A Cost Curve for Global Compute

This blog post was originally an essay submitted for the ChinaTalk Essay Contest. ChinaTalk gave me around 200 USD which went into my ChatGPT and OpenCode subscriptions.

Sidenote: I love reading ChinaTalk for independent perspectives. Their recent piece here is worth a bookmark for anyone who wants a crash course on China macro.

Viewing AI through a Commodities Lens

AI is a very interesting topic for a commodities analyst, because it has a physical reality tied to both the energy (for power to run models) and base metals complex (for building related data center and grid infrastructure). AI’s connection to commodities is also becoming more explicit with compute itself being launched on CME as a derivatives product referencing GPU rental costs.

It comes naturally that commodities frameworks can be a good lens through which to understand AI. The most fundamental framework used across commodities is a Supply and Demand balance, which allows analysts to quantify surpluses or deficits.

In this essay we look at developing a framework to understand AI supply, by building a global supply curve for compute. It becomes extremely clear how US chip controls directly shape the economic structure of AI supply.

For myself personally, another reason why I was especially interested in building this project was that, while working as a cross-commodities analyst on a power trading team over the past few years, it increasingly occurred to me that China is extremely well positioned and could potentially dominate in an AI driven world, purely by its advantage in energy costs (cough, coal). I was interested to see if this hunch can be supported by data.

The Compute Cost Curve

I ended up creating two global cost curves for compute. One curve with the same NVIDIA GB200 NVL72 technology everywhere (i.e. no chip restrictions on China), and another with export control restrictions in place where China only had access to its domestic Huawei CloudMatrix384 which is estimated to require 2.5x as much electricity for the same amount of compute. 

This is essentially an attempt at constructing a 2025 cost curve, based on data centers operational by December 31, 2025, and electricity prices effective during 2025.

Explore the curves in depth here

It becomes apparent that US export control restrictions result in vastly different pictures of global compute cost curves. US motivation for chip control becomes crystal clear.

This is in line with the great work Aqib at China Talk has already done here comparing how much AI you can get for 1$ in the US vs China by estimating the cost of constructing and operating a 400MW data center over three years. This project here builds upon that by adding data center capacity into the picture, and the cost curve here considers solely the electricity costs after a data center has been built.

It should be noted that the compute cost curves here are exploratory in nature. Coverage is far from complete as only public domain data were used, and you will find many datacenters not being represented on the curve. Certain workarounds in arriving at estimations were also made, and these are detailed in the full methodology. Nonetheless, I think these exploratory cost curves still provide a useful and quantitative lens to objectively view the global AI race, especially between the US and China.

Methodology

The main steps to put together a compute cost curve were: 1) collating data center capacities of each region, 2) collating each region’s electricity prices, and 3) converting electricity prices into compute cost via technology assumptions

1. Collating Data Center Capacities 

Data center capacities operational by 31 Dec 2025 were collated for commissioned design IT power, i.e. non-redundant IT or critical-load capacity that excludes cooling and facility overhead. 

An AI agent was used for this search, and the collated capacity data were collated in a csv file: Global facility register.csv 

This is not an authoritative list, and some workarounds were made to get data such as approximating capacities from published charts as underlying data was not available.

2. Collating Electricity Prices

Explicit datapoints stating data-center tariffs were sparse. To go around this, comparable electricity-price proxies were used based on public large-load tariffs, official state industrial averages, and official national non-domestic averages. Some examples of data points used include:

• China provincial industrial tariffs;
• EIA 2025 state industrial averages for US states;
• Hydro-Québec Rate LG for Montreal;
• UK provisional 2025 non-domestic average.

Demand-based tariffs were standardized using a 90% load factor and 730hr/mth. Currencies were converted to USD using 2025 annual-average exchange rates.

These electricity prices were mapped to the capacity regions previously found.

3. Converting Electricity Price into Compute Cost

The standardized-technology calculation uses:

• GB200 NVL72 dense-BF16 performance: 180 PFLOP/s;
• rack power: approximately 120 kW;
• standardized PUE: 1.20.

Conceptually:

Facility power = IT power × PUE

Electricity cost per 10¹⁹ FLOPs
  = facility power × electricity price
  ÷ compute performance
  × required FLOPs

The export-constrained curve multiplies China’s resulting electricity cost per FLOP by 2.5.