A follow-up to The Missing Variable. Same town, same war — one variable the first article didn’t count.
The “tokens as currency” frame is Jensen Huang’s. This article is what happens when you take it seriously.
Preface
The town had two families.
In Sergio Leone’s Per un pugno di dollari (1964), a small Mexican border town is being burned from both ends by two rival clans — the Baxters and the Rojos. A stranger rides in, sizes up the war, and sells his gun to both sides. He leaves with a fistful of coins while the town finishes burning.
A month ago I published The Missing Variable, a response to Citrini Research’s 2028 crisis scenario. That article mapped two families too. Call the Baxters Scenario A — consumer capitalism saved by some form of redistribution (UBI, credits, vouchers), digital feudalism with Prime delivery. Call the Rojos Scenario B — the oligarchic bifurcation, an automated circuit for asset owners, subsistence for everyone else.
What the first article didn’t ask is what currency the bodies were being counted in. Both families assumed dollars. Both could be wrong.
This article is the third reading of the town: the one where the money itself mutates while the two families are still shooting at each other.
One disambiguation before we start. When I say token in what follows, I mean an LLM input/output unit — the quantum of inference a model produces or consumes. Not an ERC-20 coin on a blockchain. Crypto has borrowed the word; where crypto tokens matter, I will name them explicitly.
I. What Both Families Assumed
Scenario A and Scenario B disagree about almost everything — who gets what, who owns what, who eats. They disagree about morality, about the state, about the shape of the world on the other side of the crisis. What they quietly agree on is the unit of account.
The Baxters hand out UBI in dollars. The Rojos buy farmland and nuclear plants in dollars. The lobbying is denominated in dollars. The pensions that get wiped out in Phase 2 are denominated in dollars. The transition mechanism of the missing-variable article — financial destruction, real-asset preservation — is a story told entirely inside one currency.
But fiat is not a natural fact. A dollar is a claim on labor-hours in a labor economy. It’s a receipt for an hour of somebody doing something. Take the labor economy away — which is the whole premise of the 2028 crisis — and the receipt keeps printing while the thing it names evaporates. The bullet still comes out of the gun, but there’s nothing on the other end of it.
This is the gap the stranger rides into. He looks at the two families shooting each other in a currency that no longer has a referent, and starts quoting his own prices in something else.
That something else, according to the chorus of the last two years, is a token.
II. Electricity In, Tokens Out
Huang’s compressed version, from the Dwarkesh Podcast in 2026:
The input is electron, the output is tokens. That in the middle is Nvidia.
Four sentences of corporate philosophy collapsed into one. At Nvidia’s GTC keynote earlier in 2026 he pushed it further: AI factory revenue = tokens-per-watt. Every unused watt is revenue lost. A data center stops being a building full of computers and becomes a machine for converting megawatts into measurable cognitive work.
The part worth lingering on isn’t the philosophy. It’s that Huang is already running his company on this basis. His internal heuristic, widely quoted in early 2026: an engineer earning $500K a year should be consuming around $250K a year in tokens. If they aren’t, they’re underutilizing the tool.
That ratio is a budget line. Not a metaphor, not a forecast — a line item a CFO can enforce. Tokens have climbed into the operating cost structure of the most capital-rich companies in the world. They sit next to salaries, rent, and electricity in the P&L. Deloitte’s 2026 note on “AI spend dynamics” describes the same thing from the outside: tokens migrating up the corporate budget hierarchy, crossing hourly labor on the way up.
The rest of the chorus made the same noise in 2026:
+-------------------+------------------------------+-----------------------+
| Actor | Claim | Source |
+-------------------+------------------------------+-----------------------+
| Jensen Huang | tokens = new commodity | Nvidia GTC 2026 |
| Jensen Huang | electrons -> tokens | Dwarkesh Podcast 2026 |
| Morgan Stanley | compute as new economic | MS research note, |
| | engine | 2026 |
| Deloitte Insights | tokens climbing corporate | "AI tokens spend |
| | budget hierarchy | dynamics" |
| AEI | "rise of tokenomics" | aei.org |
| SemiAnalysis | tokenomics cost model | semianalysis.com |
| Nvidia blog | tokens = AI currency | blogs.nvidia.com |
+-------------------+------------------------------+-----------------------+Nvidia has a commercial interest in pushing the frame. I’m going to ignore that. The idea stands or falls on whether a token behaves like money, not on who wants it to.
III. Is a Token Actually Money?
Classical money has three functions: unit of account, medium of exchange, store of value. A token passes one and a half. But it picks up a fourth function fiat never had.
Unit of account. ✓
Already done. API pricing is denominated in tokens. Enterprise budgets carve out token allocations. Deloitte’s 2026 note describes tokens sliding into corporate P&Ls alongside hourly labor cost. When a finance team at a big bank or a hyperscaler prices a project, they price it partly in human hours and partly in tokens. The unit of account has already switched. Most people haven’t noticed because the invoice still prints a dollar number at the bottom.
Medium of exchange. ½
Between humans, tokens don’t change hands. You pay your vendor in dollars; the vendor pays the lab in dollars; the lab produces the tokens. That’s a supply chain, not a payment network.
Between AI agents it’s a different story. The machine-to-machine agent economy — autonomous wallets, agent-to-agent settlement rails — is small but real in 2026. When two agents transact, they settle in a token-denominated flow with no human in the loop. That’s genuinely new. The first medium-of-exchange use of tokens is horizontal, between non-humans. It spreads from there.
Store of value. ✗
Tokens are perishable. Consumed at inference time, gone afterward, no natural storage form. A token is closer to electricity than to gold — a flow, not a stock.
But prepaid token balances are starting to look like a storage form. OpenAI credits, Anthropic workspace budgets, enterprise reserved capacity. Compare M-Pesa in Kenya in the 2000s: prepaid phone minutes became a de facto currency because the balance was transferable and the underlying unit was universally valued. Same shape. Prepaid inference balances are the M-Pesa minute of 2026.
Unit of cognitive work. (New.)
Fiat never had this. A dollar measures claim-on-labor with no intrinsic productivity dimension. A token measures a quantum of intelligence-production. That’s unprecedented as a monetary unit. It’s the reason Huang can say engineers should consume $250K of tokens a year — he’s denominating productivity in a unit fiat can’t express.
The conclusion: a token is not money yet, but it is already a parallel unit of account for cognitive work. That’s the slot from which historical currencies have actually emerged — tallies, beaver pelts, rice, salt, cigarettes in POW camps. All started as units of account. The medium-of-exchange function followed. The store-of-value function showed up last, usually through a derivative instrument built on top.
IV. Can You Buy Bread With Tokens?
Three concrete cases, cheapest to most structural.
Bread, about €2.
At 2026 rates, €2 buys somewhere between 200,000 and 2,000,000 tokens of a mid-tier model, depending which model. Technically there’s no obstacle to running that transaction — a prepaid token balance plus a card-shaped settlement rail gets you there.
The problem is the baker. Tokens are only valuable to her if she can use them for something she needs. If she uses AI in her business — supply-chain forecasting, accounts, regulatory filings, marketing generation — she already has a token demand and she can absorb the payment. If she doesn’t, tokens are dead weight; she’ll refuse, or convert at a haircut that eats her margin.
The threshold is mechanical. Bread-for-tokens becomes viable the moment small-business operating cost is materially token-denominated. A plausible guess is that the sectors with heaviest AI adoption — logistics, retail back-office, professional services — cross that line within five years. Bread-for-tokens at the village bakery is further out.
Rent, about €1,000 a month.
The landlord already pays for tenant-screening AI, maintenance-scheduling AI, legal-generation AI, accounting AI. Token exposure on the landlord side is already high. Rent paid in tokens has less conversion friction than bread paid in tokens. This is probably the first major consumer-scale use. Not tomorrow. But not a decade either.
A house, about €300,000.
A different problem entirely. A house is a store of value; tokens are perishable. You can’t denominate a store of value in something that evaporates when used.
You can denominate it in a claim on future token production. A bond whose coupons are paid in tokens generated by a specific AI factory. Structurally identical to oil-linked bonds or electricity futures. There’s no conceptual barrier; the instrument is buildable today.
Who would issue it? The actors already positioned: Saudi PIF, Abu Dhabi’s MGX, BlackRock’s GAIIP, hyperscalers running their own infrastructure debt. The Stargate project is essentially a token-production bond in disguise — a capital vehicle whose ultimate collateral is future inference throughput.
The pattern is simple: tokens work for flow, claims-on-tokens work for stock. Same split as electricity versus energy futures. You don’t store electricity; you store the right to receive it.
V. The Flow — Who Produces, Who Clears
Map the actors explicitly.
Producers of tokens. Physical side: Nvidia (chips), TSMC (fabrication), hyperscalers running the data centers (Microsoft, Google, Amazon, Meta, Oracle), Gulf sovereigns siting and financing them, utilities and nuclear operators supplying the electrons. Cognitive side: the model labs — OpenAI, Anthropic, Google DeepMind, xAI, DeepSeek — wrapping compute into callable inference.
Issuers of denominated balances. The model labs and the API aggregators. These are the banks of the token economy. They issue the balance, honor redemption, set the exchange rate against fiat. OpenAI credits are a liability on OpenAI’s books in the same way a deposit is a liability on a bank’s books.
Consumers. Corporations first — already happening, the part Deloitte writes about. Individuals via subscriptions today, possibly via wallets tomorrow. And a fast-growing third class: other AIs. Machine-to-machine settlement is the segment nobody has a chart for yet, and it is probably the one that matters most.
Clearing and settlement. Today, conversion to fiat happens through corporate credit cards and invoicing — slow, high-friction, intermediated. Tomorrow, plausibly either (a) native token-to-token clearing between labs, or (b) a crypto-based settlement layer riding on the AI-agent wallet narrative. Pick your flavor of infrastructure; both are being built.
Regulators. As of April 2026, nearly absent. BIS, ECB, Fed have not, as far as I can verify, published a position on token balances as a monetary aggregate. Marked unverified pending primary-source check. The absence is itself informative. The last time a parallel unit of account grew this fast with no central-bank position was the eurodollar market in the 1960s, and the consequences took two decades to surface.
[ energy ] --> [ chips ] --> [ data centers ] --> [ models ]
|
v
[ token production ]
|
+-----------------------+-----------------------+
| | |
corporate consumers consumer subscriptions M2M agent economy
| | |
+-----------------------+-----------------------+
|
v
[ clearing / settlement ]
(fiat today,
token-native tomorrow?)VI. One Token or Many?
There will not be one coin. There will be several, stratified by what the token can do.
- Reasoning tokens — the expensive ones. Hidden chain-of-thought, high wattage. Gold: scarce, high unit value, used to settle the largest questions.
- Plain inference tokens — everyday cash. The bulk of the supply, low unit value, high velocity.
- Image, audio, video tokens — specialized currencies tied to specific media. Historical analog: occupational monies. The tin tokens miners used to spend at the company store.
- Agent-action tokens — tool-use, browsing, code execution. Closer to gas in a blockchain system than to money. Transaction fees, not medium of exchange.
- Verification tokens — proof-carrying inference, attested provenance. Priced at a trust premium because they embed attestation.
Medieval Europe ran on this kind of plurality — gold for sovereigns, silver for merchants, copper for peasants, with floating exchange rates between them and nervous money-changers doing the conversion. The token economy is heading toward the same shape. Reasoning tokens convert to inference tokens at some rate. Inference tokens convert to agent-action tokens at another. Each class has its own supply, its own issuer, its own demand curve.
Whoever operates the frontier model sets the value of the top-tier token. Whoever controls the conversion between classes is doing central-bank work. Not metaphorically — literally: setting the rate between reasoning-tokens and inference-tokens is the same operation as setting the rate between high-powered money and broad money in a fiat system.
The missing-variable article flagged this in passing: AI companies would wield influence comparable to today’s central banks. This is the mechanism. Not influence over central banks. Being the central bank.
VII. The Stranger Sells to Both Families
The payoff. In the film, the stranger sells his gun to the Baxters, then sells it to the Rojos, collects from both, and lets the two families destroy each other. The town burns while he counts.
Map it to 2028.
Scenario A in token register. Redistribution happens. A UBI is issued. But it’s issued in tokens, not dollars. An allowance of inference — consumable, non-accumulable, expiring. The Baxters think they won: there’s a safety net, the population is kept fed, the unrest is contained. But the payment rail belongs to the issuer. Dollars can buy land; tokens can only buy inference. You can’t stack tokens into ownership of anything. Dependents by design. Scenario A dressed down to Scenario B’s terms. The Baxters paid in the stranger’s coin and never noticed.
Scenario B in token register. Claim-on-token instruments become a real store of value. The oligarchy converts financial wealth into tokens-as-bonds — contracts on future AI factory throughput. Whoever issues the instrument sets its yield, its supply, its convertibility. Saudi PIF, MGX, BlackRock, Nvidia itself — the candidates already named in the missing-variable article. Whoever issues the dominant token-bond class is the post-sovereign central bank. The Rojos paid in the stranger’s coin and knew it.
The stranger takes from both families. That was Leone’s joke, and it stops being a metaphor here. The same actor — the compute owner, in practice a handful of hyperscalers and sovereigns — is on both sides of the transaction. Issues the redistribution currency to the Baxters. Issues the settlement instrument to the Rojos. Profits from both flows. The town burns in fiat.
The middle-class exposure is the part that should scare a reader. In 2008, the middle class held the old settlement asset — housing equity — while the new settlement asset was being minted elsewhere. They absorbed the destruction and bought the recovery at full price. Phase 2 of the missing-variable article is the same move at larger scale: financial destruction, real-asset preservation. But Phase 2 had only one loss layer — asset value. A token transition adds a second: denomination obsolescence. You lose not only the value of your savings but the currency they’re denominated in. Two bullets instead of one.
The steel chest plate under the stranger’s poncho is still doing the same work. In the film, Ramón empties his Winchester into it and the stranger stays on his feet. In the missing-variable article, the oligarchy’s conversion of financial wealth into real assets — farmland, nuclear, compute — is the plate. Fiat bullets don’t punch through. Ramón keeps firing; fiat-denominated consumer demand runs out of referent as labor stops being the productive input; the stranger keeps standing.
Currency is usually the last thing people imagine changing. That’s why, when it changes, it finds most people with the wrong kind of money in their pockets.
VIII.
The town had two families. Both were paying the stranger. Neither was counting in the right currency.
Data Sources and Methodology
This analysis draws on publicly referenced material including: Jensen Huang, Nvidia GTC 2026 keynote; Jensen Huang on the Dwarkesh Podcast (2026); Nvidia blog, “AI tokens: the language and currency powering modern AI”; Morgan Stanley, “NVIDIA’s Jensen Huang on compute as a new economic engine” (2026); Deloitte Insights, “AI tokens: how to navigate AI’s new spend dynamics”; American Enterprise Institute, “Algorithms, Compute, and the Rise of ‘Tokenomics’”; SemiAnalysis, “Tokenomics Model”; Computerworld reporting on Nvidia tokenomics framing; RCRWireless, “Agents, inference and the new token economics”; Tom’s Hardware coverage of Huang’s engineer-token-consumption heuristic; long-form pieces by Eduardo Alvarez (“What are LLM tokens worth?”), Gilad Barkan (Wix Engineering, “The Emerging Economy of LLMs”), zhaolongzhong (“Summary of LLM Token Economies”), and the Peerism “Skill Token Economy” essay. Where a direct quote is used and the primary transcript has not been cross-checked at time of writing, the passage is marked unverified.
Internal reference: The Missing Variable: Oligarchic Agency in the 2028 Global Intelligence Crisis.