By Djellal Djouad. October 4, 2026.
The sell-side has two ways of talking about AI infrastructure, and both of them hide the same fact.
The first is capex in absolute terms. Six hundred billion here, seven hundred billion there, a number large enough to end the conversation. The second is inference output in narrative terms. The price of intelligence is collapsing, tokens are getting cheaper every quarter, so the unit economics will take care of themselves. One framing is too big to argue with. The other is too smooth to question. Neither one asks what a token actually costs to produce.
A token is a manufactured good. Its marginal cost is dominated by one input, and that input is electricity. Once you price inference off the electricity floor instead of off a narrative, the floor does not move the way the consensus assumes. Models that ignore it are not off by a rounding error. They are off by an order of magnitude the market has not yet been forced to confront.
That is the argument of "Tokens, Watts, and Geography: A Four-Tier Framework for Pricing the AI Inference Trade, 2026-2030". The framework sorts the world not by GDP and not by chip access, but by the thing that actually sets the floor: the price and availability of power at the interconnect.
Tier 1, Sovereign Capable. China, the Gulf, the Nordics, France. Power is cheap, dispatchable, and politically backed. These are the places that can run inference at the floor and export the surplus.
Tier 2, Emerging Stretched. India, South Korea, Japan. Demand is real, the grid is not ready, and the gap is the story.
Tier 3, Power Constrained Wealthy. The United States on PJM, Germany, the United Kingdom. Capital is abundant and electrons are not. This is the tier that is pricing inference as if the floor were someone else's problem.
Tier 4, Arbitrage Suppliers. Russian Siberia, Canadian hydro, Brazilian Norte and Nordeste, Australian Pilbara. Stranded power looking for a load. The release valve, and the swing factor nobody is modeling.
The reason this paper matters is not that it has a view on AI. Everyone has a view on AI. It matters because it supplies the missing floor in a larger piece of work.
"The AI Infrastructure Repricing" argues that the capex cycle is being financed in ways that look a lot like past credit cycles. "Convergent Faults" maps how private credit has synchronized that risk across the system. "The China AI Disruption Thesis" takes the geography to its conclusion. Each of those rests on an assumption about what inference costs, and until now that assumption was borrowed from the same narrative this paper takes apart. "Tokens, Watts, and Geography" is where the floor gets built. The rest of the program stands on it.
If you only read one thing from the research program this quarter, read this one first, because the others assume it.
Read the paper: https://djellaldjouad.com/papers/tokens-watts-and-geography/
$NVDA $VST $CEG $TLN $NRG $MSFT $GOOGL $AMZN $TLT $SPY
Further reading
The AI Infrastructure Repricing: https://djellaldjouad.com/papers/the-ai-infrastructure-repricing/
Convergent Faults: https://crossvolresearch.com/
Working papers and books: https://crossvolresearch.com/