This text is an opinion piece. My opinion piece. I am deliberately representing my perspective — the perspective of a user, a citizen, and also someone who pays. I have no intention of hiding the ideological coloring of individual arguments; I want to name them openly as such.
Prologue: Two truths that should not be confused
The public debate about artificial intelligence has the same problem as almost every major societal discussion of our time: it is conducted in extremes. Either AI is the salvation of humanity or its apocalypse. You are for it or against it. But this binary logic prevents the distinction that, in my view, actually matters.
Because there is not just one AI question. There are at least two.
First: As a technological stage of human development, AI is a tool of epochal significance. This development is real, it is transformative, and — as I see it — fundamentally good.
Second: It requires an infrastructure to support this technology. The economic structures, incentives, and narratives behind it are what this text aims to address.
Anyone expecting blanket AI skepticism will not find it here. Anyone willing to separate the technology from the infrastructure will find my contribution to a debate that, in my view, urgently needs to be had.
How progress happens
Let us take a brief excursion into the past. To a time when our ancestors, as early humans, roamed the steppe. It was no god who descended from the sky and handed them fire. And it was no investor who financed the invention of the first spearhead. What happened was simpler and at the same time more profound: someone from within the community observed that certain stones break more sharply than others. This insight was passed on. It was adapted and improved. No patent was needed for this, and no licensing model.
This mechanism — societal foundation enables individual insight, and individual insight flows back into society — is the driving force of human progress. It applies to the use of fire, the invention of the wheel, writing, the steam engine, the electric motor, and the internet. And it applies to the development of artificial intelligence as well.
Albert Einstein was working as a postal clerk when he developed the theory of relativity. He did not stand on the shoulders of a private investor; he stood on the shoulders of Leibniz, Newton, and Maxwell — people who had made their knowledge public, because that is how science works.
The fathers of modern AI — Hinton, LeCun, Bengio, the authors of the Transformer paper — grew up in this spirit. Their insights emerged at universities, funded by public money, published in open papers, available to all. It is no coincidence that many of the most important AI developers have left the major commercial platforms. Not because of poor pay, but because the direction seemed wrong to them. Jan Leike, a long-time safety researcher at OpenAI, wrote upon leaving the company in 2024: "Safety culture and processes have taken a backseat to shiny products." In the same year, eleven current and former OpenAI employees, together with colleagues from Google DeepMind, published an open letter warning of uncontrolled AI development and the systematic silencing of critics. [The Guardian]
AI is, at its core, a societal achievement. Jensen Huang, the CEO of Nvidia, did not invent it. He organized, capitalized, and scaled it. This ability is also valuable. But it is not creation. It is the administration of an inheritance. [Fortune]
The gatekeeper mechanism: a pattern I recognize
What is happening now is something we have witnessed as a society again and again. It is not a conspiracy — it is structure.
Between societal insight and its mass application, organizational actors and companies insert themselves. Sometimes they are publishers. Sometimes corporations. Sometimes financial institutions. They call themselves part of the infrastructure. They are — but of an infrastructure with its own interests.
The British mathematician Clive Humby coined the phrase in 2006: "Data is the new oil." He meant something sober by it: data, like crude oil, must first be refined before it yields value. The analogy caught on. What then sounded like a technical observation now takes on a second, more uncomfortable layer of meaning. The same applies to tokens.
The oil parallel: a century in the rearview mirror
The automotive industry told the 20th century that progress needs four wheels and a lot of asphalt. Cities were rebuilt. Green spaces were paved over. Public transit was systematically weakened because it threatened the car-sales model — the engine of industrial development of our time.
The argument was always the same: growth, productivity, freedom. And the argument was even correct — in part. Countries that embraced the technological leap of cars and roads and rebuilt their infrastructure accordingly saw significant gains in economic productivity. Countries without developed infrastructure drifted into economic second-tier status over the course of the 20th century. Mobility was real, measurable progress.
But the business model of the gatekeepers was always the same: the more fuel is consumed, the more we earn. Efficiency was the enemy. Anyone who built a more fuel-efficient car threatened the gas station. Anyone who bet on rail threatened the automaker. The oil industry knew for decades what combustion engines were doing to the climate — internal documents prove it. Yet investment went into lobbying, narratives, and delaying regulation. The bill was paid by others.
Now, for the first time, we find ourselves in a situation where we can generate energy cleanly. And the old gatekeepers of the fossil industry are doing everything to slow this transition — not because it is technically impossible, but because it threatens or even ends their business model.
Scene change: same structure, new industry.
"Tokens are Value" — fuel for the 21st century?
On March 16, 2026, Jensen Huang put forward a thesis at Nvidia GTC in San Jose that has since been circulating through tech blogs and podcast transcripts: "Tokens are Value." An engineer with a $500,000 annual salary who does not spend at least $250,000 of it as a token budget is, he said, suspicious. He said he would be "deeply alarmed" by such employees. [Business Insider]
This sounds like a productivity recommendation. It is demand stimulation.
Huang repeated this thesis at the start of June 2026 at Computex in Taipei — by some observers' count, perhaps forty times: agentic AI is here, it works, it makes money. And every token produced is a unit of revenue. His equation: compute, tokens, intelligence, economic output. In this world, data centers are no longer server farms — they are "AI Factories." And their primary unit of production is the token. [Semiconalpha]
That is a valid KPI — for data center operators. For no one else.
It would be like recommending that a sales department evaluate its reps by fuel consumed. Whoever drives the most gets the best bonus, regardless of whether they close deals or simply burn gas. The colleague who drives fewer miles but works more deliberately and closes better deals is rated worse by this metric. Huang's token metric is exactly that: throughput instead of outcome. Miles instead of deals. Journey instead of destination.
The price lie — why cheaper becomes more expensive
The industry presents us with a nice story in its marketing: token prices for GPT-4-equivalent capability have fallen from $20 to $0.40 per million tokens since 2022. A factor of 50. Democratization through price decline. [Token Cost Analysis]
What this story omits: agentic AI — that is, AI systems that independently plan tasks, call tools, and execute multiple steps — consumes up to 1,000 times more tokens than a simple chat prompt, according to a Stanford University study from April 2026. Greater token consumption does not necessarily lead to better results. Quality reaches a plateau, but costs keep rising. [Stanford / arxiv]
The equation:
| Factor | Value |
|---|---|
| Price per token | −50× |
| Consumption per task | +1,000× |
| Net | Costs ×20 |
This is not a failure of the technology. This is the business model. Tokens get cheaper so that more are consumed, so that more GPUs are needed. Huang, as CEO of a chip manufacturer, calls this agentic AI as a growth engine. One could also call it structural cost explosion with a marketing loop. The result is already being felt across the technology world: whoever needs to buy more RAM or storage pays the price, and the chip manufacturers rub their hands.
The more fuel your car consumes, the more they earn. The principle applied to oil producers, and it is the same in the digital revolution. Only now the fuel is called tokens.
Who actually pays, and for what?
The complete value chain is unrelentingly transparent:
Nvidia (GPUs) → AWS, Azure, Google, CoreWeave (data centers) → OpenAI, Anthropic (AI products) → businesses and users (pay)
Every tier earns its margin. But at the end there is always a real payer.
If it succeeds in shaping public discourse so that token consumption counts as proof of productivity — for the engineer, the doctor, the teacher, the social worker — then this business model is secured for decades. It is the structural logic we know from other industries: define the metric before regulation strikes. Transfer the system costs to the consumer, but keep the system profits for yourself.
Token consumption is the new personal carbon footprint of the AI industry.
Then there is the energy question — and here the oil analogy closes in a bitter way. Data centers are expected to consume around 1,050 TWh of electricity per year by the end of 2026, enough to rank among the top 20 energy consumers in the world if they were a standalone country. A complex reasoning prompt already consumes over 33 Wh — roughly 70 times more than a simple search query. [Brookings Institution]
This has a practical implication that sounds like a bad joke but is not: in Berlin-Spandau, the waste heat from an NTT DATA data center will heat a 31-hectare new development with 4,500 apartments for over 10,000 people starting at the end of 2026. In Frankfurt, the Telehouse data center in the Gallus district supplies at least 60% of the heating for around 1,300 new apartments. We are heating our homes with the waste heat of token machines. More asphalt for the cars. [NTT DATA Berlin] [Mainova Frankfurt]
What the technology could actually do
The bitter part: the solution to growing token consumption already exists technically. Intelligent model routing — the automatic decision about which model is needed for which task — can save 60 to 80 percent of token costs at consistent quality. Simple request? Small model. Genuinely complex problem? Then, and only then, the large model. The BEST-Route framework from ETH Zurich shows: 60% cost reduction with less than 1% quality loss. [Reddit / Claude Code Routing]
The idea goes further: AI systems that do not reflexively open up enormous context windows, but instead work with compact, precise contexts and intelligent summarization. Systems that are frugal — not because the user forces it, but because efficiency is built into the architecture as a goal. Fewer tokens per result, not more tokens per unit of revenue.
The technology for this exists. What is missing is the commercial interest of the major providers, because their business model is based on volume, not efficiency. The auto industry suppressed more efficient engines and models for decades. We know how that story ends, and we are paying the price.
Resistance in the stack
But resistance does not have to be loud. It can be technological. Whoever runs local models — on their own hardware, with open-weight models like Llama, Qwen, or Mistral, or in shared data centers designed for token efficiency — structurally opts out of the token economy model. Tokens are produced, yes. But they are not sold. Consumption costs electricity, and with that the only societally meaningful metric becomes relevant again: result per watt. Result per token.
This is certainly not the solution for every use case. But there is a middle path that is more important to society than any chip produced: technological literacy. The knowledge that "Token = Value" is a seller's metric. The understanding that an AI system that spins up an agentic workflow for a simple request is not more productive — it is simply wasteful and expensive. That efficiency does not mean increasing volume, but achieving goals with minimal effort.
And for companies, the decisive consequence: do not ask how many tokens your employees consume. Ask what they achieve with them. Measure the outcome. Not the miles.
Epilogue
AI technology — in its origins, in its underlying logic, in its potential — belongs to society. It grew out of society, as all technologies before it did.
The question of who controls and prices the infrastructure of this technology is a political question. It should be treated as such. Just as we eventually understood as a society that electricity, water, and roads are not private acts of grace but basic public provision.
The phrase "data is the new oil" comes from the British mathematician Clive Humby, coined in 2006 as a sober observation about raw materials that need refining before they yield value. What he did not anticipate: that his analogy would one day be overtaken by oil's dark side as well. Oil advanced the world. And it drew human society into a dependency from which we have not been able to free ourselves to this day. Tokens are the new fuel. And we know how that story ends. [Wikipedia: Clive Humby]
Jensen Huang has already answered the question of who owns the AI infrastructure. His answer is: Nvidia.
The more important question is: what is our answer?