky workflow harnesses.
Why: Thompson supplies the industrial-organization mechanism behind the deck's existing claim that open weights cap model-layer rents without eliminating scarcity or value capture.
Changelog appendix
What changed since the June version?
2026-07-19 - NYT/Census business-formation corroboration
Updated “Firm reorganization” with the bridge from firm-level mechanisms to the broader formation signal: 5.7 million applications in 2025, alongside roughly 500,000 realized business formations in 2023.
Why: the figures reinforce the shrinking-minimum-viable-firm hypothesis while keeping applications, actual firms, and aggregate productivity analytically separate.
2026-07-05 - Cordasco on what AI benefits cannot yet measure
Updated “Downstream - labour and productivity” and “Downstream - organization” to add legibility asymmetry: visible costs are measurable immediately, while the important benefits may require new practices before they can be named.
Why: this is the conceptual mechanism behind the deck's claim that tasks change first, while jobs, contracts, and productivity statistics move later.
2026-07-05 - Ramp/Revelio firm-level AI spending evidence
Updated “So is anyone actually making money?” and “AI-native firms” to add firm-level evidence on high-intensity AI adoption and headcount growth.
Why: this strengthens the buyer/economy ledger. The paper links observed AI spending to workforce records and finds high-intensity adopters grew total headcount 10.2% and entry-level headcount 12.0% over two years, while low-intensity adopters showed no significant change.
Sources: Sydney Ember, America’s Enterprising Spirit Is Booming After Decades-Long Slump; Carlo Cordasco, What we can’t measure about AI – yet; Kharazian, Simon and Stevens, A New Look at AI’s Impact on Jobs (Ramp/Revelio, June 2026).