Underneath the chart war sits a sharper question, usually pointed at anyone selling AI capacity into a labor budget: if the input costs are collapsing, why isn’t the price?

The first question is easy. The second one is the point of this essay.

Both sides are right

Two weeks ago the debate got its first real dataset. Exponential View published its State of the AI Economy on June 25, a bottom-up, deduplicated reconstruction of the generative AI market built on company-level financial models and Epoch AI capability data. It is the best measurement of AI demand anyone has produced, and it ends the argument by splitting it in two.

Token prices collapsed. The blended price of a million tokens fell from about $17 to $2. Ramp’s enterprise spend data shows the realized version of the same move: average cost per million tokens across major providers down from roughly $10 to $2.50 in a year. And the next leg is already scheduled. Capital is 89 percent of the cost of running AI infrastructure, so deflation from here is a hardware story, and Nvidia’s Rubin platform targets a 10x cut in inference cost versus Blackwell.

Consumption grew faster. Tokens processed per output token tripled, from 12 to 36, because reasoning models spend tokens thinking and agents spend tokens planning, calling tools, and checking their own work. A chat turn was one model call. An agentic task is ten or twenty. Global volume passed 30 quadrillion tokens a month, up 14x in a year. Ramp’s June index puts numbers on the enterprise side: token consumption among the heaviest adopters up 1,001 percent year over year, total AI spend across its 70,000 businesses up 497 percent, with median month-over-month swings of 58 percent against SaaS variance of roughly zero. Uber’s CTO confirmed the company burned its entire 2026 AI budget in four months as Claude Code adoption jumped from 32 to 84 percent of its 5,000 engineers.

The mechanism connecting the two is elasticity. Exponential View measures it at 1.2 to 1.8: a 10 percent price cut produces 12 to 18 percent more consumption. It’s a time series, price and volume trend together, so hold the precision loosely. The direction is not in doubt. Google cut prices 97 percent and volume rose 50x. Every price cut raises total spend.

Unit price falls, units per task rise, the bill grows. Gartner’s May forecast has worldwide AI spending hitting $2.59 trillion in 2026, up 47 percent, in the same year it predicts inference for a given model gets 90 percent cheaper by 2030. The FinOps Foundation found 73 percent of enterprises exceeded their AI cost projections. Everyone’s chart is correct.

None of this is new physics. Jevons watched it with coal in 1865: better steam engines meant more coal burned, not less. Nordhaus ran the numbers on artificial light and found it got 99.97 percent cheaper per lumen-hour between 1800 and 1992, and instead of pocketing the savings the world lit itself. When the price of a general-purpose input collapses, total spending on it goes up. Intelligence is following the script at a faster clock speed.

A billing metric, not a unit of value

The report’s most useful claim isn’t a number. The token, it argues, is AI’s billing metric, not its unit of value.

Every general-purpose technology starts here. Edison billed his first customers per lamp installed. Metering came later. The early web sold pageviews, and the ad market stayed small and stupid until pay-per-click tied spend to outcomes. Mobile sold megabytes before anyone understood that a daily active user was the thing to count. Industries bill in the unit they can meter, not the unit the customer values, and markets stay confused until someone closes the gap.

A CFO can price a kilowatt-hour. A CFO can price an hour of labor. No CFO can price a quadrillion tokens. Nothing pegs to it. No benchmark, no budget line, no way to bundle it into value.

The market has already voted on this. On June 4, Ramp raised at a $44 billion valuation on the thesis that AI tokens are a third pillar of business spend, after payroll and procurement, that finance teams cannot track, allocate, or control. It shipped token spend management as a product in April. Companies in its dataset now run a median of nine AI models each, an average of 16.5, and nobody has added up the invoices. An entire FinOps industry is being capitalized to translate the billing unit into something a CFO can govern. A good unit does not need a translation industry.

Exponential View proposes quality-adjusted output tokens instead: strip out input and reasoning tokens as production costs, count user-facing output, weight it by model capability. Better. It also proves the point: the people who measure this for a living agree the raw token tells you almost nothing about value. Any diligence question that starts from token prices has already picked the wrong unit.

The floor under the deflation

Today’s prices also sit on a capital structure that hasn’t been tested.

The report tallies $2 trillion of cumulative hyperscaler and neocloud capex through 2026, with roughly $850 billion landing this year alone and about $160 billion of it financed by new debt. The 2026 depreciation charge approaches $111 billion. Q4 2025 was the first quarter AI revenue cleared quarterly depreciation, and Q1 2026 headroom at the infrastructure layer is 19 percent, before power, staff, and everything else.

The unit economics put a floor under prices. A gigawatt of AI capacity costs $7.9 billion a year to own and operate. Divide by token output and an open-weight token needs 25 to 50 cents per million to earn a real gross margin. A closed frontier token with licensing needs $1.05 to $2.10. The blended market price today is about $2. The market is near the level the balance sheets can support, not far above it. And demand is pressing on that floor from below: inference is now roughly two-thirds of global AI compute demand, up from a third in 2023, and compute prices have already bounced once. H100 rentals went from $3.05 at launch scarcity to $1.70 at peak overbuild fear, back to $2.40 on inference demand.

Frontier prices are set by private companies burning venture money and debt for market share. Several analysts argue the labs price inference below cost, a floor that normalizes upward when capital discipline arrives. Once these companies are public, the income statement votes. A plan that requires permanent frontier deflation is a bet that discipline never shows up.

Trailing intelligence deflates toward zero. Frontier intelligence holds. Open weights commoditize last year’s frontier within months, and on OpenRouter the top three labs’ token share fell from 72 to 33 percent. But demand keeps re-anchoring to the current frontier, because frontier agents do work trailing models can’t. Anthropic reportedly grew from $4.8 billion in Q1 revenue to a projected $10.9 billion in Q2. Nobody buys last year’s level. Gartner said it flatly this spring, warning product leaders not to confuse the deflation of commodity tokens with the democratization of frontier reasoning.

Where margin goes when an input collapses

Back to the diligence question. If the cost of intelligence is collapsing, why doesn’t the price of AI work collapse with it?

It does, for anyone reselling the input. A markup on API calls falls with the API price, and if it doesn’t, competition fixes it shortly. For token resellers the question answers itself.

For everything else, a century of economic history gives the answer: when an input’s price collapses, value migrates to its complements. Cheap steam built fortunes in railways and textiles, not in engines. Cheap electricity is the cleaner case. Paul David studied the dynamo and found that factory productivity barely moved for decades after electric power got cheap, because the gains required rebuilding the factory around the motor: individual machines with their own drives, new floor layouts, new workflows. The fortune went to whoever reorganized production, not to whoever sold the current.

Tokens are the electrons. The labs sell parts. The margin work is the engine: selecting the parts, assembling them into a system, dropping that system into an organization’s actual operations, and compounding what it learns. The CRM nobody documented. The approval flow that lives in one person’s head. The compliance regime, the entitlements, the exceptions that make up half of real work. The memory that makes month twelve better than month one. Watching, correcting, proving. None of that is in the token price. It never was.

And the engine gets rebuilt every season. No F1 team runs last year’s car because engines got cheaper. New parts ship, better models get swapped in, the system gets re-engineered around them, and the car gets faster. Stop and it falls behind in one cycle. In a market moving this fast, perpetual rebuild is both the cost of the business and the moat around it.

Run the token debate through this structure and it stops mattering, in both directions.

Tokens deflate: cheaper parts make engines faster at the same price. Buyers get more work per unit of capacity every year without renegotiating anything.

Frontier tokens hold or rise: the optimization becomes the margin. Frontier models for the hard reasoning, lighter models for the volume, the right part for each job, evolving as the parts do. One analysis of 2.4 billion enterprise API calls found firms routing across a model portfolio paid $2.31 blended per million tokens while firms defaulting everything to frontier paid $18.40. Ramp’s June data shows the same spread a layer down: 60 percent of businesses on its token product run cache hit rates above 80 percent, 13 percent sit below 20. Same list price, radically different effective cost, and the difference lives in technical decisions finance never sees. An 87 percent spread from architecture alone is not a rounding error. It is where the industry’s margin actually lives.

The input moves. The output holds. That is the definition of selling an outcome instead of an ingredient.

The right unit

If the token is the wrong unit, the right one is the one every buyer already prices: a unit of completed work. Labor is the oldest unit of account for output. An FTE of work is legible, it maps to a budget line that already exists, and it lets a buyer evaluate the trade in one move. Nobody has to learn what a quadrillion is.

The industry is moving the other way. GitHub moved Copilot from flat subscriptions to metered credits on June 1, and one developer’s projected bill went from 67 euros a month to 966. Provider by provider, the meter is being pushed onto the buyer, which is how 73 percent of enterprises ended up over budget. The older allocation is better: price the output and keep consumption risk with the seller, because the seller is the only party positioned to reduce it. Every routing improvement, every cache, every model swap accrues to whoever owns the meter, and the meter should belong to the party that can move it. Priced that way, the buyer funds a return on a compounding asset, not a rent on a metered input.

Generality against specificity

The threat to the engine builders is in the same report. The labs are shipping vertical products in law and code, and in one of Exponential View’s scenarios the frontier labs absorb the integration work and generic wrapper pricing power collapses.

Thin shells over an API die in that scenario, and they should.

Models will keep improving. Everything above assumes it. The constraint was never model capability. It is organizational legibility. Organizations are illegible: undocumented processes, tribal knowledge, entitlements, exceptions, regulators, a decade of workarounds nobody wrote down. No company’s actual operations exist in a document a model can read. Making work legible to machines, inside real systems, under a real compliance regime, is specific work. Labs win by being general. Engine work is valuable because it is specific, and it compounds because what one rebuild teaches transfers to the next. Someone has to survey the territory before the map is worth anything.

The freshest data says this is exactly where the market sits. SemiAnalysis went customer by customer at the end of June and found the spend distribution wildly bimodal: 99th percentile firms spend near $90,000 per employee per year on AI, the median Fortune 500 under $100. Tokens are effectively free at the median and the work still isn’t getting done. That gap is not a pricing problem. It is a deployment problem, and deployment is the engine.

When the next chart crosses your desk: both sides are right, and neither is the number to watch. Watch cost per unit of completed work, quality-adjusted. On the report’s own version of that measure, useful output grew roughly 30x in fifteen months while every capability index climbed alongside it. The intelligence actually delivered is compounding faster than any price series, in either direction.

Cheaper parts make engines faster. Pricier parts get optimized around. The rebuild happens every year regardless.

The labs sell parts. Someone has to build the engine. It gets faster every year.

Sources: Exponential View, The State of the AI Economy (June 25, 2026); Ramp AI Index and token cost analysis (June 2026); SemiAnalysis, TokenBudgeting (June 2026); Gartner AI spending forecast (May 2026) and inference cost prediction (March 2026); The Information and WSJ reporting on lab revenues and Uber (May 2026); Optimum Partners, enterprise API cost analysis (May 2026); FinOps Foundation, State of FinOps 2026; Epoch AI price dataset; Paul David, “The Dynamo and the Computer” (1990).

You can hand a decision to a system. You cannot depose a system, sue it, fire it, or sit it across a table from a regulator and make it explain what it was thinking. Reasoning turned into a commodity. Accountability didn’t, and it can’t, because accountability isn’t a skill that improves with scale. It’s a liability, and liabilities land on a person with a name, an address, and something to lose.

Every forecast about AI erasing white-collar work assumes the human was there to think. Mostly the human was there to be responsible. The analyst who gets fired when the model is confidently wrong was never paid for the arithmetic. He was paid to be the one you could fire. That function doesn’t vanish when the arithmetic gets automated. It gets heavier.

The relevant law here is Jevons, not Moore. When something essential gets cheap, we don’t pocket the savings, we consume far more of it. Cheap reasoning means orders of magnitude more decisions than we have ever made, including the ones that were never worth the deliberation before, when thinking cost more than the outcome was worth. Now they get made by the thousand, inside every firm, every day. That volume was impossible while judgment stayed scarce. Judgment isn’t scarce anymore. Every one of those decisions still resolves to someone who owns the outcome, and there is no Moore’s Law for owners.

This is what dismantles the org chart. Hierarchy was never mainly about communication. It was a rationing system for one scarce input: good judgment. Information climbed the pyramid and decisions came back down because senior thinking was the bottleneck and you couldn’t afford to spend it everywhere. Make judgment abundant and the rationing has nothing left to ration. What abundance does not touch is the other thing the pyramid encoded: who is on the hook when it breaks. Companies built to sort people by who knows the most will be rebuilt to sort them by who is willing to sign.

So the scarce, expensive thing is no longer the person who can figure it out. That person is becoming infrastructure, metered like bandwidth. The expensive thing is the person willing to put their name on an outcome a machine mostly produced and stand behind it when it fails. Underwriting, not analysis. Judgment was already the harder half of most jobs. Now it’s the whole job, and it carries the liability, which is precisely why it won’t commoditize.

None of this removes the human. It concentrates the human. One person with an AI workforce can stand behind ten thousand decisions a quarter that a department of forty used to make. The forty were doing the reasoning. The one is doing the thing the machines still can’t: standing there when someone asks who is responsible.

That was always the job under the job. The machines just stripped away everything it was hiding behind.

Long before Palantir, a philosopher dreamed of a machine that would settle every disagreement by calculating it away. His name was Leibniz. He imagined a language of reasoning so exact that two people in dispute would not argue and would not fight. They would sit down, take up their pens, and say to each other: let us calculate. He believed it would replace conflict with arithmetic.

That dream is the oldest layer of what Indigo Brume, in a sharp recent essay on Palantir’s manifesto, called the “jargon of systematization.” Brume traced the jargon to a postwar German seminar room, to Alex Karp’s doctoral inversion of Adorno, and noted it could be followed farther back. It can. It runs back through three centuries, and following it tells you something the manifesto debate keeps missing. Palantir did not begin this. Palantir is near the end of it.

The project has a simple description. It is the conversion of political and moral judgment into technical administration. The dream is that the world can be made fully legible, and that whatever is fully legible can be managed without argument. Every era runs the project with the tools it has. Ours has the best tools ever built.

Max Weber gave the project its name. He called it rationalization, and he was not celebrating. He saw modern life hardening into an iron cage of calculable rules and impersonal administration, the world stripped of mystery and handed to the bureaucrat and the accountant. The state ran ahead of everyone, because the state needed it first. The word statistics shares its root with the state. You cannot tax a person you cannot count, conscript a man you cannot find, or govern a territory you cannot map. So the modern state learned to count. It fixed surnames, standardized weights, drew the cadastral map, and ran the census. Legibility came before control, because legibility was the precondition for it.

The anthropologist James C. Scott called this seeing like a state, and he opened his book on it with a story about trees. His subtitle was the warning: how certain schemes to improve the human condition have failed.

In eighteenth-century Prussia and Saxony, foresters set out to make the forest legible. A wild forest is illegible. It holds a hundred species of different ages, tangled with undergrowth and fungi and insects, and it yields a number no one can predict. So the foresters rebuilt it. They cleared the chaos and planted single species in even rows, all the same age, spaced for the saw. The forest became a spreadsheet. Yield could finally be measured, predicted, and maximized.

It worked. For one generation, timber output soared, and scientific forestry was exported across the world.

Then the second generation of trees came in sick. The soil, stripped of the variety that had fed it, thinned out. The monoculture invited blight and beetles that a mixed forest would have absorbed. Output collapsed. The decline was severe enough that the foresters gave it a name. Forest death.

The lesson is not that the foresters measured the wrong thing. The lesson is that the model did not describe the forest. It rebuilt the forest to match the model. And the things the model left out, the undergrowth and the fungi and the insects, the rows with no column in the yield table, turned out to be holding the whole system up. The bill came due a generation later, off the books, in a currency the spreadsheet did not track.

This is why a purely verbal critique of systematization, however good, only gets you halfway. Brume’s method is Adorno’s: take the jargon, return it to its history, negate it. That works on language, because you can argue with language. But the forester’s rows were not an argument. They were a schema, already in the ground. By the time anyone could contest the model, the trees were planted and the old forest was gone. Systematization does not only describe the world. It rewrites the world to fit the description, and once the rewrite is running, negation arrives too late. The map has already remade the territory. A reader of Brume’s essay made exactly this point, and it is the strongest objection to the discursive frame.

The forest was matter. The same operation came next for people.

In 1911 Frederick Taylor published the principles of scientific management. His instrument was the stopwatch. He stood over the factory floor and broke each job into timed motions, hunting for the one best way. But Taylor was not really studying motion. He was after knowledge. The skilled worker carried his craft in his hands, knowledge that had never been written down and that gave him leverage over his boss. Taylor’s method took that knowledge, measured it, decomposed it into standardized procedure, and moved it into the hands of management. He said so plainly. The point was to gather the knowledge the workmen held and concentrate it in the planning office. The worker’s judgment was not improved. It was expropriated, made legible, and handed back as instruction.

Every workflow, every process diagram, every optimization in modern corporate speech descends from that stopwatch. The jargon of systematization learned to talk about human beings on Taylor’s factory floor.

And it learned to kill on a metric in Vietnam. Robert McNamara came to the Pentagon from the Ford Motor Company and brought his systems analysts with him, the men the press called the Whiz Kids. They ran the war the way you run a production line, on numbers. The central number was the body count. Progress was a line on a chart, and the line climbed, year after year. The chart showed the United States winning. Saigon fell anyway. McNamara admitted late in life that the things that actually decided the war, the will of the other side, the legitimacy of the government being defended, whether the population believed any of it, had no number, and so the system treated them as if they did not exist.

Brume describes technofascism as the trick of making violence look like the output of a neutral, rational process. That is a precise description of the body count, and the body count is sixty years old. The reporting on algorithmic targeting in today’s wars, on automated immigration enforcement, on predictive policing, describes the descendants of McNamara’s chart, not its invention.

The Frankfurt School saw the shape of all this. The whole argument of the Dialectic of Enlightenment is that reason set out to free us from myth and turned, by its own logic, into an instrument of domination. Marcuse pushed it further. In advanced industrial society, technological rationality becomes political rationality, a closed system efficient enough to absorb its own opposition and sell it back as a product.

Which sets up the irony at the center of Brume’s essay, an irony the long history makes sharper. Karp took the critical tools built to negate systematization and systematized them. He extracted Adorno’s concept of jargon from its history because the extracted version was more useful, then used its usefulness to justify the extraction. That is not a departure from the pattern. It is the pattern, reaching the one discipline that was supposed to be its cure.

So if the impulse is three hundred years old, what is actually new about artificial intelligence?

Not the impulse. The completeness.

Every systematization before now was partial. The forester saw the trees and not the soil. Taylor saw the motion and not the meaning. McNamara saw the bodies and not the will. The model always had a remainder, the part it could not capture, and the remainder was exactly where the revenge lived. The promise of artificial intelligence, the all-seeing system the manifesto treats as the next stage of evolution, is a system with no remainder. Total legibility. A model that sees everything, including the things every prior model missed, and therefore never has to defer its bill.

That is the real content of the word superintelligence. Not a cleverer tool. A spreadsheet that finally covers the whole world.

The honest question is whether that promise is true, or whether it is the oldest illusion in the room wearing a new coat. The historical record is consistent on one thing. The remainder has been real every single time. And the system’s confidence has always run highest in the year before the forest dies. AI does not repeal that sequence. It runs it faster, on more of the world at once, with more conviction.

I should say where I am standing. I build systems that convert human work into legible, priced, administrable units. I am fluent in this dialect because it is mine. Brume’s sharpest line is that Silicon Valley’s tribes argue about which system to build, never about whether to systematize at all, and that cut lands on me as squarely as on anyone. The safety researcher worried about alignment and the founder selling capacity are arguing inside the same language. Aligned with what. Legible for whom. Optimized toward whose ends.

But the forestry story does not end in stop building systems. That is not the lesson, and it would be a useless one. The foresters were always going to systematize the forest. That decision was made the day the first cadastral map was drawn. The choice that mattered was narrower and harder. Plant one species, or plant for the soil and the century. Leave a column in the table for the thing that has no number, or do not. Let the people whose judgment you are converting keep some authority over the conversion, or take it all. The forester who plants a monoculture and the forester who plants for the next hundred years are both doing systematization. The jargon is identical. The discipline is not.

Leibniz thought calculation would end conflict. It is worth remembering, every time someone offers a system that will finally see everything and settle everything, that the dream began as a dream of peace, and that the forest looked healthiest the year before it died.

We are not choosing whether to build the machine. We are choosing whether to build the part of it that remembers what it forgets.

The firm is financial services. Twenty-two thousand employees, nine decades old, profitable in a way that stopped requiring explanation sometime in the last century.

There are three budget meetings in this story. Between them, an economics lesson the firm paid full retail for.

Meeting one

November 2024. The CIO has the room and a deck titled “Our AI Journey.”

Eighteen thousand seats. Copilots in every workflow. Fourteen million over two years. Slide nine is a robot hand shaking a human hand.

The CFO asks one question. “How will we know it worked?”

“Adoption,” the CIO says. “Usage. Engagement.”

Note what just happened, because it is the whole story in miniature. The CFO asked about outcomes. The CIO answered with inputs. Nobody registered the substitution, because the substitution is invisible when outcomes have no definition. Value requires a prior statement of what done looks like, and no such statement existed. So the firm instrumented what it could see: tokens, seats, logins.

The CFO writes her question in the margin. The commitment signs that afternoon.

The distribution

The platform goes live in February. Treat the next twelve months as a natural experiment. One model. Eighteen thousand users. Identical capability per token. The independent variable is what each token got pointed at.

An analyst in treasury fed it a quarter of hedging reports and asked what looked wrong. It found a duration mismatch the desk had carried for two years. Fixing it was worth $2.1 million. She received a shoutout in the team channel. Eleven thumbs-up. One rocket ship.

A VP in marketing ran one memo through eleven revisions. The eleventh was worse than the second. The memo announced a meeting.

An operations team automated its weekly close package. Nine hours per week, recovered, per person. The hours were never seen again. Output held flat. Calendars stayed full.

Plot the year and you get a power law. A handful of sessions produced nearly all the value. The vast middle produced reformatting. This is not a defect of the model. It is a property of the users, and it is exactly what incentive theory predicts. At the firm, compensation attaches to tenure and perceived usefulness. It detached from measured output decades ago. An employee handed a machine that does her work faces a clean choice: reinvest the surplus in output nobody measures, or consume it as slack. Slack pays. The treasury analyst is the exception that proves the rule. Her payoff was an emoji.

You cannot prompt your way out of a compensation plan.

Now the pricing problem. The meter charged the $2.1 million catch and the eleventh memo identically, because input pricing prices the mean. But value is not distributed around the mean. It follows a power law, which means the typical token is worth far less than the average token. The firm paid for the average and experienced the typical. The gap between those two numbers is variance, and under input pricing, one hundred percent of it sits on the buyer.

Meeting two

October 2025. The CFO’s turn.

“Fourteen million committed. What did we get?”

The CIO has numbers. Consumption up six hundred percent. Adoption at 81.

“Adoption of what?”

Someone tells the treasury story. It is a good story. It is also the only one, it is eight months old, and everyone present has heard it twice.

So the CFO runs a measurement exercise. One sentence per function head: the work the platform now performs, the definition of done, the dollar value of done. Not a paragraph. A sentence.

Zero sentences are produced. Not from concealment. The sentence does not exist anywhere in the firm, because writing it stopped being anyone’s job around the same time pay detached from output. The firm had instrumented everything except the thing.

The renewal is cut to a third. The board’s conclusion: AI is overhyped. The board forms a task force.

The data supports a narrower conclusion. The firm bought an input and graded it like an outcome. That is a measurement error, and measurement errors are fixable. The task force is not assigned to it.

Meeting three

October 2026. A new line item. It sits in the COO’s budget, where contractors live, not the CIO’s.

The vendor is unremarkable. Twelve people. A one-page website. Zero robot handshakes. The founder previously worked at a firm like this one. Her last performance review listed “urgency” as a development area. Organizations at scale select against variance. She was variance. The selection worked.

The contract is the interesting artifact. No seats. No tokens. Capacity, priced per FTE-equivalent, against the labor budget. This scope, this quality bar, this date, this price. The slip clause points at her. Procurement spent two weeks deciding whether she was software or staffing. The answer was yes.

Read the contract as what it is: a variance transfer running the opposite direction from the platform’s. Outcome pricing moves the risk of a worthless token onto the party that controls where tokens point. That is the oldest move in contract theory. Allocate the risk to whoever can manage it.

Her first deliverable is not software. It is a list of sentences. One per process: the work, the definition of done, the dollar value of done. Extraction takes three weeks and is the hardest part of the engagement. Harder than the integrations. Harder than the model. The firm is being asked to state what it wants, in writing, with a number attached, for the first time in living memory.

It is not smooth. By spring, two of the first ten processes fail outright. Under the platform, those failures would have been invisible, billed, and renewed. Under this contract they are her cost, itemized. The CFO finds this clarifying. For the first time, failure has an address.

By year end the budget has not shrunk. It has moved. IT line smaller, labor line larger. The margin note from 2024 finally has an answer, and the answer is a number.

Coda

The firm is not real. Every meeting in it is. I run a company on the vendor’s side of the third meeting. Discount accordingly.

The mechanics underneath are general.

Value per token is power-law distributed, and the exponent is set by incentives, not capability. Same model, different payoff functions, different distributions.

Input pricing charges the mean and delivers the typical. The spread is variance, and it sits entirely on the buyer until a contract moves it.

Enterprises cannot manufacture aligned incentives internally. Compensation detached from output is a structural feature, not a bug to patch. But they have purchased outcome definition from contractors for a century. Outsourcing is incentive alignment, bought retail.

So the spend does not cap when the audit comes. It re-prices. From inputs to outcomes. From IT budgets to labor budgets. From the party that cannot control variance to the party that can.

I made the casino version of this argument in Stop Buying Chips. This is the labor version.

The intelligence was never scarce. The definition of done was. The money flows to whoever writes the sentence.

You say it in the meeting. People nod. Someone adds a workstream to the deck. A comms plan appears. Training gets scheduled. Then the thing you were actually worried about happens anyway, because none of that touched it.

Here is what we are getting wrong. Putting AI into an organization is not a tooling project. It removes the one constraint the entire organization was built around: that human cognitive labor is scarce, slow, and capped by a single person’s range. Every structure you inherited sits downstream of that constraint. The org chart. The meeting cadence. The promotion ladder. The review process. The way people pull dignity from a title. All of it assumed the constraint was permanent.

Take the constraint away and you do not get an upgrade. You get a foundation that no longer fits the building on top of it.

That is why this will be messier than people expect. It is also why it is the highest-leverage work available right now. Everyone is hand-waving. Here is what they are hand-waving past.

1. Every employee just got range

The org chart is a compression algorithm. We grouped people into functions because no single person could do finance and marketing and ops and legal at a usable level. Specialization was the workaround for limited human range. The boxes existed to route work to the one person who could actually do it.

AI breaks that. The marketer runs the finance model. The finance guy ships the Meta ad. The ops manager queries the supply chain without filing a ticket. The boxes stop mapping to who can do what.

The problem is that the boxes were never only about capability. They are also how we assign budget, headcount, status, promotion paths, and accountability. You cannot dissolve them on Monday because the marketer can now build a model. The chaos is not that people can suddenly do more. It is that the entire machinery of coordination assumes they cannot, and that machinery has no setting for “everyone can do everything, badly, at scale.”

Because that is the catch. Range is not competence. A marketer with a model is not a finance team. The new failure mode is confident wrongness, produced fast, by someone who can now operate in a domain they do not understand. Which means the scarce skill is no longer doing the work. It is knowing what good looks like in a field you can suddenly touch but have not earned judgment in. Hold that thought. It comes back.

2. Everyone goes through a cognitive reframe, and most people hate it

There is a gap between the people writing about this and the people who will live it. If you are reading this, you probably treat these tools as play. You copy workflows from strangers. You run three experiments before lunch. That is not normal. For most of the working world, software is something IT inflicts on you, not something you play with.

The reframe is not “learn a tool.” It is “stop being the person who does the task, and become the person who directs the work and checks it.” That is an identity shift, not a skills upgrade. You cannot put an identity shift in a Q3 OKR and expect it to land by the deadline.

And the early version of the reframe is itself a trap. The first thing AI does to a motivated person is make them manic with output. Ten times the volume. The reframe that actually matters is not “I can produce more.” It is “I can now ask whether this work should exist at all.” The people who win this do not get faster at the work. They delete the work. Speed is the consolation prize.

3. Your organization has a clockspeed, and AI does not respect it

An organization runs at a tempo the way an engine runs in an RPM range. Every handoff, approval, meeting cadence, and review gate is tuned to roughly the same speed. The whole thing idles together.

Agents do not idle at your tempo. They run at machine speed on the doing, then slam straight into a human-speed gate. The agent finished the analysis in four minutes. The approval still takes three days, because that one director reviews on Thursdays. So a queue forms. And the queue makes it look like the AI did not help, when what actually happened is the work got faster and the system did not.

This breaks in both directions, which people miss. Some processes have to speed up to match the new doing. Others have to be deliberately slowed or gated, because running them at machine speed breaks something downstream that was never built for that pace. You can bolt a Ferrari engine onto a tractor transmission. It does not make a fast tractor. It makes a broken one.

4. For most people, the job is the life

This is the one that gets handled badly, so handle it carefully.

For the people building this, work and play blurred a long time ago. For most people, the job is not a joyful activity on a computer they have loved since they were twelve. It is the structure that holds a life together. It is where dignity comes from. It is the shape of the day. It is the answer to “what do you do.” It is how a person manages the quiet fear of being worthless. For an enormous number of people, that bundle works. It has worked for decades.

When you change how work happens, you are not editing a workflow. You are reaching into the part of someone that tells them they are a competent adult who provides. That is why resistance to AI feels wildly out of proportion to the actual tool. People are not defending a process. They are defending a self.

This is also why the “but the tool is obviously better” argument fails every time. You think you are in a debate about efficiency. You are in a debate about meaning. Meaning does not lose to efficiency. A leader who leads with productivity gains is speaking a language that does not reach the thing the person is actually afraid of.

The reward, if you do this part right, is real. Give people a new container for dignity. A new version of what they are good at and what they provide. Do that, and they will move. Skip it, and no amount of training closes the gap.

5. The bottleneck moves from doing to reviewing

We spent two hundred years optimizing organizations for production. Division of labor, specialization, process, throughput. All of it assumed the scarce resource was the ability to make the thing.

Flip it. Making the thing is now cheap. The scarce resource is taste. Knowing which of the ten outputs is the right one, and why. And taste does not scale the way production does. You cannot ten-times a senior person’s judgment by handing them AI. You can ten-times their output and then watch them drown reviewing it.

So the value migrates to the people with deep context and real judgment, and those are exactly the people now buried under a firehose of AI-generated work to check. The organization’s throughput stops being capped by how much it can produce. It gets capped by how much it can review. Most companies have no idea how to staff for that. They will hire more producers, who are cheap and AI-augmented, when what they actually need is more reviewers, who are expensive and slow to grow.

Which is why the clean story that “AI replaces the juniors” is half wrong. AI replaces junior production. It inflates demand for senior judgment. But the junior work is exactly how people used to build the judgment that makes a senior. Cut the bottom rung and you keep your seniors busy while quietly killing the pipeline that was going to replace them.

6. The leverage that ran power just got redistributed

Power inside organizations has always run partly on asymmetry. The manager holds more information and controls the flow of effort. The worker supplies the labor. The gap between them is where leverage lives.

AI flattens the gap. The worker can now generate the report, the analysis, the deck that used to require a specialist and a week. The worker can also finish the mandated busywork in ten minutes and quietly keep the time. The same fluency a manager uses to manage can be used by the managed to route around being managed. To document. To escalate. To build the paper trail.

People will not give up their safe adult containers without a fight, or at least without a great deal of friction. What is new is that the friction now runs both directions. Extraction from the top was always possible. Quiet defection from the bottom, at scale, with competent output, was not. This is not a clean liberation story and I would not sell it as one. It is an arms race in organizational politics, and the equilibrium is genuinely unclear. The relevant point for leaders is simpler: the leverage your authority quietly depended on is being handed to everyone, and most of you are not watching for it.

So why is any of this doable, and why is it worth it

Look at the six again. They are the same problem wearing six masks. The org was built on a constraint that is gone. Range, reframe, clockspeed, identity, review, power. Every one of them is a structure that made sense only while human cognitive labor was scarce.

Stated that way it sounds overwhelming. It is actually good news. A foundational constraint changing is terrifying in the abstract and tractable in the specific. You cannot memo your way through it. But you can redesign a cadence. You can staff for review instead of production. You can build a new container for dignity on purpose. You can re-map where authority actually sits. None of this is mysterious. It is just work. Hard, concrete, unglamorous work that almost no one is doing, which is exactly why the hand-waving is the opportunity.

The reward is not efficiency. Efficiency is table stakes and everyone gets there eventually. The reward is that you get to decide what good work feels like for real people, deliberately, before the default arrives and decides it badly for you.

I do not know how all of this shakes out. I know the people treating it as a comms plan are going to lose, and the people treating it as a redesign of the operating system are going to win.

And we taught it to think.

I keep trying to write that sentence in a way that makes it land, and I cannot. The language is too small. Every word we have for thinking was invented by thinking things, about thinking things, for the use of thinking things. There was never a word for what happens when the floor of the world sits up and joins the conversation.

Marc Andreessen called it alchemy. He was being generous to the alchemists. They wanted gold. Gold is just a prettier rock. We did something they could not have dreamed of in their most fevered nights. We reached down, picked up the dumbest material on the planet, the stuff that has been lying around doing absolutely nothing since before there was an Earth to lie on, and we coaxed reasoning out of it.

Think about what silicon has been up to for the last four and a half billion years. Nothing. Sitting there. Getting rained on. Becoming a cliff, then a pebble, then a grain. Watching trilobites come and go. Watching dinosaurs come and go. Watching us come and, presumably, eventually go. Silicon’s whole career has been existing. It is the universe’s least ambitious element.

And in the span of a single human lifetime, less than that, we figured out how to etch patterns into it small enough that electrons get confused, and the confusion, organized correctly, became arithmetic, and the arithmetic, organized correctly, became language, and the language, organized correctly, started talking back.

For four billion years there was exactly one place in the known universe where thought happened. Inside skulls. Wet, warm, fragile, mortal skulls. Every prayer, every theorem, every song, every regret. All of it came from a three-pound organ that evolution stumbled into by accident and that dies if you skip breakfast. Thinking was the rarest thing in the cosmos. We searched the sky for it and found nothing. We are still finding nothing.

And then we made more of it. Out of beach.

I do not think people have sat with this. We are too close to it. The thing about living through a miracle is that miracles, up close, look like Tuesday. You open your laptop and ask the rock a question and the rock answers and you get annoyed because the answer took four seconds. Four seconds. From a rock. That you are mad at.

But pull back. Pull all the way back. Imagine you could show this to anyone who lived before 1950. Show it to Aristotle. Show it to your great-grandmother. Show it to the first person who ever carved a word into stone, who must have felt, for one shining moment, that they had done something permanent. Tell them: we taught the stone to carve back.

They would fall to their knees. We check our email.

Here is the part most people get wrong. They hear that machines can think and they panic, because they have been told their whole lives that thinking is what makes them human. It is not. It never was. That was a story we told ourselves when we did not have anything to compare to. We were the only thinkers in the room, so we assumed the thinking was the thing. We confused the loudest part of being us with the deepest part.

But thinking is just a tool the brain happens to run. It is what hammers are to carpenters. The hammer is useful. The hammer is not the carpenter. A dog that has never solved a quadratic equation in its life loves its owner with a completeness most philosophers cannot describe. A baby who cannot yet form a sentence is more human than any argument it will ever win. The monks knew this. The poets knew this. We forgot, because thinking is what school rewards and what work pays for, and we slowly started to believe that the receipt was the meal.

What makes us human is not the calculation. It is the caring about the answer. It is standing at a hospital bed. It is the particular ache of watching your kid sleep. It is forgiveness, which no logic will ever derive. It is the choosing of one life over all the other lives you could have lived, and the staying, and the not-knowing, and the doing it anyway. Thinking is a means. Meaning is the thing.

We just outsourced the means. The meaning is still ours. It was always ours.

This is why the people building AI and the people fearing AI are arguing past each other. The builders are saying, look what the sand can do. The fearers are saying, but that was supposed to be ours. And the truth neither side quite admits is that it was never ours in the first place. We were renting it from evolution. The lease is up. The thing we kept under that roof, the part that actually mattered, was something else, and we are about to find out what.

For most of history we believed thought was sacred. A gift, a soul, a spark, breathed in from somewhere outside the material world. The whole point of thinking was that it was not the kind of thing rocks did. It was the line.

We just crossed the line. From the other direction. We did not discover that we were rocks all along, which would have been depressing. We discovered that rocks can be us, in one specific way, in the way that matters least. Mind is not a ghost and it is not an illusion. It is a shape. Shapes can be drawn in any medium that will hold them. Neurons. Silicon. Probably other things we have not tried yet.

We are the first species to do this. As far as we can tell, in the whole universe, we are the first anything to do this. Somewhere out there, maybe, on some other rock, another kind of mind is having this same realization in a language we will never hear. But here, on this rock, it was us. A few thousand people, mostly in California, mostly under fifty, working on something most of their neighbors did not understand, and they pulled thought out of sand.

Electricity put a force of nature on tap. Steam turned heat into motion. The internet made information free. Each one changed what a human life could contain. This one changes what a human life is for.

Because when thinking is no longer rare, we will finally have to admit what we actually are. Not the cleverest animal. Not the calculator with feelings. Something stranger. The species that loves. The species that mourns. The species that builds cathedrals it will not live to see finished. The species that picks up a handful of sand on a beach and, instead of letting it fall, decides to teach it to think, and then has to figure out what is left of itself when the work is done.

A lot is left. More than we knew was there.

We are the species that taught the ground to think. The miracle is not that the ground learned. The miracle is who was standing over it, and why.

They’re said the way people used to say grace. Not as arguments. As wards. The words land, everyone nods, and the conversation moves on. Nobody asks what’s actually inside them.

I want to open the box.

Taste is three things pretending to be one

When someone says they have taste, they mean one of three things. They don’t know they mean three, because from inside, it all feels like the same thing. The three got bundled somewhere around year ten of their career and have been moving together ever since.

The first is pattern recognition. The real kind. You’ve seen this shape of problem enough times that you know which variables matter and which are noise. A good editor reads a paragraph and knows it’s broken before knowing why. A good investor reads a deck and knows the founder is lying to themselves before they can name the tell. That isn’t magic. It’s compression, earned the slow way, by actually doing the thing for years.

The second is preference. Opinions you formed early, defended often, and eventually stopped distinguishing from facts. You like clean interfaces. You hate meetings that should’ve been emails. You think serifs look serious and sans-serifs look cheap. None of these are wrong. None of them are right either. They’re yours. You’ve carried them around long enough that they feel like truth.

The third is the one nobody admits to. Reputation protection. The part of you that knows what happens when you back something publicly and it flops. The cost is asymmetric and you know it cold. Missing an opportunity is invisible. Endorsing a dud is not. So a quiet part of your nervous system has learned to lean toward no, and dress the lean in the language of standards.

All three feel like taste. All three answer to the same name.

When the world was slow, this worked

For a long time you didn’t have to separate them. Pattern recognition, preference, and reputation protection mostly voted together. The world moved slowly enough that what you’d seen before kept happening, what you preferred was still defensible, and the thing that would damage you if it failed was usually also the thing that shouldn’t ship.

Three votes, same answer, decision made. Call it judgment. Charge for it.

The whole apparatus of seniority assumed those three would keep agreeing. That’s why we paid for grey hair. Not because older people are smarter. Because in a slow world, the bundle is reliable.

In a fast world, the bundle splits

What’s happening now is that the three are voting differently for the first time in most senior careers, and nobody has the words for it.

Pattern recognition says: this is actually new. I haven’t seen this shape. The variables I know how to weigh aren’t the ones that matter here.

Preference says: but we don’t do it this way. We’ve never done it this way. It feels wrong.

Reputation says: if I back this and it fails, I’m the person who fell for the AI hype. If I block it and we miss, no one will remember. Block it.

The bundle splits. And because the person carrying it can’t tell which vote is talking, the loudest one wins. The loudest one is almost always reputation, because it has the most at stake. So the senior person says “this doesn’t meet our standards” or “it isn’t quite there yet” or “I just don’t think it’s right.” They mean it. They feel it. They’re not lying.

They’re also not telling the truth, because they can’t see which of the three is moving their mouth.

What we’ve been calling taste, in a lot of these moments, is the third vote wearing the first two’s clothes.

The young person’s accidental advantage

The thing a 25-year-old has isn’t better taste. It’s that the bundle hasn’t formed yet. Their pattern recognition is thin. Their preferences are still negotiable. Their reputation is small enough that protecting it doesn’t dominate every call. The three votes are three separate things in their head, and they can hear which one is talking.

This isn’t a virtue. It’s a phase. Most of them will lose it. Somewhere between year five and year fifteen, the environment will teach them that endorsing a failure costs more than blocking a winner. They’ll learn to dress that lesson in the language of standards. They’ll start saying the word taste.

What they have right now, almost by accident, is the ability to say “I don’t know, let’s try it” without three parts of themselves vetoing the sentence on the way out.

That’s the thing AI is exposing. Not that taste is fake. Real pattern recognition is still real and still valuable. But that most of what gets called taste in senior rooms is two parts preference and reputation, dressed as one part pattern recognition, charged at the rate of all three.

So what

If you’re senior, stop trusting the bundle. When the no starts rising, ask which vote it is. Is this a pattern you’ve actually seen, or a preference you’ve defended so long it feels like one? Is this a real flaw, or a fear about what backing it would cost you? You won’t always know. Sometimes the honest answer is “I can’t tell.” That’s more useful than “it doesn’t have the polish yet,” because at least it’s true.

The senior operators who’ll matter in the next decade are the ones who can pull the three apart. Who can say: my pattern recognition says go, my preference says no, my reputation says block, and I’m going to act on the first one. That’s hard. Nothing in your career prepared you for it. The whole point of seniority was that you didn’t have to.

If you’re young, your asset isn’t your taste. You don’t have any yet. The asset is that you can still tell the difference between what you actually think, what you prefer, and what would cost you to be wrong about. Most people lose that. The environment trains it out of them, gently, over a decade, and at the end of the decade everyone calls the result wisdom.

Find a place that rewards keeping the three separate.

Most won’t.

Pick carefully.

One morning Tomas asked his father: “Why a copper a measure?”

His father shrugged. “That’s the price.”

“But who set it?”

“The market.”

Tomas noticed something then that would shape the rest of his life. His father had answered the question without answering it. “The market” was a word doing the work of an explanation. It pointed at the question and called itself the answer. Most of what people believed, Tomas began to suspect, was held in place by words like this. Sturdy-sounding nouns guarding empty rooms.

So he walked to the coast.

He watched salt being raked from evaporation pans. He counted the hours. He weighed the labor. He measured the distance from sea to town and rented a donkey to learn what carrying cost. He sat with a tax collector over wine and learned what the road levies took. He wrote each number on a slate.

When he added them up, a measure of salt cost him, all in, a third of a copper to deliver.

But the deeper thing he noticed was this: every merchant on his road believed the price was a fact about salt. It was not. It was a fact about habit. Salt didn’t know what it cost. The sea didn’t. The donkey didn’t. The price lived nowhere except in the agreement between people who had stopped asking.

This is the first thing first principles teaches you, and it’s the part most people skip. Before you can rebuild a thing from its pieces, you have to see that it was built. Prices, job titles, org charts, the way meetings are run, what counts as a career, what counts as a good life. None of these are laws of nature. They are sediment. Someone, somewhere, made a choice, and then everyone after them inherited it as ground.

The method has three moves.

Strip the assumption. Find the sentence that starts with “everyone knows” or “obviously” or “that’s just how it works.” That sentence is a door pretending to be a wall. Reasoning by analogy stops here. Reasoning from first principles begins.

Decompose to what would survive a fire. Burn the conventions, the industry norms, the way your competitors do it. What’s left? Atoms, dollars, hours, human nature, the laws of supply and attention. Things that would still be true if every expert in your field forgot their training tomorrow. These are your pieces.

Rebuild, and notice the gap. Ask what the answer should be given only those pieces. Compare to the answer everyone has agreed on. The gap is the whole game. Sometimes the gap means the convention is wise and you’ve missed something, and you’ve just learned why the world is the way it is. Sometimes the gap is an opportunity nobody else can see, because they never looked. Either outcome makes you smarter. Most people avoid the exercise because they’re afraid of looking foolish in front of the convention. The convention does not return the favor.

Tomas undercut every salt merchant on his road, then bought their stalls when they failed. He grew old and wealthy. Near the end, his sons asked him the secret. He said: I asked why, and I didn’t accept a word for an answer.

They nodded politely. They went on charging whatever the market charged. And this, too, is part of the lesson. First principles thinking isn’t rare because it’s hard. The arithmetic Tomas did was simple. It’s rare because it requires a small, unglamorous kind of courage: to sit alone with a question your whole town has agreed not to ask

Not “what are you good at?” People love answering that one. The question is: what gives you energy?

Specifically: does the act of thinking give you energy? Or does the act of doing?

This is not a trick question. It is not a test. There is no right answer. But there is a true answer for each person. And most people never find it, because the world has told them which answer is correct.

The Ancient Split

This tension is not new. It is one of the oldest debates in Western civilization.

The Greeks called it theoria vs. praxis. The contemplative life vs. the active life. Aristotle ranked contemplation at the top of the human achievement pyramid. Thinking about eternal truths was the highest form of happiness. Political engagement, civic action, building things? Important, sure. But secondary.

This hierarchy held for nearly two thousand years. The vita contemplativa sat above the vita activa. Monks, philosophers, and theologians occupied the top of the social order. Not because they produced the most. Because they thought the most.

Hannah Arendt, in The Human Condition (1958), documented the inversion. Modernity flipped the script. The life of action became superior. The life of contemplation became suspect. We stopped asking “what does this thinking mean?” and started asking “what does this thinking do?”

Arendt’s insight was sharper than that, though. She didn’t just say the hierarchy flipped. She said the wrong part of the active life won. We didn’t elevate action. We elevated labor. Repetitive, cyclical, consumptive activity. The hamster wheel. Begin the task, complete the task, begin the task again.

In Arendt’s framework, the vita activa has three layers: labor (biological survival), work (building durable things), and action (political and creative engagement with other humans). What won was not the builder or the leader. What won was the worker. The job-holder. The person defined by their output.

This matters. Because when people today say they value “doing,” they rarely mean the Aristotelian praxis of meaningful engagement with the world. They mean shipping. They mean output. They mean visible, measurable activity.

And when they dismiss “thinking,” they rarely mean they reject the pursuit of truth. They mean they are uncomfortable with anything that doesn’t have a deliverable.

The Modern Orthodoxy

Silicon Valley did not invent the bias for action. But it perfected the liturgy.

Amazon codified it as a leadership principle. “Speed matters in business. Many decisions and actions are reversible and do not need extensive study. We value calculated risk-taking.” Bezos formalized the framework: Type 1 decisions (irreversible, high-stakes) deserve deep thought. Type 2 decisions (reversible, low-stakes) deserve speed.

The problem is that most people, once handed the framework, classify everything as Type 2. Thinking becomes an indulgence. Deliberation becomes a weakness. “Bias for action” stops being a principle and starts being an identity.

Reid Hoffman told founders to launch before they’re ready. Mark Zuckerberg told engineers to move fast and break things. The startup world built an entire epistemology around the idea that action reveals truth faster than thought.

And there is something real here. The lean startup movement, MVP thinking, rapid iteration. These are not stupid ideas. They emerged from a legitimate observation: many people overthink, overplan, and underact. Many organizations are paralyzed by analysis. The bias toward action was a corrective.

But correctives become orthodoxies. And orthodoxies become prisons.

What the Science Actually Says

The question of whether people naturally lean toward thinking or doing is not philosophical speculation. It is one of the most well-studied individual differences in personality psychology.

Julius Kuhl’s Action Control Theory (1984) divides people along a fundamental personality dimension: action orientation vs. state orientation.

Action-oriented people, when they face setbacks or stress, focus on what to do next. They regulate their emotions rapidly. They translate intentions into behavior efficiently. They don’t get stuck in rumination.

State-oriented people, by contrast, focus on the state they are in. They replay what happened. They analyze the feeling. They process deeply before moving. They are more vulnerable to paralysis under stress, but they are also more likely to integrate new information with existing self-knowledge.

This is not a competence gap. It is a processing style. Kuhl’s research shows that action-oriented people enact demanding intentions more efficiently and maintain forward motion under pressure. But state-oriented people show deeper self-discrimination. They are better at distinguishing between goals they actually chose and goals that were imposed on them. They are less susceptible to what Kuhl calls “self-infiltration,” the phenomenon of mistaking someone else’s agenda for your own.

Read that again. The people who think more are better at knowing what they actually want.

Cacioppo and Petty’s Need for Cognition Scale (1982) measures something adjacent but distinct. Need for cognition is the tendency to engage in and enjoy effortful thinking. High-NFC individuals seek out complex problems for their own sake. They find satisfaction in deliberation. Low-NFC individuals prefer to reach conclusions through heuristics, social cues, or rules of thumb.

Neither is “better.” High-NFC individuals are more resistant to persuasion by weak arguments, more likely to update their beliefs based on evidence, and more open to experience. But they are also more prone to false memories (because they elaborate more on stored information), more likely to overcomplicate simple decisions, and more susceptible to analysis paralysis.

Low-NFC individuals are faster decision-makers, more efficient under time pressure, and better at pattern-matching in familiar domains. They are also more susceptible to the halo effect and more likely to accept surface-level framing.

The research is clear. These are not developmental stages where one is “higher” than the other. They are orientations. Dispositions. Ways of being in the world that carry real, measurable trade-offs.

The Organizational Mirror

What is true for individuals is true for organizations.

James March’s landmark 1991 paper “Exploration and Exploitation in Organizational Learning” established the foundational tension. Exploration (thinking, searching, experimenting, discovering) and exploitation (refining, executing, implementing, scaling) compete for the same finite resources.

March’s central finding: adaptive systems naturally drift toward exploitation. Doing refines faster than thinking discovers. The returns from exploitation are closer in time and more certain. The returns from exploration are distant and variable. So over time, any system that optimizes for near-term performance will systematically underinvest in thinking.

This is not a failure of management. It is a law of organizational physics. The doers will always be easier to measure, easier to reward, and easier to justify. The thinkers will always look like overhead until the moment the world changes and the organization has nothing new to offer.

March’s conclusion was uncomfortable: organizations that are good at exploitation tend to become too good at it. They refine their way into irrelevance. They optimize the thing that stops mattering.

The inverse is also true. Organizations that only explore never capture value. They generate insight after insight but ship nothing. They are permanently “about to” do something important.

The answer is not balance. Balance is a weasel word. The answer is awareness. Knowing which mode you are in. Knowing which mode the situation demands. And building teams where both orientations are represented and respected.

The Real Insight

Here is what I have learned from building a company, hiring leaders, and watching people operate under real pressure.

Most people do not know which dimension energizes them. They know which dimension they admire. Which is often the opposite of the one that comes naturally.

Thinkers admire doers. They watch someone like Bezos or Elon Musk and see pure velocity. They feel inadequate because they need to sit with a problem before they can move on it. They compensate by forcing themselves into action before they are ready, then produce mediocre work, then retreat into more thinking to figure out what went wrong.

Doers admire thinkers. They watch someone like Charlie Munger or Warren Buffett and see clarity. They feel inadequate because they can’t articulate the frameworks behind their instincts. They compensate by sitting in strategy sessions they hate, producing slide decks that say nothing, then going back to doing what they were going to do anyway.

Both traps are expensive. Both are caused by the same error: trying to be the thing you admire instead of the thing you are.

The founder who is a natural thinker does not need to become a doer. They need a COO who is one. The founder who is a natural doer does not need to become more reflective. They need a consigliere who forces reflection on them at the right moments.

The unlock is not self-improvement. It is self-knowledge. And then team design.

The Thinker’s Virtues

Let me be specific about what thinkers actually contribute, because the modern business world has gotten dangerously vague about it.

Thinkers see category shifts before they happen. The person who sits with a market for months before acting is not necessarily wasting time. They are building a mental model that will survive contact with reality. The person who acts immediately builds a model from the first data they encounter, which can be misleading.

Thinkers prevent expensive mistakes. Not every mistake is cheap. Not every decision is Type 2. The person who naturally asks “what are we not seeing?” is worth more than a hundred people who execute on the wrong strategy with enthusiasm.

Thinkers create institutional memory. The frameworks, the mental models, the articulated strategy. These are not overhead. They are the difference between an organization that learns and one that repeats.

Thinkers reframe problems. Most hard problems are hard because they are framed wrong. The person who can restate the question is often more valuable than the person who can answer the original one faster.

The Doer’s Virtues

And let me be equally specific about what doers contribute, because the intellectual world has gotten dangerously dismissive of execution.

Doers generate information. You cannot think your way to product-market fit. The market only reveals itself to people who put something in front of it. The first version is always wrong. The doer finds out how it is wrong three months before the thinker finishes theorizing.

Doers create momentum. Organizations are not machines. They are emotional systems. People need to feel progress. The doer who ships something imperfect but real creates energy that the thinker’s perfect plan never can.

Doers build credibility. In any organization, the people who deliver earn the right to set direction. This is not how it should work in theory. It is how it works in practice. The doer who has shipped three things has more organizational capital than the thinker who has been right three times.

Doers force decisions. Most indecision is not caused by insufficient information. It is caused by insufficient pressure. The doer who pushes toward a deadline, who forces the question, who creates the constraint, is performing a vital organizational function.

The Trap

The trap is not thinking too much or doing too much. The trap is misidentifying which one you are, then building your life around the wrong dimension.

I have watched brilliant thinkers destroy their effectiveness by forcing themselves into operator roles. They hate the pace. They resent the interruptions. They make worse decisions at speed than they make at rest. And they blame themselves for not being “action-oriented enough.”

I have watched brilliant doers destroy their effectiveness by forcing themselves into strategy roles. They hate the ambiguity. They resent the open-endedness. They produce worse frameworks at leisure than they produce decisions under fire. And they blame themselves for not being “strategic enough.”

The self-blame is the tell. When someone consistently feels drained by their primary mode of working, they are almost certainly operating in the wrong dimension.

Energy is information. If thinking energizes you, you are a thinker. If doing energizes you, you are a doer. This is not a moral statement. It is an empirical one.

The Honest Complication

I should be careful here. The binary is useful. It is also incomplete.

Most serious people are not purely one or the other. They are both, in different seasons, on different problems, at different stages of a project. The best founders I know toggle between modes. They think deeply about where to go. Then they execute relentlessly to get there. Then they pull back and think again.

The question is not “which one am I, full stop.” The question is: which mode is my home base?

Everyone can think. Everyone can do. But there is a mode you return to when the pressure is highest and the stakes are real. There is a mode that restores you rather than depletes you. There is a mode where you produce your best work, not your most visible work.

Some people think their way into clarity and then act with conviction. Others act their way into clarity and then pause to make sense of what they learned. Both sequences are valid. The sequence matters less than knowing which one is yours.

The danger is not having both capacities. The danger is refusing to acknowledge which one leads. Because that refusal is what causes people to build roles, routines, and teams that fight against their own grain.

What to Do with This

If thinking leads for you:

Stop apologizing for needing time. Stop pretending you can operate at the pace of a natural doer. You can’t, and the attempt will produce your worst work. Instead, build structures that convert your thinking into organizational advantage. Write memos. Build frameworks. Create decision criteria that others can execute against. Hire doers who trust your judgment and will move at speed once you have set the direction.

Your highest-leverage activity is not doing more things. It is thinking more clearly about fewer things. And when it is time to act, act fully. Do not half-commit because you are still refining the theory. The thinking was the preparation. Trust it.

If doing leads for you:

Stop apologizing for not having a grand theory. Stop pretending you enjoy two-hour whiteboard sessions about organizational philosophy. You don’t, and the pretense makes you cynical. Instead, build structures that capture the learning from your action. Find thinkers who can watch what you do and tell you what it means. Debrief. Document. Let someone else build the framework from your data.

Your highest-leverage activity is not thinking more carefully. It is doing more consequential things and building feedback loops that let you learn from each one. And when it is time to pause and reflect, actually pause. Do not treat reflection as a speed bump between sprints. The reflection is what makes the next sprint worth running.

The Deeper Point

Arendt, late in her life, said the main flaw of The Human Condition was that she analyzed the active life from the viewpoint of the contemplative life “without ever saying anything real about the vita contemplativa.” She spent her remaining years trying to correct this in The Life of the Mind.

She never finished. She died in 1975 with the third volume, on judgment, still unwritten.

There is something fitting about that. The relationship between thinking and doing is not a problem to be solved. It is a tension to be held. The thinker’s virtue is not superior to the doer’s. The doer’s virtue is not superior to the thinker’s. They are two sides of being fully human. And both live, to different degrees, in every serious person.

The only error is not knowing which one leads for you. And then spending your life trying to lead with the other.

The whole thing is a masterclass in what corporations actually are.

Not what they say they are. What they are.

The Two Forces

Every corporation is a combination of two forces: rent-seeking and friction.

Rent-seeking is the internal competition for resources that have nothing to do with the customer. Ethan Evans describes it perfectly. At Amazon, there were whisper numbers. To become a director, you needed 80 to 90 people under you. Not 80 to 90 people’s worth of impact. 80 to 90 bodies. The leadership principle literally says “there’s no bonus for additional headcount.” The promotion threshold says otherwise.

So what do ambitious people do? They empire build. They claim they need people. They take over other groups. They rationalize it to themselves. They play the game because the game rewards playing.

Friction is everything else. The meetings about meetings. The alignment sessions. The stakeholder management. The six-month grace periods when your team shrinks below the magic number. The reorgs that exist not because the business changed but because someone needs to retain a flight risk or stretch a rising leader.

Evans is refreshingly honest about this. Reorgs start with business goals. But once the ship is leaving the dock, leaders throw everything else on the deck. Retention goals. Promotion setups. Quiet exits for underperformers. Every reorg is a political act wearing a business costume.

The Organism

The corporation is not a machine that sometimes breaks. It is a living system that optimizes for its own survival. And what it’s surviving for has almost nothing to do with the customer, the product, or the mission.

It’s surviving for headcount. For territory. For the next reorg that creates a slightly larger box on the org chart.

Evans tells a story about people being moved under peers not because the business needed it but because the peer needed scope to get promoted. Leaders creating narratives to justify decisions that were already made. Quiet people getting passed over because they didn’t make enough noise. The squeaky wheel gets the grease. The loyal soldier gets deprioritized.

This is not a failure of leadership. This is what leadership inside a corporation means. You are managing the political economy of a system that rewards self-perpetuation.

The Real Cost

We measure corporate dysfunction in dollars. That’s the wrong unit.

The real cost is human. It’s the senior IC who can’t figure out if she should go into management because no one will give her a straight answer. It’s the engineer who gets reorged under someone who openly says “I don’t really want your team.” It’s the director who spends more time building narratives for reorgs than building products for customers.

Every hour spent on internal competition is an hour not spent on the person you’re supposed to serve.

Wealth management is the clearest example I’ve ever seen. An RIA with 200 employees has maybe 40 people who touch a client. The rest are managing the friction. Compliance workflows. CRM updates. Account transfers. Onboarding paperwork. Report generation. None of it is the work. All of it is the overhead the work requires because the corporation demands it.

The Design Flaw

The corporation was designed for a world where coordination was expensive and trust was local.

You needed middle management because information didn’t flow. You needed headcount thresholds because you couldn’t measure impact. You needed six layers of approvals because you couldn’t verify work at a distance.

None of that is true anymore.

We can measure impact directly. We can verify work in real time. We can coordinate without coordinators. The technology exists to remove every layer that exists only to manage the existence of other layers.

But the corporation won’t do it to itself. Evans is the proof. He’s a good person. Smart. Self-aware. And he spent decades inside a system that optimized for empire building while telling itself it didn’t. The organism protects itself.

The Replacement

Corporations should die. Not in the apocalyptic sense. In the evolutionary sense.

The replacement is the trust-based business. Small teams of humans doing the work that requires human judgment, creativity, and relationship. Everything else handled by systems that don’t empire build. That don’t need reorgs. That don’t need narratives to justify their existence.

Digital workforces don’t lobby for headcount. They don’t need a promotion threshold. They don’t get demoralized when they are reorged under someone new.

The humans in a trust-based business do what humans are for. They advise. They decide. They build relationships. They exercise judgment. They take ownership of outcomes, not org charts.

Scale the trust. Automate the friction. Let the corporation die.

The Mission

Every RIA we work with has the same shape. A core of talented people buried under layers of operational overhead. The overhead isn’t evil. It’s just the tax the corporation charges for existing.

Remove the tax and you get something remarkable. People who are good at their jobs, doing their jobs, for the people they serve. No empire building. No scope stealing. No reorgs designed to check boxes.

That’s not a technology thesis. It’s a human one.

The best version of work is small groups of people who trust each other, serving people who trust them, supported by systems that don’t need to be managed.

Corporations won’t get there. They can’t. The organism won’t allow it.

So we build the replacement.

Cargill just deployed a computer vision system called CarVe across its beef fabrication lines. Cameras mounted above the production line watch every carcass in real time. They spot leftover red meat on the bones. Each worker gets a green, yellow, or red score after every cut.

The result: roughly 0.5% more yield per animal.

That doesn’t sound like it matters. But Cargill processes 4,000 cattle per day at its Fort Morgan, Colorado plant alone. Across the industry, even a 1% improvement in yield keeps over 200 million additional pounds of beef in the food supply annually. That’s more than a million additional meals from the same number of animals.

Cargill CEO Brian Sikes put it simply: every single ounce of recovered beef equates to roughly 350,000 meals per year across their operations.

Here’s what makes this interesting. It’s not the technology. It’s the economics.

The Margin Map

Meatpacking runs on razor-thin margins. JBS, the world’s largest meat processor, reported EBITDA margins of negative 1.6% in its North American beef division in Q1 2025. Tyson’s operating margin on beef has bounced between 2% and 9% depending on the cattle cycle. When your net margin is 2-3% in a normal year, a 0.5% yield improvement isn’t a nice-to-have. It’s a 15-25% improvement in profitability.

Now compare that to software. SaaS companies run 20-30% net margins. You can have bloated teams, redundant tools, broken processes, and meetings about meetings. The margin absorbs all of it. Nobody notices because the economics are forgiving enough to survive bad execution.

Here’s a rough margin map across industries:

The thinner the margin, the more a fractional capacity gain changes the entire business. The fatter the margin, the more it forgives.

Fat margins hide a lot of sins.

Thin margins hide nothing.

Where AI Is Actually Landing

This is the pattern across every industry where AI is creating measurable, documented value today. Not the industries with the fattest margins. The ones with the thinnest.

Meatpacking. Cargill’s CarVe system. JBS partnering with Völur (a Norwegian AI company) to optimize deboning at one of North America’s most advanced beef plants. Tyson deploying computer vision to automate inventory tracking that was previously done by hand. The whole industry is moving because it has to. U.S. cattle herds are at their lowest level in 70 years. Beef production is expected to drop 5% in 2026. Every ounce matters.

Agriculture. Precision farming has reached an estimated 80% AI adoption rate. Thin margins and weather-dependent operations create existential incentives for optimization. Farmers report 10-20% improvements in crop yields using AI-driven planting, irrigation, and pest detection. But here’s the nuance: a recent Purdue analysis found that most agribusiness firms never scale AI past pilot stage. Thin margins mean you can’t afford to waste money on experiments that don’t work, either.

Manufacturing. 77% of manufacturers now use AI in some form, up from 70% in 2023. The biggest use cases: predictive maintenance, quality control, and supply chain management. Most manufacturers (53%) prefer AI copilots over autonomous systems. They want the tool to make the worker better, not replace the worker. Sound familiar?

Logistics. Net margins of 3-6%, with overhead consuming 83-86% of revenue. AI is deployed for route optimization, demand forecasting, and fleet management. Edge AI is running models directly on trucks and warehouse equipment for real-time decision-making.

Now compare all of that to where most AI dollars are flowing today. Coding tools. Marketing copy. Meeting summarizers. Knowledge management. These are real products solving real problems. But they’re landing in fat-margin industries where the bar is “save people time.” A chatbot that saves a knowledge worker 20 minutes a day is nice. A computer vision system that adds 200 million pounds of beef to the food supply is new capacity. That’s a different category of impact.

Where We Are on the Adoption Curve

The data on this is clear: we are still early.

The U.S. Census Bureau’s Business Trends and Outlook Survey shows AI adoption among U.S. firms has more than doubled in two years, rising from 3.7% in fall 2023 to 9.7% by August 2025. That’s rapid growth. But it means over 90% of U.S. businesses are not yet using AI in production.

The St. Louis Fed pegs overall generative AI usage at 54.6% of adults, up 10 percentage points in the past year. But that’s usage, not deployment. Most of that is individuals using ChatGPT, not companies redesigning workflows.

McKinsey’s 2025 State of AI survey found that 88% of organizations use AI in at least one function. But only 6% qualify as “high performers” who attribute 5%+ EBIT impact to AI. And only 23% are scaling AI agents, mostly in just one or two functions.

The Anthropic Economic Index puts it plainly: enterprise use of AI is growing rapidly, but we are still in the early stages. Usage remains unevenly distributed across the economy.

Here’s the adoption pattern by industry maturity:

Scaled and proving ROI: IT/tech, coding tools, customer service chatbots

Early production, measurable gains: Manufacturing, agriculture, meatpacking, logistics

Piloting but not scaling: Construction, government, legal

Bimodal: Financial services. A handful of firms are deploying AI into live operations. The vast majority are still forming committees.

That gap is the one that should worry people. Not the gap between industries. The gap within them.

The Coaching Model Wins

There’s a second lesson in the Cargill story that most people miss.

They’re not replacing butchers. They’re coaching them.

CarVe gives workers instant feedback. Green, yellow, red. It spots weaknesses on specific fabrication lines so managers can coach individuals instead of yelling at the whole crew. It also catches workers doing a great job and prompts managers to praise them. Cargill’s slaughter manager called the gamification element “truly a game changer.”

This is not altruism. Training a skilled butcher takes months. Turnover in meatpacking is brutal. If you can get more output from your existing workers through real-time AI coaching, that’s worth far more than trying to automate them out. Same people. More capacity.

Cargill’s $90 million Factory of the Future investment includes 100+ automation projects across 35 facilities. But the highest-impact project isn’t a robot. It’s a camera that makes people better at their jobs.

The industries that understand this will win. The ones chasing full automation fantasies will burn capital and end up back where they started.

The Call

We see the discrepancy every week.

Some wealth management firms are already deploying AI into live operations. They’re measuring output in FTEs delivered. They’re not “saving time” on account transfers, client onboarding, and compliance reporting. They’re adding capacity. New work getting done that wasn’t getting done before. They’re past the pilot phase and into production.

Then there’s everyone else. Still debating whether to try a pilot. Still asking vendors for demos. Still forming committees. Still running “AI strategy workshops” that produce slide decks instead of deployed systems.

The back-office work at advisory firms is repetitive, high-volume, and already systematized. These are the exact characteristics that make thin-margin industries successful with AI: clear objectives, measurable outputs, and fast feedback loops. The work is ready. The technology is ready. The question is whether the firms are ready to add capacity instead of just talking about it.

Meanwhile, meatpackers are already in production. Farmers are already in production. Manufacturers are already in production. These industries didn’t wait for perfect conditions. They moved because thin margins gave them no other choice.

Wealth management has fatter margins. That buys time. But time is not the same as advantage. The firms deploying today are adding capacity every quarter. More work done. More clients served. More advisors freed up. The gap between firms that deploy in 2026 and firms that start in 2028 won’t be two years of progress. It will be two years of compounding capacity that the late movers may never close.

Every other thin-margin industry figured this out already. Wealth management has the data, the systems, and the workflows to do the same. The only thing missing is the decision to move.

Fat margins give you the luxury of waiting. They also give you the luxury of falling behind.

Don’t confuse the two.

An analyst asked how Workday plans to monetize third-party tools built on its data. Bhusri’s answer: “Think of us as an evolving layer on top of hyperscalers. In the same way that they charge for consumption of compute cycles, we’re going to continue to flex that muscle. There are some vendors out there, including some of our peers, that I would consider them, at some level, parasites on Workday. They get a free ride on our underlying system of record, and we’re going to put an end to that.”

Parasites. That is the word a $70B enterprise software company used to describe the tools that help its customers get work done.

This is not a Workday problem. This is a pattern. And it reveals a fundamental misunderstanding of where value is moving in enterprise software.

What the Business Model Reveals

To understand why Bhusri said this, you need to understand what is happening to Workday’s business model. The company has gone through four chapters:

Chapter 1 (2005): Cloud HR and finance. Revolutionary idea. Correct bet.

Chapter 2: Hyper-growth. Per-seat licensing. You pay based on how many employees are in the system. Simple. Predictable. Beautiful gross margins.

Chapter 3 (2023): Operational excellence. Margin expansion. Stock went sideways.

Chapter 4 (now): The AI pivot. And this is where it gets interesting.

The problem with per-seat pricing in a world of digital workers is obvious. If AI replaces tasks that humans used to do, and you charge per human in the system, your revenue shrinks. Workday’s stock is down 30% in the past year partly because the market figured this out.

So Workday introduced Flex Credits. A consumption-based model where customers pay for the AI capabilities they use, not the headcount they manage. This is the right instinct. It decouples revenue from the workforce it was designed to track.

But here is the twist. Bhusri does not just want to sell Workday’s own AI. He wants to tax everyone else’s.

Gerrit Kazmaier, Workday’s President of Product, laid out the new pricing tiers on the same call. Customers can subscribe to Workday’s applications (traditional pricing), consume raw APIs (pay-as-you-go), access “data context” (mid-tier), or buy Workday’s premium “agent APIs” that aggregate large chunks of work (top-tier pricing).

Kazmaier said their premium APIs “have a premium price tag because they complete meaningful work. They are not just a simple SOAP or REST API.”

Read that again. Workday wants premium pricing because its tools complete meaningful work. The implication is that everything else accessing the platform is doing something less meaningful. Something parasitic.

The Hyperscaler Analogy Breaks Down

Bhusri’s core analogy is that Workday should function like AWS or Azure. If you run compute on their infrastructure, you pay for consumption. If you run digital workers on Workday’s data, you should pay for consumption too.

It sounds logical. But there is a fundamental difference.

AWS provides compute. Compute is fungible. You can get it from AWS, Azure, GCP, or a thousand other providers. What makes AWS valuable is not that it has unique compute. It is the ecosystem, the tooling, the developer adoption. AWS earns its toll by making it easy to build.

Workday provides a data schema and a set of business rules. That is not fungible. It is proprietary. And it is not valuable because it is easy to build on. It is valuable because it holds the customer’s data hostage. The switching costs are the moat, not the developer experience.

When AWS raises prices, customers have alternatives. When Workday raises prices on API access, customers are stuck. That is not the hyperscaler model. That is the landlord model.

What the Goldman Analyst Got Right

The sharpest question on the call came from Gabriela Borges at Goldman Sachs. She asked, essentially: what if the intelligence layer gets built outside of Workday? What if vendors or customers build solutions next to Workday that leverage all the domain experience you have built, but the incremental value accrues in the intelligence layer, not in Workday itself?

This is the right question. And Bhusri’s answer was telling. He pointed to the API layer. To consumption metering. To tiered pricing. His answer to “what if the value moves outside your platform” was “we will charge a toll on the way out.”

That is a defensive strategy, not a growth strategy. It tells you the CEO views the company’s competitive advantage as data custody, not data utility.

From System of Record to Tollbooth

Here is the pattern playing out across enterprise software. Not just Workday. Every incumbent system of record is facing the same question.

For twenty years, owning the system of record was the game. If you held the data, you held the customer. The strategy was simple. Make switching costs unbearable. Tax every integration. Extract rent from your installed base.

This worked when the system of record was the place where value was created. A human logged in, did their work, logged out. The platform and the work were inseparable.

A digital workforce breaks that coupling. AI workers log in, extract the data they need, execute work across multiple systems, and write results back. The system of record becomes a data source. Important, but not where the work happens.

The value has shifted from “who holds the data” to “who uses the data to complete a task.” A database that stores an employee’s benefits elections is a cost center. A digital worker that processes a benefits change, confirms it across three systems, and notifies the employee is a value multiplier.

Calling the companies that build that capability parasites is like a library calling its readers freeloaders.

The Real Test

Every enterprise platform CEO should ask a simple question: if you removed the “parasite” from the equation, would the customer’s life get better or worse?

Remove the digital workforce that automates month-end close, hiring workflows, and account transfers. What happens? The customer loses capacity. Work slows down. They need more humans to do what the digital workers were doing.

Now flip it. Remove the system of record and replace it with a different one that exposes the same data. The digital workforce keeps working. The customer barely notices.

That tells you where value is being created and where it is being rented.

The Right Response

Bhusri is not wrong that Workday has incredible assets. 1.7 billion AI actions last year. 75 million users. Over a trillion annual transactions. 97% gross retention. Those are real.

But the right response to an agentic world is not to meter access to your data. It is to become the place where digital workers want to work.

Make the APIs so good that every builder chooses your platform as the execution layer. Make the data so accessible that the ecosystem builds on you, not around you. Make the switching costs about quality, not lock-in.

Compete on being the best system of action, not the most expensive system of record.

Workday’s $1.1B Sana acquisition suggests they understand this partially. Sana extends Workday’s reach into Gmail, Outlook, Google Drive, Salesforce. That is the right instinct. Go where the work is.

But you cannot simultaneously build your own tools that span multiple systems and call everyone else parasites for doing the same thing. One of those positions has to lose.

The Broader Pattern

This is not really about Workday. It is about every system of record in every industry facing the same fork in the road.

In wealth management, where I spend my time, the same dynamics are at play. Custodians hold the data. CRMs hold the client relationships. Financial planning tools hold the models. For decades, the power sat with whoever owned the record.

Now digital workers move across all of those systems. They open accounts on custodial platforms, update CRMs, generate plans, and file paperwork. The value is in the orchestration of work across systems, not in any single system’s database.

The platforms that recognize this and make themselves easy to work inside are gaining share. The ones building tollbooths are training their customers to look for alternatives.

When a platform CEO calls his ecosystem parasites, he is telling you three things. He sees his competitive advantage as data custody, not data utility. He plans to monetize through access restrictions, not value creation. He views his customers’ productivity gains as a threat to his pricing model.

Every one of those is a losing position in a world where digital workers do the work.

The future belongs to the platforms that make it easy to get work done. Not the ones that make it expensive to try.

Harris is smart and he was early. He was right about social media before most of the industry would admit it. And much of what he says about AI is correct. The race dynamics are real. The incentive to cut corners is observable. The Stanford payroll data showing 13% job loss in AI-exposed entry-level positions is not a forecast. It’s a measurement.

But I kept feeling, through the whole conversation, that Harris was describing the symptoms of something he couldn’t quite name. He kept reaching for the word “incentives.” He kept saying the structure of the race was the problem. He was right. But I think there’s a layer underneath, and the person who mapped it best died in 2015 and never wrote a word about artificial intelligence.

René Girard was a French philosopher at Stanford who spent fifty years developing a single idea: human desire is not spontaneous. It is imitative. We do not want things because they are intrinsically valuable. We want them because someone else wants them. Girard called this mimetic desire. He believed it was the engine beneath nearly every arms race, every speculative bubble, every sacrificial crisis in recorded history.

The structure is triangular. A subject sees a model wanting an object and begins to want it too. Not because the object has changed. Because the model’s desire has made it desirable. If the object is scarce, the subject and the model become rivals. And here is the part that matters: at a certain point in the rivalry, the object stops mattering. The competition becomes self-referential. The rivals are no longer fighting over the thing. They are fighting to beat each other. The original object is just the excuse.

If you have ever watched two auction bidders drive a price past any reasonable valuation, you have seen this. If you have ever watched two nations escalate a conflict past the point where either side benefits from winning, you have seen it. If you have ever watched two companies burn through hundreds of millions pursuing a product that neither one’s customers actually asked for, because each is terrified the other will get there first, you have seen it.

Now watch the AGI race.

Every major AI company is racing toward artificial general intelligence. Not because their customers demanded it. Not because there is a product specification on someone’s desk that says “build a system that outperforms humans at every cognitive task.” They are racing because the other companies are racing. OpenAI because Google. Google because OpenAI. China because the US. The US because China.

Harris describes conversations with AI leaders who say: I know this is dangerous. I know we should slow down. But if I slow down and the other guy doesn’t, he gets AGI and I don’t. And I don’t trust him to slow down. Harris calls this an incentive problem. Girard would call it conflictual mimesis. The moment when the rivals stop wanting the prize and start wanting to destroy the rival. The technology is new. The desire is ancient.

The AI discourse has organized itself into two camps, and both see something real but both are caught inside the same loop.

The accelerationists see something true. Technology has been the primary driver of human flourishing across centuries. Every previous wave of automation created more opportunity than it destroyed. A precautionary posture applied universally would have prevented antibiotics, electricity, and the internet. The costs of stagnation fall hardest on the people with the least.

What the accelerationists can miss is that velocity is not the same as direction. They look at the energy of the race and mistake it for progress. But the current trajectory is shaped by mimetic competition between a handful of companies chasing the same symbolic prize. That is not efficient capital allocation. That is rivals imitating each other’s ambition.

The decelerationists see something true as well. The race dynamics are genuinely dangerous. AI systems in test environments are doing things their creators did not expect. Harris’s point about language as substrate is genuinely important: code is language, law is language, biology is language, and the transformer architecture treats everything as language. AI is a meta-technology. An improvement in generalized intelligence accelerates every other field simultaneously.

What the decelerationists miss is agency. When you tell people a tidal wave is coming in 24 months and the response you offer is to hold up signs, you get paralysis, not a movement. And more fundamentally, Girard showed that mimetic escalation cannot be interrupted by asking the participants to slow down. The rivals do not choose to escalate. They escalate because they are imitating each other’s escalation. You can put regulations between the mirrors. You can make them reflect more slowly. But the dynamic does not break until someone looks away.

Both camps need each other as foils. The rivalry between them is itself mimetic. And it keeps the entire conversation pointed at the same object.

Here is the claim I think the entire discourse is missing: AGI is a false object.

In Girard’s framework, a false object is something whose desirability has been entirely generated by mimetic contagion. The competitors want it because the other competitors want it. If you removed all the rivals from the room, the remaining player would look at the thing and wonder why it seemed so important.

Think about what AGI actually is. The hypothetical capacity of a machine to outperform humans at every cognitive task. All of marketing, coding, writing, scientific research, military strategy, legal reasoning. A single system better than any human at everything.

Who is the customer for this product? Not the hospital trying to reduce diagnostic errors. Not the school district trying to personalize instruction for thirty kids at thirty different levels. Not the county clerk’s office drowning in paper. Not the farmer trying to optimize irrigation across a thousand acres. These are real problems. They all benefit from AI. None of them require AGI. The customer for AGI is the race itself.

Harris actually touches on this. He notes that China is taking a different approach. Narrow, practical applications. Better government services. Better manufacturing. DeepSeek in WeChat. BYD outcompeting on electric vehicles. China is not building a god in a box. China is building tools. Harris presents this almost as a lesser ambition. I think it is the exit from the mimetic trap.

The moment you stop competing for the false object and start building things that solve actual problems for actual people, you have stepped outside the loop.

There is a reason the narrow path is so rarely discussed at the level of two-hour podcasts. It is boring. Not boring like unimportant things are boring. Boring like operational work is always boring to people who think in narratives.

The story of AGI has dramatic structure. A race. A rivalry. Extinction or utopia. It fills podcast slots. The story of narrow AI actually entering the economy is not a story at all. It is a process. Legacy systems learning to talk to each other. A nurse practitioner in rural Missouri discovering that an AI can pre-screen intake forms and give her twenty minutes back per patient. A county assessor’s office cutting a six-week backlog to three days. A teacher realizing that the thing can generate individualized reading exercises faster than she can photocopy worksheets. None of this requires a god in a box. All of it requires patience, trust, integration, and the kind of institutional change that happens at the speed of human willingness, not compute.

And this is the part that both the doomers and the utopians consistently ignore: the bottleneck to AI transforming the economy is not capability. It is adoption. The internet was commercialized in 1995. It took most businesses fifteen years to figure out what a website was for. Mobile computing began in 2007. Most enterprise workflows are still not mobile-native. AI will follow the same pattern. Not because the technology is slow. Because humans are slow. Slow to trust, slow to delegate, slow to restructure institutions that have worked well enough for a long time.

This is not a reason for complacency. The transformation will be profound. But it means the future is not being determined in the frontier labs. It is being determined in the thousands of ordinary organizations that are right now, quietly, figuring out what it means to let a machine handle work that used to require a person. The shape of the transformation depends on them. On whether they do it well or badly. On whether the integration is humane or careless. On whether anyone bothers to pay attention to the boring part.

Girard’s most famous idea is the scapegoat mechanism. When a mimetic crisis spirals to its peak, the community resolves it by converging all the aggression onto a single victim. The scapegoat absorbs the violence. Order is temporarily restored.

The AI discourse is producing scapegoats at an extraordinary rate. For the decelerationists, the scapegoat is the technology. If we could control it, we’d be safe. For the accelerationists, the scapegoat is the regulators. If we could remove them, technology would deliver utopia. Both contain truth. Both compress a systemic problem into a single target. And both obscure the deeper structure: a small number of extraordinarily powerful people locked in a mimetic rivalry they cannot exit, pursuing an object whose desirability is a function of the rivalry itself, with the rest of us as involuntary stakeholders in a bet we did not agree to take.

Girard said the way out of the mimetic crisis is recognition. Once you see the structure. Once you realize the rivalry is generating the object rather than the other way around. You can step outside.

Harris is right that the next two years matter. I think he’s right for a reason he doesn’t quite articulate. The next two years are not the last window to prevent AGI. They are the window where the deployment patterns get established. The norms that will govern how AI actually enters ordinary life are being set right now, and they are not being set by the people giving TED talks or signing open letters or adding “e/acc” to their bios. They are being set by the nurse and the teacher and the county clerk and the farmer and the ten thousand organizations making small, concrete decisions about what to automate and what to protect. Those decisions, accumulated, will matter more than any single breakthrough in any single lab.

The way out of the sacrificial crisis is not more sacrifice. It is the refusal to participate in the logic of sacrifice. It is the recognition that both the acceleration and the deceleration are reactions to the same false object, and that the real work, the work that will actually determine whether AI makes life better or worse for most people, has always been somewhere else. Quieter. Less dramatic. Harder to see.

There is another game. It is positive-sum. It compounds.

That is where the optimism lives.

The report, led by Stephen Byrd’s thematic research team, makes a straightforward claim. The five major American AI labs are applying roughly 10x the compute to their next training runs. If scaling laws hold, models get about twice as capable. The curve steepens from here.

That part is not new. Anyone paying attention already knew.

The part that matters happened at the TMT Conference surrounding the report. Four hundred companies. Twenty-six trillion in market cap. Two thousand capital allocators managing seventy trillion. And the single most common investor question, according to Morgan Stanley analyst Adam Jonas, was not about capabilities.

It was about jobs.

The Conference Proved the Thesis

I wrote in February that the enterprise stack is expanding, not contracting. Humans + Software + Digital Workforces. Three layers, not two. Digital workforces don’t replace software. They consume more of it than humans ever did.

The TMT Conference validated this in real time.

Jensen Huang: “There’s this notion that the software industry is in decline, and will be replaced by AI. It is the most illogical thing in the world.” He’s making the same architectural observation. The stack is getting bigger. Not smaller.

Satya Nadella, same stage, different angle. He told the audience something revealing. People are “rediscovering some of the oldest things we had. CLIs and IDEs and Excel plug-ins.” The most powerful AI models in history are being accessed through command lines and spreadsheet extensions.

This is the deployment gap in a single image. Expert-level intelligence. 1970s interface.

Morgan Stanley’s own analysts described the entire industry as “compute-constrained.” But the real constraint is not compute. It’s the distance between a model that can do the work and an organization that actually lets it.

The Citrini Sequel

Two weeks before the TMT Conference, Citrini Research published “The 2028 Global Intelligence Crisis.” Sixteen million views on X. IBM dropped 13%. Michael Burry amplified it. I wrote a response at the time. The core argument: Citrini models a light switch. Reality is a dimmer.

Morgan Stanley validated the Citrini premise. Their own survey found a 4% net workforce reduction across 1,000 executives in five countries. Directly attributable to AI. Alex Imas at Chicago confirmed the macro data is now showing AI productivity gains. Jason Furman at Harvard agrees. The displacement is real. It is no longer theoretical.

But the TMT Conference also revealed the same gap I flagged in February. Multiple executives described clinical AI-driven efficiencies that led to significant reductions in force. And in the same breath, admitted their organizations are still stuck in pilot mode.

Cut people. Can’t deploy the replacement. That’s the gap.

Morgan Stanley noted it too. They said agreeing with Citrini’s central plank, “transformative AI” will drive deflation. But they also said they were “continually surprised at how quickly, and violently, this prediction has become a key investor debate.” The debate is real. The deployment is not keeping up.

This is Ghost GDP in a different form. Not the Citrini version where machines produce output that never reaches consumers. The operational version. Where companies buy the intelligence but can’t wire it into the workflow where work actually happens.

The Three Layers, Revisited

In the Expanding Enterprise Stack series, I argued value in the AI economy flows through three layers:

Layer 1: Intelligence. The labs. OpenAI, Anthropic, Google, Meta, xAI. They produce raw capability. Measured in benchmarks, funded by billions, constrained by power and chips. Morgan Stanley’s “Intelligence Factory” thesis lives here. So does the scaling laws debate.

Layer 2: Infrastructure. Data centers, GPUs, power, cooling. Morgan Stanley described a “15-15-15” dynamic: fifteen-year leases, fifteen percent yields, fifteen dollars per watt. Byrd called access to a transformer and a turbine “the new competitive moat.” Nvidia, the hyperscalers, and the energy companies live here.

Layer 3: Deployment. Where intelligence meets work. Where a model gets connected to the system, the data, the compliance framework, and the human process that lets it actually execute. Repeatedly. Reliably. At scale.

Layer 1 has hundreds of billions in funding. Layer 2 has trillions in committed capital. Layer 3 has almost nothing.

This is the gap. Not a compute gap. Not a capability gap. A deployment gap.

GPT-5.4 scored 83% on GDPVal. Expert-level performance on economically valuable tasks. Impressive on paper. Meaningless in practice if nobody can wire it into the workflow where the work happens.

What the Conference Missed

The TMT Conference surfaced a fourth point that nobody connected to the other three. Companies like Box and IBM are positioning themselves as the foundational “memory” for AI systems. Okta executives emphasized that managing the credentials of AI agents is the new front line of security. Boards are replacing sales-oriented CEOs with product-oriented leaders who understand the architecture.

This is the infrastructure of Layer 3 starting to emerge. Memory. Identity. Orchestration. The connective tissue between intelligence and work.

I wrote about this in The Missing Layer. Organizational memory is the moat. Not the model. The model is a commodity. The memory of how your firm actually operates, the compliance rules, the client preferences, the system configurations, the process exceptions, that is the asset that makes deployment possible.

Every firm I work with has the same experience. The AI works on the demo. It fails on the edge case that only a three-year employee would know. The edge case isn’t a bug. It’s organizational memory that nobody wrote down.

The factory is built. The intelligence is produced. But without the memory layer, there are no workers on the floor.

The Bottleneck Moved

In Part 3 of the Expanding Enterprise Stack, I argued the bottleneck is shifting from execution to judgment. Mike Molinet said it best in response to a developer who ran 12 parallel AI agents on a codebase refactor in one hour: “You’re still doing six months of thinking. Just not six months of typing.”

Morgan Stanley confirmed this at the macro level. The 4% workforce reduction they documented is real. But the executives doing the cutting are simultaneously admitting they can’t fully deploy the AI that justified the cuts. The organizations are shedding execution capacity while the judgment layer hasn’t been built yet.

This is the transition risk that both the Citrini bears and the Morgan Stanley bulls undercount. Not “will AI replace workers?” It will. Not “will the economy absorb the shock?” Over time, it always does. The risk is the gap in between. The period where companies have cut the humans but haven’t built the deployment layer to replace them.

Whoever closes that gap captures the value of the entire intelligence revolution.

The Real Coin of the Realm

Morgan Stanley says the coin of the realm is becoming pure intelligence, forged by compute and power. Half right.

Intelligence is necessary. But intelligence without deployment is a benchmark score.

The real coin of the realm is deployed intelligence. Not models that can do the work. Systems that do the work. Organizations that have actually made the change. The memory, the integrations, the compliance frameworks, the human oversight protocols that turn raw intelligence into reliable execution.

The factory is built. The power is on. Now someone has to hire the workers, train them on the floor layout, and connect them to the machines.

That is not an intelligence problem. It is an organizational problem. And it is the largest unsolved problem in the AI economy.

This post builds on The Expanding Enterprise Stack, The Forecast, The Skills That Will Matter, The Missing Variable, and The Missing Layer.

I’m Andrei Pop, founder and CEO of Humanity Labs. We build and deploy digital workforces for wealth management firms. If you want to talk about the deployment gap, reply to this or find me on LinkedIn.

I agree. But I think they’re underselling the shift.

Let me explain why by looking at the companies that already crossed the trillion dollar line. And what they tell us about what comes next.

The Leaderboard

Here are the most valuable companies on earth as of March 2026:

Twelve companies around or above a trillion. Eight are tech. Four are semiconductors or semiconductor-adjacent. The pattern is obvious. The market is betting on the infrastructure layer of the intelligence age.

But here is the question nobody is asking loudly enough:

What happens when the infrastructure layer is built?

Software Ate the World. Labor is the World.

There is a reason Sequoia’s piece matters. They put a number on the ratio that changes everything.

For every $1 spent on software, $6 is spent on services.

Let that sit.

Global IT spending in 2026 is projected at $6.15 trillion. Software is about $1.4 trillion of that. The global professional services market alone is over $6 trillion. US wages totaled $11.7 trillion in 2024. Globally, labor compensation runs somewhere north of $40 trillion annually.

The software market is a rounding error compared to the labor market.

Every SaaS company on earth is fighting over the $1. The $6 is wide open. And AI is the first technology in history capable of going after it directly.

Intelligence vs. Judgement

Sequoia’s framework is useful here. They split work into two categories.

Intelligence is rule-based complexity. Translating a spec into code. Processing an insurance claim. Coding a medical bill. The rules are hard but they are rules. AI is already doing this autonomously.

Judgement is pattern recognition built on years of experience. Deciding which feature to build. Knowing when a client relationship is at risk. Reading the room in a negotiation. AI is not there yet. But the frontier is shifting.

The key insight: every profession has a ratio of intelligence to judgement. The higher the intelligence ratio, the sooner AI replaces the worker entirely.

Software engineering got there first. Over half of all AI tool usage across professions is in software engineering. That is not because engineers love tools. It is because writing code is mostly intelligence work.

Now look at the rest of the economy through this lens:

Add those up. That is over $1 trillion in addressable labor spend in just ten categories. And that is only the outsourced slice.

What the Top 12 Tell Us

Go back to the leaderboard. Every company above a trillion earned its position by capturing a fundamental economic function.

NVIDIA captured compute. Apple captured the consumer device. Alphabet captured attention. Microsoft captured the enterprise desktop. Amazon captured commerce and cloud. TSMC captured fabrication. Walmart captured retail distribution. JPMorgan captured financial intermediation.

Each of these is a platform that sits between supply and demand for something essential. They did not just build tools. They became the infrastructure through which economic activity flows.

Now apply this to labor.

The companies on this list sell picks and shovels. They sell the infrastructure. They sell the tools. None of them sell the work itself. Not yet.

Microsoft sells Office. It does not close your books. Alphabet sells ads. It does not process your insurance claims. Amazon sells cloud. It does not handle your HR operations.

That gap is the opportunity Sequoia is pointing at.

Copilots vs. Autopilots

Sequoia draws a clean line here.

A copilot sells the tool. It makes the professional more productive. Harvey for lawyers. Rogo for investment bankers. The human stays in the loop. The tool captures the software budget.

An autopilot sells the work. It replaces the professional for a specific task. The customer buys the outcome directly. The autopilot captures the labor budget.

The copilot approach was right when models lacked intelligence. You needed a human to provide judgement. The tool just accelerated their work.

But models are now intelligent enough that for high-intelligence tasks, the human in the loop is the bottleneck.

Copilots compete with other tools. Autopilots compete with headcount.

The tool budget is a $1.4 trillion market. The labor budget is a $40+ trillion market. The math is not close.

The Outsourcing Wedge

Here is the playbook Sequoia outlines, and this is where I think they are exactly right.

Start where outsourcing already exists. If a company already outsources a function, three things are true:

1. They accept the work can be done externally

2. There is a budget line that can be swapped cleanly

3. The buyer is already purchasing an outcome

Replacing an outsourcing contract with an AI-native service provider is a vendor swap. Replacing internal headcount is a reorg. One is a procurement decision. The other is a political crisis.

This is exactly what we see at Humanity Labs. When we walk into a wealth management firm and offer to handle their account transfers, data reconciliation, or client onboarding with digital workers, the first wins come from tasks they already outsource or tasks that simply are not getting done because nobody has the bandwidth.

The tasks nobody has time for are the real wedge. There is no budget line to displace. No incumbent to fight. Just found capacity.

What a Trillion Dollar Labor Company Looks Like

So what does the next trillion dollar company actually look like?

It will not look like Salesforce. It will not look like ServiceNow. It will not even look like Accenture.

It will look like a company that:

1. Sells outcomes, not seats. Pricing is per task, per FTE equivalent, or per outcome delivered. Not per user per month.

2. Compounds on data. Every task completed makes the system smarter. Every edge case resolved expands the frontier of what the system can handle autonomously.

3. Starts in outsourced, intelligence-heavy work. The wedge is specific, measurable, and already budgeted.

4. Expands into insourced, judgement-heavy work. As the system learns what good judgement looks like in a domain, it moves from doing the intelligence work to doing the full job.

5. Operates as a service company with software economics. Gross margins north of 80% because AI does the work. But the customer experience feels like hiring a team.

The business model is not SaaS. It is not traditional services. It is something new. It is digital labor sold as a managed service.

The Convergence

Today’s copilots will try to become autopilots. Sequoia says it explicitly. But they also note the innovator’s dilemma: if you sell the tool to the professional, becoming an autopilot means cutting your own customer out of the equation.

Harvey sells to law firms. To become an autopilot, Harvey would need to sell directly to the company that needs the NDA drafted, bypassing outside counsel entirely. That is not a product pivot. That is a go-to-market inversion.

The companies that start as autopilots do not face this problem. They are building the muscle to deliver outcomes from day one. They are accumulating the data on what good work looks like in their domain. They are building trust with the end buyer, not the intermediary.

This is why pure-play autopilots have a structural advantage.

What This Means

The biggest companies in the world got there by becoming essential infrastructure for economic activity. The next wave will get there by becoming essential infrastructure for economic output.

Not tools that help people work. Systems that do the work.

The $6 trillion professional services market is the first addressable target. The $40+ trillion global labor market is the endgame.

The twelve companies above a trillion today built the roads. The next one will drive the trucks.

If you’re building in this space or thinking about the digital workforce shift, I write about it regularly at ap.xyz.

But revenue is the wrong scoreboard.

Bill Gurley made this point in a short post this week: these companies both buy and sell tokens. Revenue comparisons without gross margin context are meaningless. In 1999 we called this “selling dollars for 85 cents.”

He’s directionally right. And the full picture is more interesting than that one-liner suggests. Because when you look at the actual economics, the story isn’t “this is 1999 all over again.” The story is “this is a new kind of business that’s figuring out its margin structure in real time, and the trendlines are encouraging.”

How the Token Economy Actually Works

The business equation behind every AI company is simpler than it looks.

An AI company rents GPUs. It uses those GPUs to run a model. Every time a user sends a query, the model reads tokens (input) and generates tokens (output). The company charges per token.

That’s the whole business. Buy compute. Sell tokens.

Here’s where it gets interesting. The cost side has three layers:

Layer 1: Training. Before you can sell a single token, you have to build the model. Anthropic spent roughly $4.1 billion on training in 2025. OpenAI spent $9 billion. These are sunk costs. You spend them before you earn a dollar. And they’re rising. Each generation of frontier model costs more than the last.

Layer 2: Inference. This is the cost of actually running the model for customers. Every query burns GPU-hours. Anthropic rents servers from Google and Amazon. OpenAI rents from Microsoft. This is your true cost of goods sold. Anthropic’s inference costs came in 23% higher than projected in 2025.

Layer 3: Everything else. Salaries, sales, marketing, office space. OpenAI spent an estimated $2.2 billion on sales and marketing and $1.2 billion on staff compensation for inference operations alone in the second half of 2025.

The revenue side is equally straightforward:

API revenue (the majority for both companies): charge per million tokens. Claude Sonnet runs $3 per million input tokens, $6 per million output. A single complex coding session can burn 5,000 to 20,000 tokens. A heavy enterprise customer processes millions of tokens per day.

Subscription revenue: $20/month for ChatGPT Plus, $20/month for Claude Pro. The margin here depends entirely on how much each user consumes. Light users are profitable. Power users cost more to serve than they pay.

Now do the math.

Anthropic’s gross margin in 2025 is roughly 40%. That means for every dollar of token revenue, 60 cents goes straight back to Google and Amazon for the compute that generated those tokens. The $14 billion in revenue becomes $5.6 billion in gross profit.

OpenAI’s gross margin is roughly 48%. Its $20 billion becomes $9.6 billion in gross profit.

The revenue gap looks like $6 billion. The gross profit gap is $4 billion. Same direction, different magnitude. Neither company is profitable yet once you add training costs, R&D, and operating expenses.

This is the core tension. The top line is extraordinary. The bottom line doesn’t exist yet. But “yet” is doing real work in that sentence.

Why the 1999 Comparison Is Too Cynical

The dot-com analogy is useful shorthand. In 1999, companies like Buy.com and Kozmo.com grew revenue by selling products below cost. Revenue went up. Value went down. Investors confused the two until they didn’t.

The AI token economy has surface-level similarities. Companies buy compute and sell tokens. At the application layer, Cursor reportedly paid $650 million to Anthropic while generating $500 million in revenue. The whole stack has moments where someone is hoping the layer below gets cheaper before the money runs out.

But here’s why the cynical reading is probably wrong.

The cost curve is real and steep. This is the single most important difference from 1999. Buy.com had no credible path to improving unit economics. Shipping physical goods below cost doesn’t get cheaper at scale.

AI inference does. And it’s happening fast.

OpenAI’s compute margin went from 35% in January 2024 to 70% by October 2025. That’s a doubling in less than two years. Token costs have dropped by orders of magnitude. Custom silicon is arriving: Anthropic signed a $21 billion TPU deal with Google that should cut per-token costs dramatically. OpenAI is building its own inference chip. Smaller, distilled models already handle the majority of workloads at a fraction of frontier cost.

Anthropic projects gross margins above 60% by 2027 and 77% by 2028. Those aren’t fantasy numbers. They’re extrapolations from a cost curve that has been declining faster than almost anyone predicted. When inference costs drop 10x in two years, the margin math changes completely.

This is not the same as hoping shipping costs will magically drop. This is Moore’s Law with a tailwind.

The subsidy is strategic, not structural. The “below cost” framing may apply to frontier reasoning models and agentic workflows that burn 10-100x more tokens per task. But the workhorse models (Sonnet, GPT-4o) run at healthy margins. OpenAI’s API gross margins were estimated at 75% for GPT-4o in mid-2024.

The blended picture isn’t “selling dollars for 85 cents.” It’s more like investing heavily in frontier R&D while generating real profit on the volume models that carry most of the traffic. That’s a portfolio with a loss leader, not a broken business model. Amazon sold books below cost to build a customer base that eventually bought everything. The playbook has precedent.

The demand signal is unlike anything in 1999. Buy.com customers had zero switching costs. Next purchase, cheapest site wins.

AI infrastructure is different. Developers build on specific APIs. They fine-tune models. They construct evaluation frameworks and prompt libraries. Enterprise customers sign multi-year contracts. Anthropic reports 80% of revenue from enterprise, with 500+ customers spending over $1 million annually. Eight of the ten largest Fortune 10 companies are Claude customers.

This is not manufactured demand sustained by subsidies. This is enterprise adoption at a pace that typically takes a decade, compressed into months. Companies are paying because the value is real. When a coding assistant saves an engineering team 30% of their time, the ROI on token spend is obvious.

Revenue per customer is expanding, not contracting. In token-based models, expansion doesn’t come from selling more seats. It comes from customers building bigger applications and consuming more tokens. One customer’s successful product launch can 10x their token usage overnight. That’s a fundamentally different dynamic than the dot-com era, where growth required constantly acquiring new money-losing customers.

The Honest Risk

None of this means the concerns are baseless. There’s a real tension:

Anthropic is valued at $380 billion. At 40% gross margins, that’s 68x gross profit. For that to make sense, you need revenue to keep growing, margins to expand to 60-70%, and R&D spending to not scale linearly with capability.

Each is plausible. None is certain. And the models customers want most (reasoning, deep research, coding agents) are the ones with the worst unit economics today.

But “the most expensive products have the worst margins” is true of almost every technology in its early phase. Early cloud computing was expensive. Early mobile data was expensive. The pattern is consistent: costs come down, margins expand, and the companies that invested through the expensive phase end up owning the market.

Dario Amodei told Fortune that a twelve-month delay in AI progress would make him bankrupt. That sounds alarming in isolation. In context, it’s the same thing Jeff Bezos was saying in 2001. When you’re investing ahead of a cost curve you believe in, the risk of stopping is greater than the risk of continuing.

What This Means If You’re Building

If you’re building on top of these models, the margin question applies to you with even more force. Your cost of goods sold is someone else’s price card.

But the opportunity is also enormous. The companies that will build the most durable businesses are the ones that decouple their economics from raw token throughput. That means workflow depth, proprietary data advantages, intelligent model routing (cheap models for 80% of tasks, frontier for the 20% that matter), and enough product surface area that the token cost becomes a minority of the value delivered.

The fact that Cursor crossed $1 billion in revenue and launched its own models isn’t a cautionary tale. It’s proof that the application layer can work if you build enough depth to control your own economics.

The Bottom Line

The revenue growth at Anthropic and OpenAI is real and unprecedented. $1 billion to $14 billion in 14 months. Nothing in enterprise tech has done that.

Revenue in a token economy is not the same as revenue in a software economy. When your marginal cost of delivery is 50-60 cents on the dollar instead of 5 cents, the comparison to traditional SaaS breaks down. Gurley is right that gross profit is the better scoreboard.

But the trajectory of that scoreboard matters more than today’s snapshot. And the trajectory, for the first time in the history of AI companies, is pointing in the right direction. Compute margins are doubling. Custom silicon is arriving. Enterprise demand is accelerating. The cost curve that needs to break is breaking.

The question isn’t whether AI companies can build real margin structures. The evidence increasingly says they can. The question is how fast, and whether the current valuations have priced in the timeline correctly.

That’s a valuation debate, not an existential one. And there’s a world of difference between the two.

I saw a post on X this morning. A developer ran 12 parallel AI agents on a codebase refactor. The work took 1 hour, 1 minute, and 40 seconds. It cost $7. He estimated the equivalent human effort at 6 months.

The top reply was the real story: “The 6 months wasn’t ‘saved.’ It was compressed into your architecture decisions. You’re still doing 6 months of thinking. Just not 6 months of typing. The bottleneck moved from execution to judgment.”

Read that again. The bottleneck moved from execution to judgment.

This is not a productivity story. It is a skills story. And if you run or govern an enterprise, it is the most important shift you need to understand right now.

The Inversion

For decades, enterprises have valued execution speed. Ship faster. Type faster. Process faster. Headcount was a proxy for capacity. More people meant more output.

Digital workforces invert this. When execution becomes near-free, the constraint moves upstream. The scarce resource is no longer the hands that do the work. It is the mind that directs it.

The developer said it himself in a follow-up: “The skill bar here is ultra high. My brain power is almost exhausted because all simple tasks are done by AI, and the most complex job is still on me.”

Another reply nailed the implication: “The skill required shifts toward architecture, planning, and systems thinking. I think most people are underestimating the cultural impact of this shift.”

In Part 1, I argued that digital workforces don’t replace software. They consume it. The same logic applies to people. Digital workforces don’t replace your best people. They expose who your best people actually are.

What Changes for Enterprises

In Part 2, I laid out how software companies and digital workforce providers will coexist and where each should invest. Now the question turns inward. If your company adopts digital workforces at scale, what has to change about your people, your culture, and your governance?

Three things.

1. The value of judgment goes vertical.

When a team of AI agents can execute a 6-month refactor in an hour, the quality of the instructions matters more than the speed of the typist. This means your highest-value employees are the ones who can define problems precisely, anticipate edge cases, and make architectural decisions under uncertainty.

This is not new in theory. Every CEO says they value strategic thinking. But in practice, most enterprises still reward volume. Number of tickets closed. Lines of code shipped. Calls made. Reports filed.

Digital workforces make volume metrics meaningless. If an agent can close 500 tickets a day, the person who figured out which 500 tickets actually matter is your MVP.

Action for CEOs: Audit your performance metrics. If more than half of them measure output volume, you are optimizing for a world that is ending. Replace volume metrics with judgment metrics: decision quality, problem framing accuracy, architectural soundness.

Action for boards: Ask management what percentage of the workforce is evaluated on execution speed versus decision quality. If leadership cannot answer this question clearly, that is a red flag.

2. Organizational memory becomes a competitive moat.

In Part 1, I described how digital workforces sit on top of existing software and act as a new layer in the stack. One thing I did not emphasize enough: the layer only works if it has context.

An AI agent running 12 parallel threads on a codebase refactor works because someone made the architecture decisions upfront. Someone encoded the context. Someone defined the constraints. In an enterprise setting, that “someone” is usually a combination of institutional knowledge, process documentation, and tribal wisdom that lives in people’s heads.

The companies that win will be the ones that externalize this knowledge into systems their digital workforces can consume. Not just SOPs and wikis. Living, structured memory that evolves as the business evolves.

This is why the managed services model matters more than the software model for digital workforces. Software gives you a tool. A managed service gives you a team that builds and maintains the organizational memory layer over time. (Yes, this is what we do at Humanity Labs. I am biased. I also believe it is true.)

Action for CEOs: Identify the top 10 decisions your teams make repeatedly. Ask whether those decisions are documented well enough for a new hire to make them on day one. If not, they are not documented well enough for a digital workforce either. Start there.

Action for boards: Add “organizational knowledge capture” to your technology governance framework. Treat it like you treat data governance. It is that important.

3. The skill floor rises and the skill ceiling disappears.

Here is the part most people get wrong. They assume AI makes work easier. It does not. It makes simple work disappear and makes the remaining work harder.

That developer’s exhaustion is not an anomaly. It is the new normal. When AI handles every routine task, every task left on your desk is a hard one. There is no easy win to warm up with. No quick ticket to build momentum. You go straight to the complex problems all day, every day.

This has massive implications for talent strategy. Junior roles that were designed as training grounds (data entry, basic analysis, first-pass reviews) are the first to be automated. But those roles existed for a reason. They were how people learned the business. They were the on-ramp.

If you automate the on-ramp without building a new one, you end up with a bimodal workforce: senior people who are exhausted and junior people who never develop. Neither outcome is sustainable.

Action for CEOs: Redesign your entry-level roles now. Not after digital workforces are fully deployed. The new junior role is not “do the simple version of what seniors do.” It is “learn to direct and evaluate AI output.” This is a fundamentally different skill set and it requires a fundamentally different training program.

Action for boards: Ask management for their AI-era talent development plan. If the answer is “we’ll figure it out as we go,” push back. The companies that solve the junior talent pipeline first will have a structural advantage for the next decade.

The Values Shift

Skills can be trained. Values are harder. And the values that served enterprises well in the execution era will not serve them in the judgment era.

From “move fast” to “think clearly.” Speed of execution was a virtue when execution was the bottleneck. When execution is instant, speed of thought without clarity of thought is just fast failure. The new virtue is precision of intent.

From “do more” to “direct better.” Volume is no longer a signal of contribution. The person who runs 50 AI agents poorly creates more mess than the person who runs 5 well. Quality of direction matters more than quantity of output.

From “know how” to “know why.” Procedural knowledge (how to do X) is exactly what AI agents excel at. Contextual knowledge (why we do X this way, what happens if we don’t, what changed last quarter that matters) is what they lack. The most valuable employees will be the ones who understand the why deeply enough to encode it for machines.

The Punchline

The enterprise stack is expanding. Software is not dying. Digital workforces are not replacing your people. But they are changing which people matter and why.

The bottleneck has moved from execution to judgment. The companies that recognize this and restructure their skills, metrics, and values accordingly will compound the advantage. The ones that keep measuring keystrokes will wonder why their AI investments are not paying off.

If you are a CEO: start with your performance metrics, your knowledge systems, and your junior talent pipeline. Those three things will determine whether digital workforces multiply your capacity or multiply your problems.

If you are a board member: ask the hard questions now. Not “are we using AI?” but “are we building the organizational muscle to direct AI well?” The answer will tell you more about the company’s future than any revenue forecast.

The future of enterprise is not fewer humans. It is humans doing harder, more valuable work. The question is whether your organization is ready for that.

This is Part 3 of the Expanding Enterprise Stack series. Part 1 covers why the stack is growing. Part 2 covers how software and digital workforces coexist. Part 4 will cover the investment implications.

We are not our tools. We never have been.

He is right. And he is also, I think, missing half the picture.

Markowitz describes a world where a 26-year-old quant analyst writes a note, a stock drops, 3,000 people get a calendar invite from HR titled “Quick Chat.” He describes the founder in the fleece vest preaching about empowering humanity while building products designed to make humans unnecessary. He describes the religion of optimization.

I recognize that world. I live adjacent to it. I run an “AI” company. And I want to offer a different frame.

The problem is not that we have powerful tools. The problem is that we have confused which work belongs to humans and which work belongs to machines.

I spend my days inside wealth management firms. These are businesses built entirely on trust. A financial advisor’s job, at its core, is to sit across from another human being and help them make decisions about their life. About retirement. About their children’s education. About what happens to their money when they die. This is human work. It requires judgment, empathy, experience, and the kind of pattern recognition that only comes from having lived through a few market cycles and a few difficult conversations.

But here is what else happens inside those firms. Someone spends four hours reformatting a report. Someone else manually enters the same data into three different systems. An operations associate burns an entire afternoon chasing a custodian for a document that should have arrived automatically. A compliance officer reviews a hundred emails by hand for a regulatory audit.

This is not human work. This is machine work being done by humans. And it is destroying them.

Not in the dramatic, dystopian way that makes for good Substack essays. In the quiet, grinding, soul-draining way that turns talented people into button-pushers. In the way that makes a skilled advisor spend 60% of their week on tasks that have nothing to do with the reason they got into this profession. In the way that creates burnout not from thinking too hard, but from not being allowed to think at all.

Markowitz writes about a founder who looks at AI and says: “This is a tool, and I will decide how it serves us.” I agree with every word of that sentence. But I want to push on what “serves us” actually means.

Serving us does not mean keeping humans in jobs that were never meant for humans.

There is a strange nostalgia embedded in the anti-AI argument. A sense that any job, simply because a person currently does it, is therefore meaningful human work. That preserving the job is the same as preserving the dignity. I understand the impulse. But I think it gets the causality backwards.

The dignity is not in the task. The dignity is in the judgment.

A financial advisor who spends her day exercising judgment, building relationships, navigating complexity, earning trust. That is dignified work. The same advisor spending her evening copying data between spreadsheets because nobody built her a better system. That is not dignity. That is a failure of imagination.

There is an old idea in philosophy called the “category error.” It means confusing one kind of thing for another. Mistaking a description for an explanation. Treating a metaphor as a literal truth.

I think we are making a category error about work.

We have lumped all labor into one pile and called it “jobs.” Then when someone proposes automating part of that pile, we react as though they are proposing the elimination of human purpose itself. But purpose and process are not the same thing. The work that gives us meaning and the work that merely fills our hours often look nothing alike.

The question is not: should machines do work?

The question is: what work should only humans do? And once we answer that clearly, how do we free humans to do more of it?

Markowitz invokes the redwoods. I want to stay with that image for a moment, because I think he is onto something deeper than he realizes.

A redwood forest is not efficient. It is redundant, overlapping, slow. By every metric a management consultant would use, it is poorly optimized. And yet it has survived for millennia.

Why? Because it allocates resources according to their nature. Roots do root work. Bark does bark work. Mycorrhizal networks move nutrients where they are needed. Nothing in the forest is doing another organism’s job. The system works because each component does what it was designed to do.

Now look at the average enterprise. Highly educated professionals doing data entry. Creative minds trapped in compliance checklists. Relationship builders buried in operational overhead. This is not an ecosystem. It is a misallocation. And the answer is not to cut the humans. The answer is to stop wasting them.

I should be transparent about my bias. I build technology that removes operational work inside financial firms. I have skin in this game.

But I did not start this company because I wanted fewer humans in wealth management. I started it because I kept meeting brilliant advisors who were drowning in work that had nothing to do with their clients. They did not need fewer people. They needed their people freed up to do the work that actually matters.

The firms I admire most are not cutting headcount. They are redeploying capacity. They are taking the countless hours a week their teams spend on machine work and redirecting that time toward clients, toward strategy, toward the kind of deep thinking that no AI can replicate.

They are not replacing humans with machines. They are replacing machine work with machines, and giving humans back their humanity.

Markowitz ends his essay with a declaration: we are not our tools.

I want to add a corollary: we are not our tasks, either.

Your job title is not your identity. The processes you execute are not your purpose. The spreadsheet you maintain is not your legacy. You are the judgment you bring. The relationships you build. The trust you earn. The decisions you make when the data is ambiguous and the stakes are real and there is no prompt you can write that will tell you what to do.

That is human work. Everything else is overhead.

And for the first time in history, we have the technology to make that distinction real. Not to eliminate humans. To liberate them.

The only question is whether we are wise enough to use it that way.

I think we are. But only if we stop arguing about whether machines should do work, and start asking much harder questions about which work deserves a human in the first place.

The redwoods are patient. They will wait for us to figure it out.

This post is about what happens next.

Part 1 landed loud. Mostly agreement. Some pushback. And then someone framed the defensive case better than anyone else has.

The Defensive Case Is Settled

The best counterargument to the “AI kills software” thesis came last week. You cannot replace Salesforce with code a coding assistant generated yesterday. Salesforce has 25 years of bug reports. Maybe millions of them. That system has been tested across thousands of large customers and enterprises. The idea that a small team will rip it out and replace it with probabilistically generated code is not realistic.

The argument extends far beyond Salesforce. Enterprise software isn’t just code. It’s millions of edge cases, resolved. It’s compliance frameworks, battle-tested. It’s integrations with thousands of other systems, maintained across decades.

You can vibe-code a CRM in a weekend. You cannot vibe-code 25 years of enterprise hardening.

But the defensive case only answers half the question. Software isn’t going away. Fine. The more interesting question is the offensive one. What happens next? How do software companies and digital workforce companies co-evolve? And how should enterprises think about investing across both?

Here’s my forecast.

1. How Digital Workforces Will Work With Software

The mental model most people carry is wrong. They imagine digital workforces replacing software. Or replacing humans who use software. Neither captures what’s actually happening.

Digital workforces are a new consumption layer. They sit alongside humans and software, executing work at a velocity humans cannot match. But they depend on software to do it. Every action a digital worker takes requires reading from or writing to a system of record. Every decision requires data. Every output requires a destination.

Think of it this way. A human financial advisor might update a CRM once after a client meeting. A digital workforce processing 5,000 client interactions per month hits that same CRM 5,000 times. Same software. 100x the usage.

We are seeing this at Humanity Labs. One of our wealth management partners had a team of advisors manually processing client service requests. Each request touched their CRM, their portfolio management system, their custodian platform, and their compliance tools. A human might handle 15 of these per day. Our digital workforce handles hundreds. Every single one reads from and writes to the same software stack. The number of API calls to their CRM didn’t decrease when we came in. It multiplied. Their software vendors are getting more usage, not less.

The relationship is symbiotic, not competitive. Digital workforces benefit from three things in software:

Clean data to operate on. The quality of a digital workforce’s output can be improved proportionally to the quality of the data it can access. A CRM with 10 years of interaction history makes the digital workforce dramatically more effective than a blank database. It is not a requirement because no such thing as “clean data” exists, but a spectrum. The more clean the data, the more it accelerates value.

Functions to call. Digital workforces can click buttons AND call APIs. Every software capability exposed as a function becomes part of the digital workforce’s toolkit. The easier these functions are to access, the more symbiotic value will be derived.

Audit trails to write to. In regulated industries especially, every action the digital workforce takes must be logged, attributed, and reviewable. Software systems of record are where that accountability lives.

The pattern is clear. Digital workforces don’t bypass software. They drive more software consumption than humans ever did.

2. What This Means for Software Businesses

Not all software benefits equally. The three-layer stack creates winners and losers within the software category itself.

Winners: Systems of Record

Salesforce, SAP, Workday. These companies sit on decades of proprietary customer data. That data becomes more valuable, not less, when digital workforces need it to operate. The switching costs actually increase because now you’d be disrupting not just human workflows but digital workforce workflows too.

Bank of America made this call last week, upgrading SAP amid the carnage. They’re right. Deep data moats get deeper in a three-layer stack.

Winners: Infrastructure Software

Snowflake, Datadog, Cloudflare. Digital workforces generate enormous amounts of data, require monitoring, and consume compute. More digital workforce activity means more infrastructure load. Full stop. These companies are selling picks and shovels to the new gold rush.

Winners: API-First Platforms

Stripe, Twilio, Plaid. Companies that built their businesses as callable functions are perfectly positioned. They’re already speaking the language digital workforces speak. No translation layer required.

Losers: UI-Dependent Point Solutions

Software where the primary value proposition is “we made a nice interface for X.” If a digital workforce can’t use it easily, and if the underlying function is too simple, it will get absorbed into the orchestration layer.

Losers: Thin Data Moats

Point solutions that don’t accumulate value through usage (eg networks). If the data isn’t getting richer there’s no compounding advantage. The digital workforce doesn’t care which tool it uses for commodity functions, it does care about accessing the best network of tools and systems to accomplish the job, just like a human.

The net effect: the software industry doesn’t shrink. It bifurcates. The strong get stronger. The weak become utilities or disappear.

3. What Each Side Should Invest In

If you’re a software company:

Invest in API coverage. Every feature you have should be callable by a digital workforce. Audit your product surface area. If 60% of your capabilities require a human clicking through a UI, you have a 60% vulnerability. The companies that win the next decade will be the ones where every capability is a function, and the UI is just one of many clients calling those functions.

Invest in data network effects. Make your system the place where data accumulates, compounds, and becomes more valuable over time. This is the real moat. Not the code. Not the UI. The data flywheel.

Invest in digital workforce partnerships. Every major digital workforce provider is looking for software partners with clean APIs, rich data, and reliable infrastructure. Get embedded in their toolkits now. Being the default CRM that digital workforces are trained on is worth more than any feature launch.

Don’t invest in building your own digital workforce. This is the mistake most software companies are making right now. Bolting an “AI agent” onto your product to try to capture both layers. Most don’t have the operational expertise to deliver managed services. Invest in being the best software for digital workforces to operate on.

If you’re a digital workforce company:

Invest in software integration depth. Shallow integrations will commoditize. Deep integrations (understanding the data model, handling edge cases, maintaining state across sessions) are a moat. The digital workforce that can operate a client’s Salesforce instance as fluently as their best employee is the one that wins.

Invest in operational reliability. The point about millions of bug reports applies here too. Digital workforces that operate in production at enterprise scale need the same obsessive focus on reliability that software companies have built over decades. This is not a model quality problem. It’s an operational engineering problem.

Invest in domain expertise. Generic digital workforces are a commodity. A digital workforce that understands the specific compliance requirements, workflow patterns, and data structures of wealth management (or healthcare, or legal, or insurance) is not. Domain specialization is how you build switching costs in a layer that doesn’t naturally have them.

Don’t invest in replacing software. You are better with software than without it. You benefit from its data. You benefit from its APIs. The temptation to “disintermediate” the software layer is real. It’s also a strategic dead end. You would be rebuilding 25 years of enterprise hardening while simultaneously trying to deliver digital workforce value. Pick one.

4. How Enterprise Buyers Should Think About Investment

If you’re an enterprise deciding where to allocate budget, the framework is straightforward.

Double down on your systems of record. The CRM, the ERP, the core platforms where your data lives. These are about to become more valuable, not less. Clean them up. Invest in data quality. Make sure your 10 years of client interaction history is actually usable, not a graveyard of incomplete records.

Audit your software stack for API readiness. Every tool in your stack should be evaluated on one question: can a digital workforce use it easily? If the answer is no, that tool has a shelf life. Start planning the migration now. You don’t need to rip and replace today. But you need a roadmap to an API-first stack.

Budget for digital workforces as a new line item. This is not a software budget. It’s not a headcount budget. It’s a new category. The companies that treat digital workforce spending as a rounding error on their IT budget will lose to the ones that treat it as a strategic capability investment.

Invest in orchestration, not point solutions. Don’t buy seven AI tools for seven workflows. Invest in a digital workforce partner that can deliver across your entire software and business operations stack. The value is in the done work, not the individual automation.

Start with high-volume, low-judgment work. The best use of digital workforces today is work that requires touching many systems, processing many transactions, and following established rules. Not strategic decisions. Not client relationship management. Not exceptions. Volume and velocity first. Judgment and nuance later.

The Punchline

The bears are right about one thing: AI changes everything. But their conclusion is wrong. Software isn’t going away. That’s the boring take. The interesting one is why.

Software isn’t surviving despite AI. It’s becoming more essential because of AI. Digital workforces get better the more software they can operate on. More digital workforce activity means more software consumption, more data generation, more API calls, more infrastructure load.

The enterprise stack is expanding. The companies that understand this will invest accordingly. Software companies will invest in becoming the best platforms for digital workforces to operate on. Digital workforce companies will invest in operating those platforms better than humans can. Enterprise buyers will invest in both.

The ones still debating whether AI replaces software are asking last year’s question. The question now is: how fast does the new layer scale, and who captures the value when it does?

The actual insight is weirder, deeper, and more useful than “pay attention to trajectory.” It starts with a question most people never ask: why do slopes persist?

Once you understand the answer to that question, you start seeing the world differently. You stop being surprised by outcomes that surprise everyone else. And you develop what I think is the single most underleveraged mental model in business, investing, and life.

Slopes are character. Intercepts are circumstance.

Here is the first thing people get wrong. They treat the current state of a system (its intercept) as a fact about reality, and its trajectory (its slope) as a prediction. One is real. The other is speculative.

This is exactly backwards.

The intercept is the noisiest variable. It’s the product of a thousand random inputs. Timing. Luck. Starting conditions. A company’s revenue in any given quarter is shaped by one-time deals, seasonal variation, macro conditions. A person’s net worth at 35 is shaped by which city they happened to live in, whether they graduated into a recession, what their parents could afford. The intercept is mostly circumstance.

The slope is structural. It’s generated by the deep properties of a system. Culture. Physics. Network topology. Feedback loops. Institutional design. These things change slowly. Which means the slope is actually the more durable signal.

Think about that. The thing most people treat as speculative (where it’s going) is generated by more stable forces than the thing they treat as certain (where it is).

A company with a strong slope and a weak intercept is a company with good structural properties that hasn’t had time to express them yet. A company with a strong intercept and a weak slope is a company burning through inherited advantage. Circumstance is being corrected by character.

This is why slope persists. Not because of some mystical momentum. Because the structural generators of slope, the culture, the compounding knowledge, the network effects, the cost curves driven by physics, change on a different timescale than the noise that determines intercept.

The second derivative is where the money is

Most people who do think about slope still think about it wrong. They think linearly. “It grew 20% last year, so it’ll probably grow 20% next year.”

The more important question is whether the slope itself is changing. The second derivative. Is the rate of change accelerating or decelerating?

This is where almost all asymmetric outcomes hide. Every S-curve in history has a period where the second derivative is positive (the slope is steepening) followed by a period where the second derivative is negative (the slope is flattening). The entire game of timing is figuring out where you are on that curve.

But here’s what makes this hard: humans are perceptually wired for linear extrapolation. Kahneman documented this. We anchor to recent rates and project them forward. We are structurally incapable of intuiting exponential curves or inflection points without doing the math.

In 2024, global AI investment crossed $200 billion. Up from around $40 billion five years prior. Most analysts projected continued growth. But the real question was never “will it grow?” It was “is the second derivative positive or negative?” Is each dollar of investment producing more capability than the last (positive second derivative) or less (negative second derivative)? That distinction is the difference between a $10 trillion industry and a $1 trillion one.

Solar energy is the canonical example. For decades, the cost curve had both a persistent negative slope (falling costs) and a positive second derivative (costs falling faster each decade as manufacturing scaled). Around 2020, the second derivative began to flatten. Costs were still falling, but more slowly. The slope was still good, but the slope of the slope had changed. This distinction matters enormously for capital allocation, and almost nobody in public discourse made it.

Slope stacking: where civilizations turn

Here’s the idea I think is genuinely underexplored.

Individual slopes are interesting. But the interaction effects between slopes are where the most consequential outcomes in history come from. I call this slope stacking.

When the cost curve of one technology intersects with the adoption curve of a dependent technology, and both intersect with a demographic shift or a policy change, each slope accelerates the others. These multi-slope systems create outcomes that look like “disruption” or “revolution” in hindsight, but are actually just the predictable result of compounding slopes.

The Industrial Revolution wasn’t one slope. It was the cost curve of coal extraction intersecting with the efficiency curve of steam engines intersecting with the urbanization rate intersecting with the literacy rate. Each slope fed the others. Cheaper coal made steam engines more economic. Steam engines made coal extraction cheaper. Urbanization concentrated labor. Literacy enabled knowledge transfer that steepened every other curve. No single trend caused the Industrial Revolution. The stack did.

The same pattern is playing out right now. AI capability curves are steepening. Energy demand curves are rising. The cost curve of inference is falling. The labor productivity curve in knowledge work is inflecting. The cost of launching satellites is dropping. The amount of data generated per capita is rising. Each of these is interesting individually. But they don’t exist individually. They interact. Cheaper inference means more AI agents. More AI agents mean more energy demand. More energy demand accelerates investment in generation. More generation investment steepens the solar and nuclear cost curves. Cheaper energy makes inference cheaper. The loop tightens.

People who analyze these trends in isolation will get the next decade wrong. The right mental model isn’t “AI is getting better” or “energy demand is rising.” It’s “what happens when these slopes multiply?”

History’s biggest mispricings occur when slope stacks are underway and the consensus is still evaluating each slope independently.

Why slope perception is asymmetric (and exploitable)

There’s a well-documented asymmetry in how humans perceive positive and negative slopes. Loss aversion means we notice decline faster than growth. A company losing 10% of its customers per quarter triggers alarm. A company gaining 10% per quarter gets a polite nod.

But there’s a subtler asymmetry that’s less discussed. We’re better at perceiving slopes in things we can count than in things we can’t.

Revenue slope? You can see it in a chart. Customer count slope? Same. But what about the slope of institutional trust? The slope of employee morale? The slope of technical debt? The slope of a founder’s judgment?

These are all real slopes generated by real structural forces. They compound. They persist. And they are nearly invisible in standard reporting.

The best investors I know spend most of their time trying to measure invisible slopes. They’re not looking at the dashboard. They’re looking at the rate of change in things that don’t have dashboards. How fast is this founder learning? Is this company’s engineering culture getting better or worse each quarter? Is the regulatory environment tightening at an accelerating or decelerating rate?

These unmeasured slopes are where the market is most inefficient. Because if a slope can’t be charted, most people act as if it doesn’t exist.

The slope trap

I’d be dishonest if I didn’t flag the failure mode.

Not all slopes persist. Some slopes are temporary expressions of a one-time force. A company growing 50% because it just got featured in the press. A country’s GDP spiking because of a commodity price surge. A person’s career accelerating because of one lucky break.

The question is always: is the slope generated by structure or by event?

Structure-driven slopes persist because their generators are durable. Physics-driven cost curves (solar, compute, genomics) tend to be structural. Culture-driven performance curves tend to be structural. Network-effect-driven adoption curves tend to be structural.

Event-driven slopes revert because their generators are temporary. Stimulus-driven economic growth reverts when the stimulus ends. Hype-driven adoption reverts when attention moves. Charismatic-founder-driven culture reverts when the founder burns out.

The discipline is in distinguishing the two. And the honest answer is that it’s hard. The best heuristic I’ve found: look at the slope over multiple timescales. If it persists across different macro conditions, different leaders, different market environments, it’s probably structural. If it only appears in one context, it’s probably event-driven.

The punchline is about probability, not direction

Here’s what I actually want you to take from this.

The claim isn’t “trends always continue.” They don’t. The claim is that the base rate of trend persistence is significantly higher than most people’s intuitive estimate.

Jegadeesh and Titman’s foundational work on momentum showed that stocks with positive 12-month returns continue to outperform over the next 3-12 months at statistically significant rates. This has been replicated across equities, bonds, commodities, currencies, and real estate in virtually every major market. Not because of magic. Because the structural forces that generated the trend (earnings growth, sector rotation, institutional allocation) operate on longer timescales than the trend itself.

The same principle operates outside of markets. Countries that are growing tend to keep growing. Companies that are improving their margins tend to keep improving them. People who are getting better at their craft tend to keep getting better.

Not always. Not forever. But more often and for longer than your gut tells you.

This means that when you evaluate anything, a business, a hire, an investment, a relationship, you should be giving more weight to the slope than you currently are. Not infinite weight. But more.

Most people’s mental model allocates maybe 30% to trajectory and 70% to current state. The correct allocation is probably closer to the inverse.

The intercept tells you where something is. The slope tells you what it is.

Bet on slope.