The Mature Token Law: From "Whoever Spends the Most Tokens Wins" to "Whoever Converts Them Wins"

There is a specific pleasure in being quoted, and a specific discomfort in being quoted accurately for three years straight. In 2023 I wrote The Agent Token Manifesto. Its chapter on token economics made a claim designed to be argued with: your CFO's instinct to minimise token consumption will destroy your competitive position, because in Software 3.0 the companies that burn more tokens strategically will dominate their industries — token spending isn't a cost centre, it's your R&D budget for continuous evolution.

The compressed version travelled further than the chapter did: whoever spends the most tokens wins.

That sentence has now been stress-tested by the entire industry, loudly, in public, with nine-figure invoices attached. The test results are worth being honest about. The instinct survived. The maximand did not. A manifesto that preaches compounding learning and then refuses to update its own most-quoted claim would be ignoring its own doctrine — so this piece does in public what the doctrine says to do with every evaluated output: feed it back in, and ship the stronger version.

Here is the mature form of the law, the reasoning that forces it, and the test for whether your token budget is buying advantage or just heat.

The law we wrote, and what it actually said

It matters that the original claim was not a licence for reckless spending, because the revision is not a retraction. Chapter 5 of the manifesto said strategic token burning requires clear objectives, bounded constraints, value validation — "agents that burn tokens without progress get adjusted or terminated" — and reinvestment loops that route budget toward what works. The seed of conversion was already in the text. But the headline optimised for bluntness, and bluntness is what propagated. The law as remembered — as quoted in pitch decks, as implemented in engineering orgs — was: more burn, more win.

It was deliberately blunt because the enemy at the time deserved bluntness. The enemy was the instinct to treat machine cognition as a metered utility: minimise consumption, wait for prices to fall, pilot cautiously, keep the invoice small. Against that enemy, the law holds up fine. What nobody had yet seen in 2023 was the second failure mode — the one that appears after an organisation takes the advice.

Then the industry ran the experiment

Between February and May 2026, the law's compressed form escaped into the wild under a new name: tokenmaxxing. Y Combinator partner Diana Hu made it advice: "Maximizing token usage, not head count, will be the critical shift. The best companies will be the ones that are tokenmaxxing."¹ Jensen Huang told the All-In podcast he would be concerned if an engineer earning $500,000 did not consume at least $250,000 worth of tokens.² Individual burn became a badge: reports of an OpenAI engineer processing 210 billion tokens in a week, an Anthropic employee running a $150,000 monthly Claude Code bill.² Someone finally built the thing the blunt law implies: a scoreboard. Meta's internal leaderboard — nicknamed "Claudenomics" — tracked token usage across 85,000 employees. In one 30-day window they consumed 60.2 trillion tokens, roughly $900 million at public API pricing, "with the vast majority attributed to tokenmaxxing rather than productive work."³

What did the burn look like up close? "They run AI agents on repetitive, meaningless tasks. They ask questions about documents they have already read. They build feature prototypes that will never ship. They let agents run until they error out, consuming tokens the entire time."³ Amazon shut down its own usage leaderboard after discovering employees inflating their scores the same way; its SVP of engineering had to write the sentence "Please don't use AI just for the sake of using AI."³ Engineering-analytics firm Jellyfish measured the shape of the curve: the top 10% of users burn roughly ten times as many tokens as average developers and produce about twice the output.³ Uber's COO said the company has not found a clear link between increased AI spending and the delivery of successful products.³

The backlash wrote itself. HubSpot's CEO: "Outcome maxxing >> token maxxing."⁴ Appian's CEO compared tokenmaxxing to the Soviet practice of judging chandeliers by their weight.⁴

Here is the part that matters, and it is the reason this piece is a revision rather than a surrender: both sides of that debate are right, and neither has named the structure. The proponents are right that more cognition matters — but they have not said what shape of spending compounds. The critics are right that raw burn is a gameable vanity metric — but "outcomes" is not an operational alternative; it is a gesture at one. The proponents named the directional claim. The critics named the symptom. The gap between them is where the mature law lives.

The instinct still stands

Before the revision, the defence — because the easiest misreading of this piece is "so we can cut the AI budget," and that reading loses just as surely as the leaderboard did.

The economics that motivated the original law have only steepened. Epoch AI's measurements across benchmarks found LLM inference prices at a fixed capability level falling between 9x and 900x per year, with a median around 50x.⁵ Gartner forecasts that inference on a trillion-parameter model will cost over 90% less by 2030 than it did in 2025 — and, in the same breath, that total inference spend will rise, because agentic workloads consume 5–30x more tokens per task than chatbots and consumption is outrunning unit cost.⁶ The volume curve agrees: Google's Gemini processed 480 trillion tokens a month in May 2025 and 3,200 trillion a month by May 2026.³ The leaderboards died; the burn kept growing. That is not a contradiction. It is the sorting mechanism.

So let me restate the part of the original law that survives, in the original spirit: cognition should be funded, not rationed. A company trying to minimise token consumption while its competitors use machine intelligence as continuous research and development is optimising the wrong resource. The manifesto's own urgency chapter made a prediction that has aged well: token costs would fall for everyone, so cheap tokens themselves would confer no positioning — everyone gets them. If the fuel gets 50x cheaper every year for every player, the fuel cannot be the advantage. Something else has to be.

What broke: undirected cognition thickens the fog

The failure mode that the original law did not name is now measurable from three directions at once.

Start with the mechanism. AI moved the scarcity, and it moved it asymmetrically:

When producing an answer required weeks of human effort, the ability to produce answers was valuable. When anyone can produce a hundred plausible answers before lunch, the advantage is no longer answer number 101. The advantage is choosing the question worth answering. It is recognising which possibility deserves to become real. It is rejecting attractive but structurally weak futures. It is making a consequential decision while the opportunity to act still exists.

Burn without that chooser does not merely waste money. It actively degrades the environment the decisions have to be made in: undirected cognition generates more possibilities and thickens the fog. Every unfiltered draft, every speculative analysis nobody asked for, every agent left running until it errors out adds to the pile that someone's scarce attention must now triage.

The external record has caught up with each clause of that claim. The MIT Media Lab's Project NANDA report found that despite $30–40 billion in enterprise GenAI spending, 95% of organisations are seeing no business return — while over 80% have piloted the tools.⁷ Adoption is nearly universal; conversion is rare. And the fog has a name now: researchers at BetterUp Labs and Stanford called it "workslop" — AI-generated content that masquerades as good work but lacks the substance to advance the task.⁸ Forty per cent of surveyed workers had received it in the previous month, and each piece cost its recipient an average of nearly two hours to deal with — a roughly $186-per-month invisible tax per employee, about $9 million a year across a 10,000-person organisation.⁹

Read that carefully, because it is the old law's failure mode in one statistic: tokens were spent lavishly, and the cost landed on the one resource that cannot be bought at any per-token price — a colleague's attention. Undirected generation does not retire judgement. It relocates the burden of judgement downstream, with interest.

Even the machine-side research says the same thing in its own dialect. Test-time-compute scaling — spending more inference per question — genuinely buys capability, which is why the "fund cognition" instinct survives contact with the frontier. But the same literature documents formal saturation points, beyond which additional samples plateau; and the more elaborate undirected strategies "tend to concentrate errors on consistent wrong answers," so that majority voting stagnates.¹⁰ More burn past the point of saturation does not just buy nothing; it can buy confident agreement on the wrong answer. The scaling papers found what the leaderboards found: raw volume is not the variable.

Why the variable moved: everyone got the same fuel

Two structural facts force the revision. Neither is a matter of taste.

First: the fuel is symmetric. Cheap cognition is not a private input. The same price collapse that funds your analysis is funding your customers' analysis, your competitors' redesigns, and the entry cost of your next attacker. Cheap thinking does not make strategy easier. It creates more affordable moves for everyone: customers, competitors, constructors and companies that do not exist yet. A resource that every player can buy at a collapsing price cannot be the axis of advantage — it can only be the ticket to the table. Nvidia's own framing at GTC 2026 makes the point from the supplier's chair: "tokens are your new commodity… this is your token factory, this is your revenue."¹¹ For the vendor, the token is the product and burn is revenue. Importing the supplier's metric as your strategy is how you end up with a chandelier weighed by the kilogram.

Second: the complement is fixed. MIT economist Christian Catalini states the mechanism plainly: "As AI becomes capable of executing more work at machine speed, the scarce resource moves from intelligence to verification… We have built an engine that scales faster than its brakes."¹² Agent fleets reproduce at compute speed; human expertise grows at biological speed. Andrej Karpathy's version, from the practitioner's chair: the scarce thing is shifting toward understanding, taste, eval design and knowing when the model is off the rails — "You can outsource your thinking, but you can't outsource your understanding."¹³ The developer data shows the squeeze mechanically: teams using AI merge 98% more pull requests — and review time per PR rises 91%, with PRs 154% larger.¹⁴ The work did not disappear. It moved from generation, which is abundant, to judgement, which is not.

Put the two facts together. One input to consequential cognition — machine generation — is abundant, symmetric, and getting 50x cheaper a year. The other inputs — the question chosen, the evidence gathered, the decision made, the capability retained — are scarce, asymmetric, and priced in a currency no provider discounts. When one input to a production function becomes free, advantage moves entirely to the complements. The law had to be restated in terms of the complements, or it would start selecting losers.

The mature token law

Our early thesis was deliberately blunt: whoever spends the most tokens wins. The instinct remains right. Cognition should be funded, not rationed. But token burn alone is not an advantage — undirected cognition generates more possibilities and thickens the fog. The mature law is:

Whoever converts the most tokens into better questions, faster evidence, stronger decisions and compounding capability wins.

And its operating rule fits on an index card:

Spend machine cognition lavishly.
Spend human attention ruthlessly.
Maximise consequential cognition.

Notice what the asymmetry does. It preserves everything the original law got right — the machine side of the budget stays lavish, uncomfortable, R&D-shaped. And it names what the original law left implicit: the scarce budget was never the tokens. It was always the human attention deciding which of the machine's outputs deserves to touch reality. Conversion is the discipline of pointing a fixed pool of attention at an effectively unlimited pool of generation.

The law has four roles in it, and they are roles, not synonyms:

• Tokens are the fuel. Necessary, purchasable, and getting cheaper for everyone on a published schedule. Fund them like R&D, exactly as the original manifesto said. Just stop reading the fuel gauge as a speedometer: litres consumed was never the measure of a journey.
• Judgement chooses the direction. The question is the highest-leverage artefact in the whole system, because every downstream token inherits it. A thousand agents pointed at the wrong question produce fog with excellent production values. Choosing the question worth answering — and rejecting the attractive but structurally weak futures — is where human attention earns its keep. This is the discipline our canon develops as disciplined cognition: spend inside an apparatus that knows what to reject.
• Evidence establishes contact with reality. Generation that never meets a test is opinion at scale. The conversion step is forcing each candidate answer through something that can refute it — a dataset, a customer, a deployed system, an adversarial check — before it consumes anyone's attention. Catalini's caution applies here too: a checker that shares the generator's blind spots produces synthetic confidence, "one model's plausible mistake becomes another model's approved answer."¹² Evidence means contact with something outside the model.
• Capital determines whether the journey actually begins. A converted token chain ends in a decision that moves money, people, or a roadmap. If the analysis was brilliant, the evidence real, and nothing was funded, stopped, or changed — no conversion happened. The fourth output, compounding capability, is what the decision leaves behind: the evaluations, priors, and validated judgement that make the next cycle start further ahead. That residue is the subject of its own piece in this series, on callable assets.

Restated as a budget instruction: your token line should keep growing. What changes is the ledger it reports to. Not "tokens consumed" — questions retired, evidence returned, decisions changed, capability accrued. Those are the four things a token is for.

The same budget, spent twice

The brief version of the proof is a case from our own engineering practice, documented in Designing Loops, Not Prompts. Same task — auditing pull requests. Same class of budget. Two shapes of spend.

Loop A — the undirected version. Fifty-five agents run a tournament over three candidate pull requests. No verify phase; the agents are scored against each other. Every output is fluent. The token meter spins impressively. The result, in our own post-mortem's words: $400 of agents agreeing. No question was sharpened, because the question was never chosen — "which PR do the agents prefer?" is not a question about the repo. No evidence came back, because nothing the agents said was ever tested against the code that actually had to run. No decision improved, because consensus among correlated judges is not information — strip the refutation step and a tournament is just expensive agreement. And nothing compounded: the next audit starts from zero.

Loop B — the converted version. Five agents, structured as audit → rule → verify. The question is explicit: does each finding survive refutation against the real repository? Priors from the last run are loaded at the start — the loop begins where the previous one ended. Each ruling is attacked by an adversarial checker whose job is to break it against reality before a human ever reads it. The result: the same useful answer at a tenth the cost — and it left priors behind for the next run.

Walk the four roles through both loops and the mature law stops being abstract. Same fuel. Loop A converted none of it: no chosen question, no contact with reality, no decision a human could trust, no residue. Loop B converted all four: the question was selected, the evidence was adversarial, the decision was grounded, and the capability — the priors, the rules, the checker — carried forward. The difference was never the token count. Loop A burned more.

The industry's own corrections are converging on the same shape, one output at a time. Amazon replaced its token leaderboard with "normalized deployments" — AI-assisted code that actually ships.³ Salesforce introduced "agentic work units" to translate tokens and compute into completed work.⁴ Jellyfish advises tracking cost per merged pull request rather than total consumption.³ Each of these is one cell of the conversion ledger — shipped code is a decision that survived evidence; a work unit is a retired task. None of them yet names the full chain from question to compounding capability. That chain is the law.

Running the mature law: the conversion audit

The restated takeaway is that you can now audit a token budget the way you audit any other capital deployment. Four questions, asked of the last 30 days of spend:

Two pointers for the "how", because the mechanics are developed elsewhere in this canon and re-teaching them here would break the piece's own rule about spending attention ruthlessly. The operational version of "stop burning where nothing converts" is the saturation rule: when a better model stops producing a better result on your use case, you have hit a ceiling that more tokens will not fix — undisciplined burn is usually brute-forcing a saturating rung, and the move is to change the dimension of the work, not the volume. And the loop-level allocation discipline — where to point spend across phases, routing and independent verification — is developed as token discipline in Designing Loops, Not Prompts, whose summary sentence is the one to carry: the token-maxing is the theater; the apparatus is the asset.

Why we revise in public

A fair question: why not quietly stop saying the old line? Because the manifesto's core claim was never really about tokens. It was about compounding: evaluated output becomes durable input to the next cycle, and organisations that run that loop pull away from organisations that reset. A doctrine that runs that loop on its products but not on itself is decoration. The original law has now been through the most expensive evaluation cycle the industry could have designed — leaderboards, nine-figure burn, a public backlash — and the evaluated output is in. Feeding it back into the law is the manifesto behaving as specified.

It is also, frankly, the pattern this whole body of work keeps finding: the blunt claim first, the correction second, and the doctrine stronger for the correction. A claim that arrives fully formed has usually had its qualifications quietly removed. This revision has siblings: the same abundant-cognition shift that matured the token law also re-rolls every company's inherited advantages at the industry level — that argument is made in The Re-Roll — and determines which of your assets AI can actually multiply, which is The AI Carry-Forward Test. Each piece revises a different part of the same early instinct. None of them repudiates it.

The 2023 law told you which resource to stop rationing. It was right, and it is still right. The 2026 law tells you what that resource is for. Whoever spends the most tokens buys the most raw cognition. Whoever converts the most tokens — into better questions, faster evidence, stronger decisions and compounding capability — wins.

FAQ

So the original law was wrong?

No — it named the right resource and a blunt maximand, on purpose, against an enemy (token rationing) that deserved the bluntness and still does. The chapter that carried it already required value validation and reinvestment; what the compressed line lost was the conversion structure. The revision moves the variable being maximised, not the instinct. If you read this piece as permission to cut the AI budget, read it again: the rationer loses to the converter just as surely as the burner does.

Isn't "conversion" just "spend wisely" in a suit?

"Wisely" has no units. Conversion has four: questions retired, evidence returned, decisions changed, capability accrued. You can put those on a dashboard next to the token line and audit them monthly — that is the difference between an adverb and a law. The Loop A/Loop B case shows the units doing real work: identical task, and the four-column ledger cleanly separates the $400 of agreement from the answer at a tenth the cost.

Does conversion slow the machine side down?

No. Nothing in the mature law throttles generation — the asymmetry explicitly says to spend machine cognition lavishly, and the volume data — total token volume growing more than sixfold year on year even as the leaderboards died³ — is the law working, not failing. Conversion disciplines where the fixed resource lands: human attention goes to choosing questions, weighing evidence and making decisions — not to triaging fog.

Where do I learn the actual mechanics?

They are already written, deliberately not re-taught here: disciplined cognition and the apparatus that can reject (The Cognition Dimension Ladder), loop-level token discipline — phases, routing, independent verification (Designing Loops, Not Prompts), and what happens to strategy when everyone's thinking gets cheap at once (Cheap Thinking Makes Strategy Harder). This piece is the keystone that connects them back to the law they matured out of.

What should I do this week?

Run the conversion ledger over your last token invoice — the four questions in the box above, answered in writing, with names attached. If the answers are strong, raise the budget. If they are weak, do not cut it; build the verify phase and the question discipline first, then raise it anyway.