Author: Luo Yihang
In January 2009, an anonymous individual invented something called a "token." You contribute computing power to obtain tokens, which circulate, are priced, and traded within a consensus network. The entire crypto economy was born from this. More than a decade later, people are still debating whether these tokens actually have any value.
In March 2025, a man in a leather jacket redefined something called a token. You invest computing power, produce tokens, and those tokens are immediately consumed during an AI inference and reasoning process: thinking, reasoning, writing code, making decisions. The entire AI economy is thus accelerated. No one argues whether these tokens have any value, because you just used millions this morning.
Two tokens, same name, same underlying structure: computing power goes in, and valuable things come out.
In March 2026, I sat in the NVIDIA GTC venue and listened to Jensen Huang's keynote speech, which was almost entirely devoid of product promotion. Yes, he announced Vera Rubin, a product combining a CPU and a GPU. But this time, he didn't talk about chip specifications or manufacturing processes; he spoke about a complete economics of token production, pricing, and consumption:
Which model corresponds to which token speed? Which token speed corresponds to which pricing range? And which pricing range requires what level of hardware to support it?
He even helped the CEOs and decision-makers in the audience who controlled the company's checkbook to prepare a data center computing power allocation plan: 25% for the free tier, 25% for the mid-range tier, 25% for the high-end tier, and 25% for the high-premium tier.
Yes, this time he didn't specifically sell any particular GPU series, just like he did with Blackwell two years ago. But this time, he's selling something bigger. After two hours, I think the one thing he really wanted to say was: "Welcome to consume tokens, and only Nvidia's factory could produce them."
At that moment, I realized that this man, and the anonymous person who mined the first token 17 years ago, were doing the exact same thing in terms of structure.
The same set of conversion rules
The anonymous author, using the pseudonym "Satoshi Nakamoto," wrote a nine-page white paper in 2008, outlining a set of rules: invest computing power, complete a mathematical proof (Proof of Work), and receive crypto tokens as a reward.
The brilliance of this rule lies in the fact that it doesn't require anyone to trust anyone else—as long as you accept these rules, you automatically become a participant in this economy. This rule is correct, after all, it brings so many treacherous people together.
On the stage of GTC 2026, Jensen Huang did something structurally identical.
He presented a graph illustrating the relationship and tension between inference efficiency and token consumption: the Y-axis represents throughput (how many tokens are generated per megawatt of power consumption), and the X-axis represents interactivity (the speed of tokens perceived by each user). He then labeled five pricing tiers below the X-axis: Free uses Qwen 3, $0/million tokens; Medium uses Kimi K2.5, $3/million tokens; High uses GPT MoE, $6/million tokens; Premium uses GPT MoE 400K context, $45/million tokens; and Ultra, $150/million tokens.
This image could almost serve as the cover of Jensen Huang's white paper on "token economics".
Satoshi Nakamoto defined "what constitutes valuable computation"—achieving a SHA-256 hash collision is valuable. Jensen Huang defined "what constitutes valuable reasoning"—generating tokens at a specific speed for a specific scenario, given power consumption constraints, is valuable.
Neither Satoshi Nakamoto nor Jensen Huang directly produced tokens; they defined the token production rules and pricing mechanisms.
One sentence that Lao Huang said on stage could almost be directly written into the summary of the token economics white paper—
Tokens are the new commodity, and like all commodities, once it reaches an inflection, once it becomes mature, it will segment into different parts.
Tokens are the new commodities. Commodities will naturally stratify as they mature. He's not describing the current situation; he's predicting a market structure and then precisely positioning his hardware product line across each layer of that structure.
The production processes of the two types of tokens even have a semantic symmetry: mining is called mining, and reasoning is called inference.
The essence of both cryptocurrency mining and inference is turning electricity into money. Miners spend electricity to mine crypto tokens and then sell them; inference models and AI agents spend electricity to generate AI tokens and then sell them to developers for millions. The intermediate links are different, but the two ends are the same: on the left is the electricity meter, and on the right is the revenue.
Two ways to write scarcity
Satoshi Nakamoto's most important design decision wasn't the Proof of Work, but the cap of 21 million Bitcoins. He created artificial scarcity through code—no matter how many mining machines flood in, the total number of Bitcoins will never exceed 21 million. This scarcity is the value anchor of the entire crypto economy.
Huang Renxun, however, used the laws of physics to create natural scarcity. He said:
"You still have to build a gigawatt data center. You still have to build a gigawatt factory, and that one gigawatt factory for 15 years amortized... is about $40 billion even when you put nothing on it. It's $40 billion. You better make for darn sure you put the best computer system on that thing so that you can have the best token cost."
A 1GW data center will never become 2GW. This isn't a code limitation; it's a law of physics.
Land, electricity, heat dissipation—each has physical limits. How many tokens this factory you built at a cost of $40 billion can produce over its 15-year lifespan depends entirely on the computing architecture you put inside.
Satoshi Nakamoto's scarcity can be forked. If you don't like the 21 million coin cap, fork a new chain, change it to 200 million coins, call it Ethereum or whatever you want, and release a white paper while you're at it. And people have actually done this, and they enjoy it immensely.
The scarcity created by Lao Huang cannot be forked. After all, you can't fork the second law of thermodynamics, you can't fork the power grid capacity of a city, and you can't fork the physical area of a piece of land.
But whether it's Satoshi Nakamoto or Jensen Huang, the scarcity they created led to the same result: a hardware arms race.
The history of cryptocurrency mining is: CPU → GPU → FPGA → ASIC. Each generation of dedicated hardware renders the previous generation obsolete. The history of AI training and inference is repeating itself: Hopper → Blackwell → Vera Rubin → Groq LPU. General-purpose hardware started, dedicated hardware became the standard. Nvidia showcased the Groq LPU at this year's GTC, a deterministic dataflow processor released after acquiring Groq. Static compilation, compiler scheduling, no dynamic scheduling, 500MB of on-chip SRAM—its architectural philosophy is that of an ASIC for inference. It does only one thing, but does it to the extreme.
Interestingly, GPUs played a key role in both waves.
Around 2013, miners discovered that GPUs were better suited than CPUs for mining crypto tokens, leading to a sell-out of Nvidia graphics cards. Ten years later, researchers found that GPUs were the best tool for training and inference AI models, again causing Nvidia data center cards to be sold out. As a type of processor, GPUs have served two generations of the token economy.
The difference lies in the fact that the first time, Nvidia passively benefited, and that was the end of it. The second time, however, when the main battleground for AI computing power consumption shifted from pre-training to inference, Nvidia quickly seized the opportunity to proactively design the entire game, becoming the writer of the AI game rules.
The world's most profitable shovel
The most profitable companies during the gold rush weren't the prospectors, but Levi Strauss, who sold shovels. Similarly, the most profitable companies during the cryptocurrency mining boom weren't the miners, but Bitmain and Jihan Wu, who sold mining rigs. And in the AI pre-training and inference wave, the most profitable companies weren't the base models and agents, but Nvidia, which sells GPUs.
But to be honest, Bitmain and Nvidia are no longer the same companies in their respective industries.
Bitmain only sells mining machines; Nvidia was once a supplier to Bitmain. What cryptocurrency you mine, which mining pool you use, and at what price you sell it are all irrelevant to Bitmain. It's a pure hardware supplier, earning one-time profits from the equipment.
Nvidia is different. It doesn't just sell hardware. Now, especially since the explosion of inference-side AI in 2025, it has deeply defined what to mine with this GPU, how to price tokens, who to sell tokens to, and how data centers should allocate computing power... All of these are in Nvidia's presentation slides: he divides the market into five tiers, each tier corresponding to what model, context length, interaction speed, and price... Nvidia has standardized and formatted the future market where AI inference drives everything.
Around 2018, global computing power was concentrated in a few large mining pools—F2Pool, Antpool, and BTC.com—which competed for computing power share, but the source of mining machines was highly concentrated in Bitmain.
Just like Nvidia today, 60% of its revenue comes from competing "hyperscalers" such as AWS, Azure, GCP, Oracle, and CoreWeave, while 40% comes from decentralized AI natives, sovereign AI projects, and enterprise customers. Large "mining pools" contribute the majority of revenue, while small "miners" provide resilience and diversification.
The two ecosystems have identical structures. However, Bitmain later encountered competitors—Whatsminer, Innosilicon, and Canaan Creative were all eroding its market share. Mining machines are relatively simple ASIC designs, giving pursuers a chance. But shaking Nvidia's dominance seems increasingly difficult: 20 years of CUDA ecosystem, hundreds of millions of installed GPUs, NVLink sixth-generation interconnect technology, and the decoupled inference architecture integrated with Groq—Nvidia's technological complexity and ecosystem barriers render most competing tools ineffective.
This could take 20 years.
The fundamental fork of the two tokens
What fundamentally distinguishes cryptocurrencies from AI training and inference tokens is people's motivations and psychology in using them.
The demand side of crypto tokens is speculation. Nobody "needs" Bitcoin to get their job done. All white papers claiming blockchain tokens can solve your problems are scams. You hold crypto because you believe someone will buy it from you at a higher price in the future. Bitcoin's value comes from a self-fulfilling prophecy: it has value if enough people believe it does. This is a faith economy.
The demand side for AI tokens is productivity. Nestlé needs tokens to make supply chain decisions—its supply chain data refresh frequency has decreased from 15 minutes to 3 minutes, reducing costs by 83%, a value that can be directly mapped to P&L (Production and Lending). Nvidia's engineers already require tokens to write code instead of manually; their research teams need tokens for scientific research. You don't need to believe that tokens have value; you just need to use them, and their value will be proven through use.
This is the most fundamental difference between the two types of tokens. Crypto tokens are produced to be held and traded—their value lies in not using them. AI tokens are produced to be consumed immediately—their value lies in the moment they are used.
One is digital gold, which becomes more valuable the longer you hoard it; the other is digital electricity, which is burned as soon as it is produced.
This difference determines that the AI token economy will not experience the same bubble as the crypto token economy. Bitcoin's dramatic rise and fall stems from the fact that the price of speculative commodities is driven by sentiment. However, token prices are driven by usage and production costs. As long as AI remains useful—as long as people continue to use Claude Code to write code, ChatGPT to write reports, and agents to run business processes—the demand for tokens will not collapse. It doesn't rely on faith; it relies on indispensability.
In 2008, the Bitcoin white paper required repeated justification for the value of a decentralized electronic cash system. Seventeen years later, the debate continues.
In 2026, token economics sparked no controversy; it became a consensus without even needing proof. When Nvidia CEO Jensen Huang stood on the GTC stage and said, "Tokens are the new commodity," no one questioned it. Because everyone in the audience had spent millions of tokens that morning using Claude Code or ChatGPT. They didn't need to be convinced that tokens had value—their credit card statements already proved it.
In this sense, Huang is truly a copy of Satoshi Nakamoto, the one who left behind a monopoly on mining machine production, defined the use cases and usage standards for tokens, and held an annual show at the SAP Center in San Jose to show people how powerful the next generation of "mining machines" supporting AI training and inference are.
Satoshi Nakamoto possessed a captivating, deliberate charm. He designed the rules, handed them over to the code, and then vanished. This was the romance of the cypherpunks. Huang, on the other hand, was more like a businessman than any scientist. He designed the rules, maintained them personally, and continuously added to them, building a fortified moat around his business.
The token you saw in the past because you believed, you can now see without believing. It is the next after Watt, Ampere, and Bit.
