Three Numbers for AI's Power Draw in China: 170bn, 196bn, and +44%

In one line: For the same figure — "2025 national compute-centre electricity consumption" — the National Energy Administration says 170 billion kWh, while CAICT says about 196 billion kWh. That is a gap of roughly 15%. Meanwhile, in Q1 2026, electricity use in internet data services grew +44.0% year-on-year.

1. The three numbers

Two totals for the same 2025, about 26 bn kWh apart (≈15%).

This is not a case of "one is wrong" — both are official. The divergence typically stems from what counts as a "compute centre" (self-built vs edge facilities, whether non-IT supporting load is included, etc.). When citing it, you must name which basis you are using — otherwise the same sentence can be "proved" by either number.

2. Why both figures belong in the record

This is precisely where trustworthy data separates from single-source repetition:

• Cite only one → the reader never learns a second exists, and will conclude you are wrong when they encounter the other.
• Record both, label the basis → the reader gets a range plus the reason for divergence. That is information a decision can actually rest on.

How we handle it: both entries are in the dataset, each with its source, each retrievable by record ID (tbm-155 and tbm-156). We do not merge them, and we do not average them.

3. The growth line matters more than the total

The totals diverge; the direction of growth does not:

For scale, China's total 2026 electricity consumption is projected to exceed 10 trillion kWh again, with peak load of 1.575–1.6 billion kW.

In other words, AI electricity demand is growing at a multiple of the national total — which is the pressure the power system felt before "compute–power coordination" became policy.

One more demand-side indicator: national average daily token calls rose more than 1,000-fold in two years (100 bn → 100 tn → 140 tn, official statistics).

4. Policy: from "East Data, West Computing" to "compute–power coordination"

The progression is clear: from "put compute in the west" (spatial) → to "make compute run on green power" (energy mix) → to "compute–power coordination" (treating compute as a dispatchable resource for the power system).

5. Infrastructure and efficiency today

Global comparison: as of June 2025, global computing-device capacity totalled 4,495 EFlops, of which intelligent compute accounted for 85% (CAICT basis).

⚠️ China's "1.59 million PFlops (FP16)" and the global "4,495 EFlops" are not on the same statistical basis (the former is intelligent compute at FP16 precision; the latter is total computing-device capacity). We do not convert between them; they are presented side by side only.

6. An easily missed downward signal: green certificate prices are falling

The other side of compute electricity demand is green power. And the price trend in the green certificate market runs against the intuition of "surging green-power demand":

Price ladder by year of electricity generation (compared at the same point in time):

Two readable patterns:

1. Freshness premium — at a single point in time, current-year certificates trade far above older ones (August: 6.32 vs 0.61, roughly a 10× spread). The market pays a premium for "this year's green power".
2. Across-the-board decline within four months — all three vintages fell (2026: 7.76 → 6.32; 2025: 5.71 → 3.21; 2024: 1.51 → 0.61).

Meanwhile, the Q1 2026 average certificate trading price was CNY 5.13 (+2.17× YoY), and 2025 national green-power traded volume reached 328.5 bn kWh (+38.3%) (this entry is confidence B — not verified word-for-word in an official transcript).

How to read this set: volumes are rising and the year-on-year price is up, but the four-month price trend is down. The two signals do not contradict each other — they point at different horizons. The former is an annual view; the latter is current supply and demand.

7. Green-power direct connection: three projects already running

Beyond policy, concrete projects are operating:

The last case deserves separate attention: it ties "compute moving west" to a dedicated power corridor — compute follows the electricity, not the other way round. That is "compute–power coordination" in its most concrete form.

8. Two sources to keep labelled

The dataset includes an entry stating that electricity accounts for about 56.7% of data-centre operating costs (tbm-158) — this is a brokerage research basis, confidence B, not official statistics. We label it and draw no conclusion from it.

Likewise, GPU compute rental price ranges (tbm-170) come from a compilation of public price quotes, confidence B.

Sources

Data floor: every figure in this article is drawn from public government / international organisation / official statistical sources. No commercial data-vendor sources are used.

Level key: L1 = government / international organisation / official statistical agency; L2 = international standards body / regulated exchange / state think-tank; L3 = industry association / corporate self-disclosure / official-media citation (grey — usable with original source named); L4 = commercial data vendor / self-media (prohibited by project policy).

Basis statement: tbm-155 and tbm-156 are two coexisting official bases differing by roughly 15%. This article records both and neither merges nor averages them. Please name the basis when citing.

Data current as of 2026-10-10. Each record is retrievable by record ID via the site's data index or the MCP endpoint.