AI is now billed by the token, generative compute's 'kWh,' whose price hides three cost levers (equipment, electricity, contract) that turn identical per million quotes into very different bills.
China is now consuming 140 trillion AI tokens a day, more than 1,000 times the daily volume from early 2024. That curve is being read as proof that AI compute has become a commodity. The per-million-token price, the only number on the vendor invoice, is not what it appears to be.
The token, the unit in which large language models charge for the text they read and write, is being lifted into the same role electricity's kilowatt-hour plays in a household bill. The analogy is now institutionally backed. China's National Data Administration issued a standardized Chinese term for the token in March 2026, the first time a regulator has put its weight behind the unit. State Council briefings, reported by TechNode, put March 2026's daily token throughput above 140 trillion, up more than 40% from the end of 2025. Volume is real. Comparability, the property the per-million number is supposed to deliver, is not.
Domestic Chinese models now advertise headline rates as low as single-digit yuan per million tokens, with some open-source providers charging even less. The bill a CFO actually receives is a different object. Three entangled cost drivers shape every quote, and no vendor breaks them out on the invoice.
The first is equipment cost. Top-end AI accelerators are not sold by the rack anymore. Chip-industry commentator Wu Hao, writing in a feature on the chip vertical at leiphone.com, describes the tier above single-rack servers as clusters of 32 or more boxes tied together by fast interconnect. Only about 30 customers in China can afford to rent a full cluster. Most providers lease capacity, not hardware. Idle boxes, peak-rate boxes, and boxes reserved for training runs that did not happen still get amortized into the units that did sell. A vendor that paid full price for that hardware is recouping depreciation regardless of how efficient the model looks on a benchmark.
The second is operating cost, and the biggest lever inside it is geography. Tokens are produced in data centers. Data centers pay different electricity tariffs depending on province, and Chinese policy increasingly steers compute capacity toward lower-cost Northwest grids. A vendor running on Inner Mongolia power is not running on Guangdong power. The delta shows up in every unit shipped out, but the customer's invoice does not show where the power came from.
The third is contract length. Vendors who sold a multi-year capacity contract during the 2024 GPU shortage locked in costs. Vendors who bought spot capacity in 2026 are paying down the depreciation on hardware whose sticker price has since fallen. A long contract means the vendor is shielded from those falls, and can stay cheap. A spot-tied provider reflects them on the next quote. Either quote can be the cheaper-looking one, depending on which side of the cycle the buyer reads.
A provider with locked-in Northwest capacity and a multi-year GPU deal can underbid a fully-loaded competitor without being more efficient. The market currently has both kinds of provider in it. According to Wu Hao's analysis, the visible price compression on per-million tokens is the result of this structural subsidy as much as it is the result of better models. When the contracts roll, the cheap tokens get repriced.
For a procurement lead, the practical move is to stop reading the per-million number as a comparable. Three questions to put to each vendor: what share of your hardware is leased versus owned, and at what contract length; what province does your inference capacity run in; and what is the term and renewal structure on the deal being quoted. The answers turn the per-million figure from a price comparison into a cost conversation.
The unit looks stable while the volume is rising so fast. The NDA's March 2026 token standardization has locked that readability into the public procurement conversation. Long-capacity contracts signed in 2024 and 2025 start rolling over in late 2026 and 2027. When they do, the bills on per-million tokens that look cheapest today will be the ones most likely to reprice first.