What it costs to run AI yourself, from a desk to a data centre: NVIDIA, AMD, Intel, Apple and China’s chipmakers, with graphics-card prices in 17 countries and full price histories in the US, Europe, the UK, China and Japan, each converted to dollars at that day’s exchange rate.
What a card costs if you buy it today: the cheapest board actually in stock in each of 17 countries, in US dollars at today’s rate, coloured by how much more (red) or less (blue) it costs than in the United States. Advertised prices for boards that are out of stock are not prices and are left out; hatched countries have nothing in stock. The list re-sorts from cheapest to dearest; its bar runs from the cheapest board in stock to the typical (median) one.
The whole table at once: each cell is that card’s price against the US, same colours as the map. Click a card to put it on the map; click a country to sort the cards by it.
The same machine in five markets, in US dollars at each date’s exchange rate. Shelf prices in Europe, the UK, China and Japan include VAT and US prices do not; switch to without sales tax for a like-for-like comparison.
The 2026 memory shortage reversed the usual curve: most machines cost more now than on launch day.
The chips that train and serve frontier models, from US and Chinese vendors. Purchase prices are reported estimates (no vendor publishes a list price); rent is the live weekly median where an open rental market exists.
Solid: NVIDIA; dashed: AMD; dotted: Intel. Median on-demand price across every provider listing the GPU that week; the year-on-year change appears only where at least 3 providers list it. Data through 2026-11-02. Chinese accelerators are not in an open rental dataset; their reported cloud prices are in the table above.