China AI Atlas
A Tech Buzz China research product
Key Takeaways
AI infrastructure model

China–US AI Capacity Calculator

Compare useful compute as deliverable power rises, then see where sites or accelerator supply force each curve to flatten.

Illustrative starter ranges. These are not country estimates.
Capacity response

What more deliverable power can produce

The horizontal axis is a hypothetical power sweep, not a time forecast. Curves flatten when another ceiling binds.

United States and China AI capacity response curvesUseful delivered compute (ZFLOP/s) as deliverable facility power increases from zero to 40 gigawatts. The United States is a solid blue line with circles. China is a dashed red line with squares. Shaded areas are deterministic scenario bounds.0.005.0010.015.020.025.00816243240Accelerator ceiling · 13.7 GWAccelerator ceiling · 9.4 GWUSChinaDeliverable facility power available (GW)Useful delivered compute (ZFLOP/s)
US ChinaShading = entered scenario boundsDashed vertical line = comparison point
United States
10.1 ZFLOP/s
Range 4.03 ZFLOP/s – 21.8 ZFLOP/s
China
4.14 ZFLOP/s
Range 1.51 ZFLOP/s – 9.49 ZFLOP/s
Base scenario at equal power: US 2.44× China

Advanced assumptions

Facility, hardware, delivery, and workload ranges

Facility conversion

LowBaseHigh
Load factor
%

Average annual load divided by the usable facility-power ceiling.

PUE
×

Total facility energy divided by IT equipment energy; lower is better.

Accelerator share
%

The balance covers CPUs, memory, storage, and networking.

Hardware and delivery

LowBaseHigh
Accelerator draw
kW

Average device draw; other IT and cooling loads are modeled separately.

Utilization
%

Delivered accelerator math as a share of peak while the facility is loaded.

System delivery
%

Losses from communication, storage, scheduling, and inference software.

Workload conversion

LowBaseHigh
Active parameters
billion

Parameters actually used per generated token; MoE can be below total size.

Model overhead
×

Allowance for attention, context, routing, batching, and other work.

Deliverable power is swept directly on the chart’s horizontal axis. It is not varied again inside the shaded scenario envelope.

Methodology

How the curve is calculated

  1. 1. Find the binding ceiling.
    Usable facility GW = min(site GW, deliverable power GW, accelerator-supported facility GW).
  2. 2. Convert facility power to accelerator power.
    Accelerator GW = usable GW ÷ PUE × accelerator share of IT power.
  3. 3. Calculate useful compute.
    ZFLOP/s = accelerator GW × TFLOP/s/W × utilization × system delivery.
  4. 4. Convert compute to inference output.
    TFLOPs/token = 0.002 × active parameters (billions) × model overhead.

For every power point, the band checks the low/high corners of the inputs relevant to the selected output and reports the minimum and maximum. The extremes assume simultaneous co-movement and are not probabilities.

Token output covers autoregressive inference. Measured wall-power results, such as MLPerf Inference Power, are preferable to unmatched vendor peak specifications when they cover the relevant workload.

Accessible data

Read the plotted values

Sampled United States and China low, base, and high values from the capacity response chart
PowerUS lowUS baseUS highChina lowChina baseChina high
0.0 GW0 ZFLOP/s0 ZFLOP/s0 ZFLOP/s0 ZFLOP/s0 ZFLOP/s0 ZFLOP/s
6.7 GW2.49 ZFLOP/s4.9 ZFLOP/s8.82 ZFLOP/s1.43 ZFLOP/s2.94 ZFLOP/s5.46 ZFLOP/s
13.3 GW4.03 ZFLOP/s9.8 ZFLOP/s17.6 ZFLOP/s1.51 ZFLOP/s4.14 ZFLOP/s9.49 ZFLOP/s
20.0 GW4.03 ZFLOP/s10.1 ZFLOP/s21.8 ZFLOP/s1.51 ZFLOP/s4.14 ZFLOP/s9.49 ZFLOP/s
26.7 GW4.03 ZFLOP/s10.1 ZFLOP/s21.8 ZFLOP/s1.51 ZFLOP/s4.14 ZFLOP/s9.49 ZFLOP/s
33.3 GW4.03 ZFLOP/s10.1 ZFLOP/s21.8 ZFLOP/s1.51 ZFLOP/s4.14 ZFLOP/s9.49 ZFLOP/s
40.0 GW4.03 ZFLOP/s10.1 ZFLOP/s21.8 ZFLOP/s1.51 ZFLOP/s4.14 ZFLOP/s9.49 ZFLOP/s