By Djellal Djouad
A token of LLM output is a manufactured good. Its marginal cost is dominated by electricity. And electricity, unlike GPUs and unlike capital, is not priced globally. It is priced by interconnect, by climate, by the political alignment between the grid and the data center. That makes the AI token, structurally, the first software commodity in history whose floor is set by where it is born.
The market is pricing AI infrastructure as if a megawatt in Virginia and a megawatt in Inner Mongolia were the same thing. They are not. And they are diverging in front of the consensus, in plain sight, with public data.
Today we publish a working paper that maps the divergence:
Tokens, Watts, and Geography — A Four-Tier Framework for Pricing the AI Inference Trade, 2026–2030 (Zenodo, open access, 15 pages, 51 references)
The world's AI-relevant geographies group into four tiers, defined by delivered electricity cost and the ability of the grid to dispatch new datacenter load:
Tier 1 — Sovereign Capable. China (62 GW of operating nuclear, 37 GW under construction, target 110 GW by 2030), the Gulf (Stargate UAE's 5 GW campus, HUMAIN's 6.6 GW path to 2034, Barakah nuclear already running), the Nordic countries (Microsoft's US$ 6.2 billion Norwegian commitment, free-cooling PUE below 1.10), France (nuclear at 70 percent of generation). Effective delivered electricity in the US$ 30–80 / MWh band. State policy treats AI inference as a strategic export industry.
Tier 2 — Emerging Stretched. India has US$ 310+ billion of sovereign-AI commitments from Reliance, Adani, and Tata, on a grid that supports 1.5 GW of installed data-center capacity today. South Korea has committed US$ 65 billion through 2027 and is building 3 GW of new transmission into the Yongin semiconductor cluster. Japan has SoftBank's Portsmouth nuclear consortium, and industrial electricity at US$ 110–140 / MWh. The ambition is real. The grid is not yet there.
Tier 3 — Power Constrained Wealthy. The United States in the PJM zone, Germany, the United Kingdom. The PJM annual capacity auction has cleared at or near the regulatory cap three years in a row — US$ 269.92, US$ 329.17, US$ 333.44 per megawatt-day in 2025/26, 2026/27 and 2027/28 — and for the first time in the history of the RTO, the 2027/2028 auction came in short of its reliability target by 6,623 megawatts. Hyperscaler capex of US$ 725 billion in 2026 cannot manufacture an electron that the grid is not built to deliver.
Tier 4 — Arbitrage Suppliers. Russian Siberia is the cheapest electricity on earth — wholesale at US$ 20–35 / MWh, free-cooling nine months a year, PUE under 1.15 in Norilsk and Yakutsk. It is already being exported, quietly, to Chinese cloud workloads. Canada's Quebec hydro, Brazil's Norte/Nordeste, Australia's Pilbara sit in the same bucket. The cheapest tokens that reach the global market in 2026 are, in part, the product of electrons that crossed a national border before they did.
What the framework prices
The unit-economics identity is simple. The marginal cost of a token equals the energy required to produce it, times the delivered local electricity price, plus capex amortization. Two of the three terms are essentially geographic. The third — amortization — is global today but will be amortized down by 2027–2029, exactly the window in which most sell-side terminal-margin models assume gross margins expand to 55 percent.
When you replace the consensus "flat global electricity" assumption with the four-tier weighted average, the US-anchored portion of the AI infrastructure trade falls from a terminal margin of ~55 percent to 35–42 percent, and terminal ROIC drops from 22 percent to 15–17 percent. On a 17× multiple, this is a 30–50 percent miss on terminal value for the US-anchored leg.
Shanghai's instrument
The Shanghai Futures Exchange disclosed in late May 2026 that it is designing AI-token futures. Trade press has framed the contract as a domestic Chinese hedging tool. The paper argues it is something else: the world's first tradable instrument on the eight-way cross-tier electricity spread of the global AI inference market, denominated in tokens rather than megawatt-hours. The geography is the underlying. The token is the wrapping.
This is the same transformation that took crude oil from a spot-cargo trade to a structured commodity in the 1980s, and the same kind of instrument that priced the regional spread of natural gas in the 2000s. The contract will not list tomorrow. But it is being designed, today, by an exchange that already runs the world's deepest copper and steel-rebar contracts. SHFE understands what it is doing.
Five testable predictions, Q3 2026 – Q4 2027
The paper closes with five falsifiable calls. Three of them:
PJM's 2028/2029 capacity auction will again clear at or above the cap, and the backstop reliability mechanism will be triggered before year-end 2026.
At least one US hyperscaler will publicly disclose a multi-gigawatt commitment to a non-US sovereign-capable jurisdiction — the Gulf, the Nordic countries, or France — by Q2 2027. The Microsoft–G42 200 MW announcement of mid-2026 is the first instance; the next round will be measured in gigawatts.
A first publicly disclosed cross-border power agreement between a Russian Siberian operator and a Chinese hyperscaler cloud will appear, denominated in token-output units rather than megawatt-hours, during the second half of 2027. The arbitrage layer will move from grey to formal.
Read the paper
doi.org/10.5281/zenodo.20543628 — open access, 15 pages, 51 cross-region references.
Also on Academia.edu, with a 5-minute audio summary.
Comments to info@crossvol.com.
This paper is the geographic and energy-floor extension of The China AI Disruption Thesis (CrossVol Research, May 2026). Where that paper argued the sell-side consensus is six months late on the bilateral US–China trade, this one argues the bilateral framing itself is incomplete.
CrossVol Team & Djellal Djouad
Further reading