The AI Token
Standard.

Joshua Haldeman, Elon Musk's grandfather on his mother's side, ran the Canadian arm of Technocracy Incorporated from 1936 to 1941. He wore a grey-green uniform, called himself by the registry number 10450‑1, and went to court for advocating that the nation print energy‑certificates instead of dollars. The party that descended from those certificates is, today, a footnote in the intellectual history of money. The grandson is, today, the second-richest person on earth.

The plan, drafted by a possibly-self-invented engineer named Howard Scott, was straightforward: count the kilowatt-hours North America produced in a year, divide by the adult population, issue each citizen a non-transferable book of receipts, and price every good — bread, surgery, an hour of education — in joules. Money would be replaced by a measurable, physical, productive unit. The grey-green uniforms were optional.

Almost a century later, Haldeman's grandson built the most expensive kilowatt-hour-consuming machine in the history of human industry and called it a factory. The buyer of those kilowatt-hours, however, is not Tesla. The buyer is a five-trillion-dollar company in Santa Clara whose CEO has read every Christensen book ever published, whose product is a chip, and whose actual business — though no one in the value chain quite says it out loud — is the manufacture of tokens. Words. Vector embeddings. Generated thoughts.

What the book argues. What the book leaves implicit.

Slava Solodkiy's The NVIDIA Innovator's Dilemma is, formally, a Christensen-style audit of a five-trillion-dollar monopoly. Its argument runs like a watch.

NVIDIA's seventy-five-percent gross margin functions as a tax on its largest customers. Every hyperscaler CFO has a fiduciary obligation to evade that tax. The math of custom silicon is overwhelming and recurring: a one-to-three billion-dollar ASIC programme that buys back twenty billion of annual margin compounds in favour of "build" rather than "buy." The "job to be done" in AI compute has shifted from frontier-model training (where NVIDIA's interdependent architecture is unbeatable) to inference at scale (where it is over-engineered). The CUDA moat is being routed around, not breached, by Triton, vLLM, MLIR, and the "DeepSeek moment" of April 2026.

Five new value networks are emerging at the edge of the market that NVIDIA's organisation cannot pursue from inside its own profit-and-loss statement — DePIN, brownfield energy, sovereign AI and metastates, digital identity, and untapped human capital. The Christensen prescription is classical: spin out autonomous units with separate P&Ls, separate metrics, separate physical locations, and explicit permission to ignore the parent's seventy-five-percent margin defense.

This is correct. It is also incomplete.

The book, read carefully, contains a deeper argument that it does not quite make explicit, possibly because making it explicit would require a different book. The deeper argument is about what the NVIDIA economy is actually denominated in — and therefore what the next reserve currency is going to be.

Each constraint priced. Each repriced.

Read the book's chronology against itself.

In 2023, the binding constraint on artificial intelligence was chips. By late 2024, NVIDIA had ramped Hopper and Blackwell production fast enough to clear that bottleneck. The new constraint became power: a five-hundred-megawatt data centre in Loudoun County, Virginia, could not be powered before 2032 because the grid interconnection queue had collapsed.

But by the time The NVIDIA Innovator's Dilemma went to press, the bottleneck had migrated again. The chips were available. The power was being repowered. The new binding constraint was memory — specifically, high-bandwidth memory, and behind it, the entire DRAM stack. Solodkiy's parallel project, the DRAM.GOLD appendix, exists because by 2025 a modern AI data centre is, increasingly, a vault of high-bandwidth memory with some silicon attached.

Three bottlenecks in three years: chips, power, memory. Each was the real constraint at the moment everyone was solving the previous one. Each got priced. Each got repriced. The market kept moving.

The interesting question is what gets bottlenecked next. And the answer, the moment you write it down, is unembarrassed by ambiguity: result.

The token already has a price. Treat it as the unit of account.

Begin with what an "AI token" actually is, technically. It is a unit of model output: the smallest billable atom of generated text or vector. It already has a price — denominated in dollars, paid per million, varying by model. It already has a supply curve, varying by silicon. It already has a market structure, varying by hyperscaler. By the most ordinary economic test, the token already functions as a unit of account inside the AI value chain. The question is only whether anyone will say so out loud.

Here is the move. Treat the token not as a metric but as the unit of account the AI economy is implicitly already using.

What is each link in the AI value chain actually selling? NVIDIA sells silicon that produces tokens. The hyperscaler sells the kilowatt-hour that produces tokens. DePIN sells the consumer GPU that produces tokens. The brownfield repower sells the megawatt that produces tokens. The annotation labour pool sells the human-feedback signal that improves tokens. The model lab sells the architecture that cheapens tokens.

The fact that they look like five different markets is an artefact of the value chain not yet having rationalised onto its natural unit.

Right impulse. Wrong unit. Wrong substrate.

Howard Scott was not stupid. Frederick Soddy — the 1921 Nobel laureate in chemistry who proposed in Wealth, Virtual Wealth and Debt (1926) that money should be tied to physical energy — was not stupid. Buckminster Fuller, who described energy as "the true currency of the universe" decades later, was not stupid. They had the structural impulse right. They had the unit wrong.

Almost. Three things broke their version.

// FAILURE MODE 01 · METER GRANULARITY

A kilowatt-hour can be measured at the meter. It cannot be tagged at the moment of useful use. A kilowatt-hour of light pumped into an empty room is, by their accounting, identical to a kilowatt-hour of light pumped into a room where someone is having a brilliant idea. The unit is too coarse to discriminate.

// FAILURE MODE 02 · UNIFORMITY

Money has to discriminate. A pound earned by a Bangladeshi garment worker buys the same loaf as a pound earned by a London banker; that is fine, that is the unit of account doing its job. But a pound of meaningful work discriminates fiercely, and the technocratic certificate could not encode that discrimination. The unit was too uniform.

// FAILURE MODE 03 · SUBSTRATE

The technocrats had no programmable layer on which to issue their certificates. They proposed paper. They had to. Paper money has been an IOU since the Mesopotamian temple receipts of the third millennium BC, but paper is bad at conditional logic. A kilowatt-hour certificate that is supposed to expire in sixty days, only spendable on housing, only redeemable inside a specific province, has no implementation layer in 1936. The substrate was too dumb.

The unit migrated. The substrate caught up.

The token meters at the granularity of one good thought — per-token billing already exists, has existed since the GPT-3 API, and is now the standard unit of every commercial inference contract. It is non-uniform: a token from an instruction-tuned frontier model is structurally a different good than a token from a small open-weights model, and the market already prices the difference. It runs on a programmable substrate: every commercial inference call already passes through APIs, identity layers, billing rails, and increasingly cryptographic attestation.

David Graeber's anthropology, plus the deeper Babylonian and Sumerian record, plus a generation of monetary historians from Adam Tooze to Stephanie Kelton, agree on one structural point: money was never invented as a substitute for barter. The Smithian myth that one day a tribesman with two extra goats traded them for a third tribesman's bushel of grain, and from that moment on the species reasoned its way to coinage — that myth is wrong. Money was invented as credit: an abstract claim on future production, denominated in whatever was the binding constraint of the local economy. Babylonian temples kept barley credit. Egyptian kingdoms kept grain credit. Spanish empires kept silver credit. Industrial Britain kept naval-power credit (sterling). Postwar America kept oil-and-Treasury credit (the dollar).

Each historical reserve currency is, at heart, a wager about what the next century's binding constraint will be. The world now runs on inference. Inference is the consumption of tokens. The reserve currency of an inference economy is therefore, by elementary logic, a denomination of tokens.

This is not an analogy. This is the mechanics of monetary substrates working the way they have always worked.

Reread the five pillars as the organs of a monetary system.

Solodkiy's two corpora — The NVIDIA Innovator's Dilemma and DRAM.GOLD — fit together cleanly once you see the denomination problem. DRAM is the new gold, because the binding physical constraint on token production has migrated from compute to memory bandwidth. The token is the new dollar, because the unit of account in which the AI economy clears is the token.

Reread the five pillars not as five Christensen disruptions but as the architecture of a monetary system, and they snap into a different shape.

This is not a metaphor. This is what the five pillars are.

NVIDIA does not mint tokens. NVIDIA sells the press.

Solodkiy's argument is that NVIDIA is the dominant supplier in the existing value network and is structurally incapable of pursuing the new value networks emerging at the edge. That is correct. The upgrade is this: the new value networks are not five separate markets. They are the components of the next monetary system. NVIDIA selling chips into them is the equivalent of De La Rue printing currency for fifty central banks; profitable, yes, but not the same business as the central bank.

This is structurally weaker than it sounds, because over time, in every industrial economy, the sovereign capacity to issue currency migrates down the value chain, towards whoever owns the demand for the unit. The hyperscalers and frontier labs sit on top of that demand. Every new ASIC programme — Google's Ironwood TPU, AWS Trainium 3, Microsoft's Maia 200, Meta's MTIA, OpenAI's Broadcom-fabricated XPU — is the announcement of a new mint.

The historical analogue is not Cisco in the year 2000. The historical analogue is the United States in 1944. It is what happens to whoever is sitting in the wrong chair when a reserve currency is renominated.

Whatever the equivalent of a Federal Reserve — programmable, multilateral, jurisdictionally aware, identity-bound at the wallet layer and identity-blind at the bearer layer — turns out to be in the AI economy, somebody is going to occupy that chair. The question is whether NVIDIA tries to be that, or watches it be built by someone else who will then buy NVIDIA's silicon at a higher gross margin than NVIDIA charges today, but on the buyer's terms.

NVIDIA's competitive analogues for the chip business are AMD, Broadcom, the hyperscaler ASIC programmes. NVIDIA's competitive analogues for the mint business are categorically different: every model lab that issues tokens, every sovereign that runs its own stack, every metastate that issues identity-bound credentials. The competitive set is wider, the moat is different (network effects of issuance, not silicon performance), and the regulatory exposure is enormous.

But the prescription is the same one Christensen wrote in 1997 and Solodkiy quotes correctly across three chapters: spin it out, fund it generously, give it explicit permission to ignore the parent's metrics, and accept that the most valuable thing NVIDIA does in the next fifteen years may be the thing that does not look like a chip business at all.

It would be a misreading to take the DNA-computing chapter as a flourish.

Christensen's frame, applied honestly, predicts that every substrate transition is its own dilemma. Silicon disrupted vacuum tubes. CMOS disrupted bipolar. Integrated graphics disrupted discrete. CUDA disrupted standalone graphics. Biology will disrupt silicon. Professor Ross King's Adam and Eve robot scientists are the early proof of concept for a substrate transition that runs on a different physics, with a different cost curve, on a different timeline.

When that substrate transition arrives — the realistic timeline is 2030–2040 for early industrial application, 2040–2055 for material substitution at the high end, and 2050–2070 for general-purpose biological computing displacing silicon at scale — the token will renominate again. A "biotoken" will not be denominated in the cost of one good thought delivered on time on a TSMC 3nm process. It will be denominated in the cost of one good thought delivered on time on a self-replicating biological substrate.

This is Haldeman's revenge. Not because the technocrats were right about the kilowatt-hour, but because they were right about the structural impulse: tie money to a measurable, physical, productive unit, and let the unit migrate as the physics migrates. They wrote it on the wrong unit. We are about to write it on the right one. And then, when biology takes over, on the next right one.

The deep continuity is not the unit. The deep continuity is the impulse to denominate value in something a civilisation can actually measure.

The technocrats lost. Their unit was wrong. Their substrate was wrong.
Their politics was eccentric. Their leader was a possibly-fake engineer in a grey-green uniform.

And then, almost a century later, the great-grandson of the man punished by a fiat-currency court for advocating energy-as-money built the most expensive kilowatt-hour-consuming machine in the history of human industry, and sold the output of that machine — denominated in tokens, settled on a programmable substrate, identity-bound at the wallet, identity-blind at the bearer, expiry-bound by smart contract, jurisdictionally aware by design — to a civilisation that had finally learned how to measure thought.

The next reserve currency of the world is not going to be the dollar, nor a basket of central bank digital currencies, nor a return to gold, nor Bitcoin, nor any of the candidates the conventional debate has lined up. The next reserve currency is going to be a denomination of useful thought, delivered on time. The token. Whoever issues it owns the next century the way the dollar owned the last one.

Christensen's dilemma is, in this frame, the smaller of the two stories. The bigger story is that the monetary system is being renominated underneath the dilemma, and that the company that solves Christensen will have, as its reward, the chance to compete for the mint.

The cost of one good thought is the only honest unit of account a civilisation that runs on inference can have. The technocrats were right that money should be tied to a measurable, physical, productive unit. They were a hundred years early because the unit they could measure was not fine-grained enough, the substrate they had to issue on was not programmable, and the civilisation they were addressing was not yet computing all of its valuable work.