A man walks into a bar and says: “Give me a dollar and I’ll give you two back.” Sounds like a good deal. You’d probably take it. But then another man walks in and says the same thing — except he’ll give you three dollars back. And a third offers four. Suddenly the first man has a problem. Not because his deal got worse. Because everyone else’s got better.
This is, more or less, what is happening right now in artificial intelligence. The commodity is tokens — the basic unit of computation that AI models consume and produce, roughly corresponding to fragments of words. The economics around them are undergoing one of the most dramatic shifts in the history of technology. And the companies racing to generate them most efficiently are sitting on something that looks increasingly like a licence to print money.
The falling cost, the rising demand
For most commodities, cheaper production and higher demand don’t coexist for long. Markets equilibrate. Margins compress. Someone undercuts someone else until profit evaporates.
Tokens are different — at least for now.
The cost of producing a token has fallen by roughly 99% over the past three years. What cost a dollar to generate in 2023 costs fractions of a cent today, driven by better model architectures, more efficient chips, and the relentless compounding of engineering ingenuity. This trajectory shows no signs of slowing.
At the same time, demand for tokens is not merely growing — it is exploding. Every company that deploys an AI assistant, every developer building an agent, every enterprise automating a workflow, consumes tokens by the billion. As models become more capable, use cases multiply. As use cases multiply, consumption compounds. The market for tokens is not a fixed pie being divided more efficiently. It is a pie that doubles in size every few months.
The result is a widening gap between what tokens cost to produce and what they are worth to the buyer. In economics, that gap has a simple name: profit. And right now, it is enormous.
Reading the signals in the earnings reports
You don’t have to take this on faith. The evidence is written clearly in the quarterly reports of every major technology company.
Microsoft, Google, Amazon, and Meta have all committed to capital expenditure programmes that, taken together, run into hundreds of billions of dollars over the next few years. These are not acts of faith or vanity. These are companies with sophisticated finance teams and activist shareholders. They do not deploy capital at this scale without clear evidence of return.
What the reports reveal, between the lines of cautious corporate language, is that the companies generating tokens at scale are achieving positive and improving returns on that investment. The constraint is no longer whether tokens are profitable to produce. It is how many tokens can be produced — and how quickly infrastructure can be built to produce more of them.
The race, in other words, has shifted. It is no longer a race to prove AI works. It is a race to own the infrastructure that generates the commodity that everything else runs on.
The new constraint
In a gold rush, the limiting factor is not the price of gold. It is the rate at which you can pull it out of the ground.
We are in that moment now, for tokens. The man in the bar offering you two dollars for one is being outbid — not because the deal is bad, but because the deals are getting better faster than anyone anticipated. Big Tech is not spending hundreds of billions because AI is interesting. It is spending because the numbers work, and because whoever builds the deepest mines wins.
The question is no longer whether tokens are valuable. It is who controls the pick and the shovel.
Judah Taub is Managing Partner at Hetz Ventures, one of Israel’s leading cybersecurity venture funds, and a donor to nonprofit organizations across the US and Israel
