Tokens Are the New Salt and Iron

200 Milliseconds Around the Earth

A SWIFT cross-border wire transfer takes one to five business days. A token API call, from prompt submission on one side of the planet to completion arriving on the other, takes roughly 200 milliseconds to 2 seconds. The difference is five to six orders of magnitude.

That gap is not a technical curiosity. It is a structural discontinuity. Every time you send a prompt to GPT-5.5 or Claude or DeepSeek, you are participating in a transaction that crosses borders, converts currency into intelligence, and delivers productive output, all before your finger lifts off the Enter key. No customs officer inspects it. No central bank clears it. No foreign exchange desk touches it.

The unit of this transaction is the token. And it is time we understood what that actually means.

Salt, Iron, Oil, Fiat: The Genealogy of Strategic Resources

Every era has a resource so foundational that controlling it means controlling the economy. Control the resource, and you control the tax base, the military capability, and the distribution of power.

In the Han Dynasty, it was salt and iron. Emperor Wu, on the advice of his finance minister Sang Hongyang, nationalized both industries around 119 BCE. The rationale was straightforward: salt was essential for food preservation, iron for agriculture and warfare. Private merchants had been profiteering; the state needed revenue for military campaigns against the Xiongnu. The resulting monopoly funded an empire and provoked a debate, the Discourses on Salt and Iron of 81 BCE, that echoed through two millennia of Chinese economic thought.

In the twentieth century, it was oil. The discovery that petroleum powered not just lamps but battleships, tanks, and aircraft transformed it from a commodity into a strategic asset. Control over oil supply shaped World War II, created OPEC, anchored the petrodollar system, and drove decades of geopolitical maneuvering from the Middle East to the South China Sea.

In the post-Bretton Woods era, it was fiat currency itself. The ability to issue a reserve currency gave the United States what Valery Giscard d'Estaing called an "exorbitant privilege": the capacity to print money that the rest of the world had to hold. Monetary sovereignty became the ultimate form of strategic resource control.

Tokens are all three of these at once, with one additional property that none of them possessed: they move at the speed of light.

Each generation increases the abstraction level of control. Salt and iron were physical goods, seized at mines and workshops. Oil was a physical commodity but traded through financial instruments. Fiat currency was pure abstraction backed by state power. Tokens are abstraction backed by algorithmic capability: the "credit" of a token comes not from a government's balance sheet but from the quality of the model that produces it. GPT-5.5 tokens cost more than GPT-4 tokens because the model is more capable, not because a central authority decreed it.

The Escape Velocity of Tokens

Here is where tokens diverge from every strategic resource that preceded them: they can bypass the control mechanisms that governed salt, oil, and money.

Consider an American developer who wants to use DeepSeek, the Chinese AI lab whose models rival the best from OpenAI and Anthropic. DeepSeek's API is priced in renminbi and hosted on Chinese infrastructure. The developer cannot directly pay in dollars. Instead, they purchase access through an API relay service, a third-party proxy that accepts dollars, routes the request through its own infrastructure, and pays DeepSeek in renminbi on the backend.

From the developer's perspective, they spent dollars and received intelligence. From DeepSeek's perspective, they earned renminbi from a domestic partner. No dollars entered China's foreign exchange system. No renminbi left it. Yet economic value crossed a geopolitical boundary in 200 milliseconds.

This is not hypothetical. In 2026, Chinese developers use "shadow APIs" to access US models like Claude and GPT-5.5 through proxy servers. American developers use relay services to access Chinese models. The pattern is identical. And the structure of these relay services is not new. It is, functionally, the same innovation that Shanxi bankers (票号, piao Hao) pioneered in Qing Dynasty China: you deposit silver in Beijing, receive a paper note, and redeem it in Taiyuan. The silver never moved. The value did.

That last line is the key insight. The institutional infrastructure for governing token flows has not been built. It will be, because every strategic resource in history eventually acquired a governance layer. Salt got state monopolies. Oil got OPEC. Money got central banks. Tokens will get something analogous. The question is what shape it takes and who builds it first.

The Token Standard: A Bipolar World

As of 2026, meaningful frontier model production is concentrated in two countries. The United States has OpenAI, Anthropic, Google, Meta, and xAI. China has DeepSeek, Alibaba (Qwen), ByteDance, and the state-backed ecosystem. No other country or region has both the model training capability and the compute infrastructure to produce frontier-grade tokens at scale.

This is not merely a technology gap. It is a structural feature of the token economy that mirrors the geopolitical structure of the early 21st century.

The token production chain looks like this:

Chip Design → Fabrication → Data Center → Model Training → API Service → End User
  (NVIDIA/     (TSMC)       (Hyperscalers)  (OpenAI,       (Pricing)    (Consumption)
   Huawei)                                   DeepSeek, etc)

The United States controls chip design (NVIDIA, AMD) and leverages TSMC in Taiwan as a fabrication partner. China is building domestic alternatives (Huawei Ascend, SMIC) while navigating US export controls that restrict access to the most advanced fabrication processes. As of January 2026, the US Commerce Department's Bureau of Industry and Security revised its rules on H200 and MI325X exports to China, shifting from "presumption of denial" to conditional case-by-case licensing with stringent oversight.

But export controls on chips are not export controls on tokens. A model trained on NVIDIA H100s in Oregon can serve tokens to any IP address on earth. A model trained on Huawei Ascend 910Bs in Shenzhen can do the same. The strategic resource is not the chip, it is the token that the chip helps produce. And tokens, as we have seen, escape the physical constraints that make chip export controls enforceable.

If we think in terms of a "token standard", analogous to the gold standard or the dollar standard, the implications are significant.

Token producers are monetary authorities. When OpenAI prices GPT-5.5 at $15 per million input tokens and then drops it to $5, every business on earth that uses GPT-5.5 experiences an instantaneous cost reduction. This is functionally equivalent to a central bank cutting interest rates. Global "production costs" denominated in tokens fall. But unlike the Fed's rate decisions, which require FOMC meetings, congressional oversight, and months of forward guidance, OpenAI's pricing change requires only a blog post. One company, one decision, global impact.

Token prices are falling faster than Moore's Law. Between 2024 and 2026, the cost of equivalent intelligence dropped 90-97%. DeepSeek V4 charges $0.30/$0.50 per million tokens (input/output), compared to Claude Sonnet 4.6 at $3/$5. The cheapest frontier model is now 42 times less expensive than the priciest flagship for the same volume. This deflationary pressure is a form of productivity expansion that traditional monetary economics has no framework for. When the cost of intelligence, the most fundamental input to all productive activity, drops by two orders of magnitude in two years, the concept of "inflation" needs rethinking.

Token-importing nations face a new dependency. Every country without frontier model capability is a net token importer. They purchase intelligence from US or Chinese providers, paying in dollars or renminbi. The capital outflow is real, but it does not show up in traditional trade statistics because tokens are classified as digital services, not commodities. A nation that imports 90% of its oil knows exactly how vulnerable it is. A nation that imports 90% of its AI tokens may not even recognize the dependency.

The Governance Paradox

Here is the central tension of the token era, and it is a real one, not a logical inconsistency.

Tokens must be governed. Any resource this strategically important will eventually attract state regulation, because states cannot afford to leave uncontested the control of something that determines military capability, economic productivity, and information sovereignty. China's requirement for model registrations and content filtering is one form of this. The EU's AI Act is another. The US approach, focused on chip export controls and data privacy, is a third.

But tokens resist governance. Their escape velocity, the ability to route through proxy services and bypass both geographical and financial boundaries, makes them inherently harder to regulate than physical goods or even traditional financial flows. You cannot put a customs checkpoint on an HTTPS request. You cannot freeze a token wallet the way you freeze a bank account, because there is no wallet: there is only a stateless API call that exists for 200 milliseconds.

This creates a governance paradox: the more important tokens become, the harder they are to control; the harder they are to control, the more important they become to control.

The historical parallel is instructive. The salt and iron monopolies of the Han Dynasty were not implemented overnight. They were the result of decades of debate between Legalist technocrats (who argued for state control) and Confucian reformists (who argued for free markets). The Discourses on Salt and Iron of 81 BCE was the recorded proceeding of that debate. It took generations to settle on a governance model that balanced state revenue with market efficiency.

Tokens are in the early stages of this same cycle. The current regulatory approaches, chip export controls, model registrations, content filtering, are what you get when you apply old tools to a new problem. They target the physical substrate (chips) or the output (generated content) rather than the resource itself (tokens). It is as if the Han Dynasty tried to regulate salt by controlling who could own cooking pots.

The governance anchor for tokens will most likely land at the API service and relay layer, not at the chip or model layer. This is where token flows become observable and regulable. The historical precedent is the evolution from Shanxi banks to modern commercial banks: you do not ban the flow, you license the intermediaries. A future "token settlement system" might require relay services to hold licenses, maintain reserve compute capacity, and report cross-border token flows, analogous to how modern banks report cross-border capital flows through systems like SWIFT.

But that system does not exist yet. And until it does, tokens will continue to flow across borders at the speed of light, carrying economic value and productive intelligence, invisible to the institutions designed to govern the movement of goods and money.

If Sang Hongyang Were Alive Today

Sang Hongyang, the architect of the salt and iron monopolies, understood something fundamental: whoever controls the strategic resource of an era controls the distribution of power within it. He was not making a moral argument. He was making a structural one.

If he were alive today, he would recognize the pattern immediately. A resource that is universally needed, concentrated in production, and essential for both economic output and military capability. A small number of producers who can set prices and thereby influence the cost structure of every other economic activity. A population that depends on the resource but has no visibility into, or control over, its production and distribution.

He might also recognize the governance challenge. The salt and iron monopolies were enforceable because salt mines and iron foundries were physical assets that could be seized. Tokens have no physical form. The production facility (a data center) can be located anywhere. The distribution mechanism (an API) can be proxied and relayed. The consumer (anyone with an internet connection) can access it from any jurisdiction.

The problem has not changed. The constraints of the solution space have shifted entirely. And the institutions that will eventually govern token production and distribution have not yet been invented.

Two thousand years ago, it took a century of debate to build the institutions for governing salt and iron. A century ago, it took decades to build OPEC and the petrodollar system for governing oil. The token era moves faster. The institutions will need to be built in years, not decades, because the resource they govern moves at the speed of light.

The question is not whether token governance will emerge. It is whether it will emerge fast enough to matter.

FAQ

What exactly is a token in the AI context? A token is the fundamental unit of computation in large language models. Roughly speaking, one token corresponds to about three-quarters of an English word. When you send a prompt to an AI model, your input is broken into tokens, processed through the model, and the output is generated token by token. You pay for tokens consumed, typically priced per million tokens.

Why compare tokens to salt, iron, and oil? Salt and iron were the foundational resources of agricultural economies: essential for food preservation and tool-making. Oil was the foundational resource of industrial economies: essential for transportation, manufacturing, and warfare. Tokens are the foundational resource of AI economies: essential for intelligence production, which is increasingly the bottleneck for all other productive activity. In each case, control over the resource translates directly into economic and geopolitical power.

Can't anyone produce tokens with open-source models? Technically yes, which is why the monopoly is softer than salt or oil monopolies. Open-source models like Meta's LLaMA series are available to anyone with sufficient GPU hardware. But producing frontier-grade tokens, the kind that power state-of-the-art reasoning, coding, and multimodal tasks, requires massive compute clusters, specialized talent, and billions in training investment. The gap between "can run a model locally" and "can produce frontier intelligence at scale" is comparable to the gap between "can drill a well" and "can operate a refinery."

How do API relay services work? An API relay service acts as an intermediary between a user in one jurisdiction and an AI provider in another. The user pays the relay service in their local currency. The relay service routes the API request through its own infrastructure and pays the AI provider in the provider's preferred currency. From the user's perspective, they accessed the model. From the provider's perspective, they served a domestic customer. No cross-border financial transaction appears in either country's official records.

What would a "token central bank" look like? The exact form is speculative, but the functional requirements can be inferred from historical precedent. It would need to: license and oversee token intermediaries (relay services, API aggregators); maintain visibility into cross-border token flows; establish reserve requirements (possibly compute reserves rather than financial reserves); and set "token policy" analogous to monetary policy, adjusting the terms of token production and distribution to achieve strategic objectives. Whether this institution is a government agency, an international body, or a new form of public-private hybrid remains to be seen.

References

• Discourses on Salt and Iron (盐铁论), Huan Kuan, 81 BCE. Wikipedia
• Sang Hongyang Biography. ChinaKnowledge
• James Kynge: "AI tokens are the new oil." CGTN, April 2026
• "The 2026 AI Price War Explained." AI Magicx Blog
• "AI API Pricing Dropped 80% in 2026." TokenMix Blog
• "Anthropic finally beat OpenAI in business AI adoption." VentureBeat
• "US Export Controls and China's 'Good Enough' AI Stack." Geopolitical Monitor
• "China devs buy banned US AI despite DeepSeek V4." Cybernews
• Chatham House: "AI export controls are not the best bargaining chip." Chatham House, April 2026
• "Nvidia chip smuggling reveals enforcement limits." Fortune, May 2026
• Ramp AI Index, March 2026. Ramp