If you missed our best recent articles:
“When a measure becomes a target, it ceases to be a good measure.”
- Goodhart’s Law
There are technological revolutions that begin with a very simple promise: things that were once expensive will become cheaper.
The internet brought the cost of information distribution to near zero.
The cloud made the cost of renting compute far more flexible than building servers from scratch.
SaaS transformed software from a massive capital investment into a monthly subscription.
And now, AI promises something even greater: to turn intelligence-or at least the portion of intelligence measurable in tokens-into something that can be rented, priced, and consumed in small, granular units.
But it is precisely here that the story begins to get more dangerous than the headlines suggest.
Because once intelligence is measured in tokens, the token is no longer just a technical concept. It becomes a price. It becomes a KPI. It becomes an invoice. It becomes a target for businesses to optimize. And when a measure becomes a target, human behavior begins to revolve around it.
That is when Goodhart’s Law meets AI.
A token does not appear out of thin air. Behind every response from ChatGPT, Claude, Gemini, or DeepSeek lies a very long physical chain: GPUs, HBM memory, data centers, electricity, cooling water, transformers, copper, concrete, land, zoning, transmission lines, and billions of dollars in debt to finance that entire infrastructure layer.
In other words, the world is trying to turn intelligence into a digital commodity. But to create that commodity, it requires an extremely physical industrial foundation.
This is the first paradox.
Token prices are falling very rapidly. Claude Opus 4 processes one million output tokens for about $25. DeepSeek V4 Pro does the same for less than $1. The gap between U.S. frontier models and Chinese efficiency models is no longer 10–20%. It is 10x, 20x, or even more. Meanwhile, OpenAI and Anthropic are entering their pre-IPO phase, at the exact moment both need to prove that AI labs are not just cash-burning laboratories, but businesses with sustainable unit economics.
This is the second paradox.
While token prices are falling, token consumption behavior within enterprises is increasing in very strange ways. At Amazon, employees use AI agents to run nearly meaningless tasks just to climb internal token consumption leaderboards. Meta, Uber, Walmart, and many other large enterprises are facing their own versions of the same problem. Token consumption is viewed as proof of AI adoption. But once the token becomes a target, it no longer accurately measures productivity. It begins to measure the ability to game the system.
This is the third paradox.
To serve that volume of tokens, the world needs ever more electricity, chips, copper, transformers, and data centers. Microsoft has a massive Azure backlog it cannot deliver due to power shortages. The grid interconnection queue in the U.S. stretches for years. Transformer lead times have increased from months to years. Carbon futures are reflecting a very clear expectation: when AI data centers truly come online, the electricity market will no longer have the comfortable slack it once had.
Therefore, the AI story in 2026 is no longer as simple as:
“Will AI change the world?”
The answer is almost certainly yes.
If the answer is yes, this will be one of the most significant infrastructure investment cycles of the 21st century. AI will not just be a new software layer, but a new production foundation-much like electricity, railroads, the internet, or the cloud before it.
Who pays the bill for that process, for how long, and will the cash flow from productivity appear in time before physical costs, capital costs, and technology depreciation begin to squeeze the entire ecosystem?
In today's article from Viethustler, I will cover five parts:
Part I - The Token Economy: the price of intelligence is collapsing, and why the OpenAI–Anthropic price war could worsen the economics of the entire industry.
Part II - Tokenmaxxing: when businesses mismeasure AI adoption, employees learn to game the system, and the current revenue of AI labs may contain a layer of unsustainable demand.
Part III - AI Renting & True ROI: why the pay-per-token model disrupts SaaS economics, and why most businesses have yet to reach the stage of true productivity.
Part IV - The Raw Material Bottleneck: AI is software on the surface, but it is industrial infrastructure at its core.
Part V - Convergence: as token prices get cheaper, infrastructure costs get more expensive, and demand becomes more distorted-who is the one left holding the bill?
