The number of AI chips in the world is doubling roughly every nine months
Epoch AI counts about 20 million AI chips in data centres today and expects around 200 million by the end of 2028. That is the single number that explains most of what happens next.
There are roughly 20 million AI chips running in data centres worldwide right now, and according to the research group Epoch AI that figure is doubling about every nine months. Carry that forward and you land at something like 200 million by the end of 2028, ten times today’s total. The New York Times built a piece around those estimates this week, drawing on Epoch alongside Cleanview and SemiAnalysis.
Doubling times are hard to feel intuitively, so it helps to translate. Nine months means that by the end of next year, more AI computing capacity will have been installed than exists in the world today. It also means that the models being trained in 2028 will have access to an amount of compute that nobody has run an experiment on yet, which is why lab researchers talk about the next few years the way they do. Two honest caveats belong next to the number. First, these are estimates. Nobody publishes a global chip census, so analysts reconstruct it from manufacturing capacity, packaging constraints and company disclosures, and the various groups do not agree exactly. Second, a chip installed is not a chip usefully employed. Capacity can be built faster than demand for it arrives, which is the central worry behind the current debate about whether the AI buildout is a bubble.
What is behind this. Compute is the one input in AI that behaves like a commodity. You cannot buy better research ideas, but you can buy more chips, and for the past six years more chips has reliably produced more capable models. That is why the spending looks the way it does: the companies involved are not buying capacity for the product they ship today, they are buying the ability to run the experiments that decide who is ahead in 2028. It is also why the physical constraints have moved from silicon to electricity and land. A doubling in chips means a doubling in power draw, and power stations take longer to build than data centres do.
What this means for you: if you use AI tools, the near-term effect is the pleasant one. More capacity means falling prices, which is exactly what the last two months of price cuts have looked like, and it means the free tiers keep getting better. The medium-term effect is worth watching more carefully, because a buildout this size gets financed with debt against expected revenue, and if the revenue arrives more slowly than the chips do, the correction lands on the companies whose tools you have built your workflow around. Nothing to act on today. But when someone tells you AI progress is about to plateau, or that it is about to go vertical, this is the number the argument actually turns on.
Sources
Google put an image generator in Google Earth, then pulled it a day later
Nano Banana 2 let anyone type a prompt and get a fake satellite view of a real place. Researchers produced refugee camps and a nuclear plant that do not exist, and Google rolled the feature back within roughly 24 hours.
