This is the part of the AI boom that people keep trying to talk around: it’s not “software eating the world” anymore. It’s electricity and concrete and steel eating your budget.
So when Nvidia’s Jensen Huang goes to the G20 and starts framing AI in “dollars per million tokens” instead of “cost per kilowatt-hour,” I don’t hear a fun nerdy rebrand. I hear a warning shot. The story isn’t just that AI is expensive. It’s that the way we think about cost is shifting from “How much power does this use?” to “How much money can we squeeze out of each unit of output?”
That sounds normal—of course businesses price what they sell. But there’s something a little unsettling about the flip. Once you monetize energy through tokens, you’re not just optimizing machines. You’re shaping what gets created, who gets to create, and how dependent everyone becomes on the people who own the pipes.
Huang also talked about the scale of AI infrastructure in a way that should land with anyone who’s ever rolled their eyes at “just scale it.” He said one gigawatt of AI capacity is roughly double the investment of the traditional leading-edge fabrication plants that used to be treated like the big, scary mega-projects—those $25 billion monuments to industrial ambition. That’s a bold comparison, and it’s doing a specific kind of work: it’s telling governments and big companies, “You’re not funding a feature. You’re funding a new class of national infrastructure.”
Here’s my take: that framing is smart, and it’s also dangerous.
Smart, because it tells the truth about the real bottleneck. AI isn’t limited by clever ideas right now. It’s limited by compute, power, cooling, and the ability to build giant facilities fast. If you’re a country trying to stay competitive, or a company trying not to get crushed by the next wave, this is the part you can’t hand-wave. You need physical capacity, and you need it yesterday.
Dangerous, because once you accept “tokens” as the unit of value, you start turning everything into toll roads. The question becomes: who owns the toll booths?
Imagine you run a small startup and you’ve built a product that relies on AI. Today your costs might feel like “cloud bills.” Tomorrow it’s “token bills,” and pricing is set by whoever controls scarce compute and scarce energy. If they decide to raise prices, or change access rules, or prioritize certain customers, you don’t have many options. You can’t just spin up a competing gigawatt in your garage. Your whole business becomes a passenger on someone else’s infrastructure train.
Or imagine you’re a school system trying to give students AI tools. If tokens are money and money is scarce, you end up rationing intelligence like it’s printer ink. The kids who get more access learn faster. The kids who get less fall behind. We’ve seen this movie with internet access and devices. Tokenized AI could make that gap feel even more personal, because it’s not “you have slower internet,” it’s “you get fewer chances to think with the tool.”
Now flip it: imagine you’re a hospital, or a call center, or a public agency drowning in paperwork. Token pricing could be the cleanest way to budget and measure value. You can track usage, set limits, and tie spending to output. From that view, “dollars per million tokens” is just clarity. It’s a way to make the costs legible.
I get that argument. I just don’t buy that it stays clean.
Because tokens aren’t just “usage.” They’re a proxy for attention, productivity, and decisions. When the meter is running, people change their behavior. Managers start asking teams to cut prompts, shorten outputs, avoid “wasting tokens” on exploration. You don’t just optimize cost—you compress thinking. The incentives push toward faster, cheaper answers, even when the best work is messy and slow.
And at the national level, the gigawatt talk turns AI into an arms race without calling it one. If a gigawatt is the new badge of seriousness, then countries and mega-corps will chase capacity the way they chase ports, pipelines, and chip supply. That can create jobs and new industries, sure. It can also create a world where a handful of players set the terms of intelligence access the way a handful of players once set the terms of energy access.
There’s also a quieter consequence: if AI’s economic model is “monetize energy through tokens,” then AI companies are incentivized to make you use more tokens, not fewer. The ideal customer isn’t the one who gets a quick correct answer and leaves. It’s the one who integrates AI into every workflow, every chat, every document, every decision—always consuming. That doesn’t mean the tools are bad. It means the business model pulls in a direction that doesn’t always match the user’s best interest.
What’s not clear yet—and this matters—is how much of this token pricing becomes a stable “utility” model versus a chaotic pricing battlefield. If the market gets competitive, token costs could drop and access could broaden. If capacity stays tight and demand stays explosive, tokens become the new scarce resource, and scarcity always finds a way to become power.
I’m not against building the infrastructure. I’m against pretending that this is just a technical scaling problem. It’s a governance problem, a pricing problem, and a control problem, wrapped in a shiny “innovation” story.
If “dollars per million tokens” is the new language of AI, who should get to decide what a token is worth?