Until the early 2000s, running software meant buying physical servers, placing them in a data centre, supplying cooling and electricity, and replacing hardware when it failed. If traffic spiked tenfold, you did not have the capacity; if traffic fell, you still paid. Amazon opened up to outsiders in 2006 the excess capacity it had built to handle its own retail holiday peaks. On 14 March 2006, S3 (Simple Storage Service) — object storage at a few cents per GB per month — went live. On 25 August 2006, EC2 (Elastic Compute Cloud) — virtual servers rented by the hour — launched. Together these two services made up the first large-scale "as-a-service" infrastructure offering.

In another corner of the same year, Doug Cutting and his team at Yahoo! released Apache Hadoop, a data-processing framework inspired by Google's 2004 MapReduce and 2003 GFS papers, spread across thousands of commodity servers. Hadoop pushed the phrase "big data" into corporate vocabulary: every company was now accumulating terabytes of log, click and sensor data and needed a way to process it. AWS plus Hadoop made the equation possible: "store it, scale it out, pay only for what you use."

Microsoft Azure joined in 2010 and Google Cloud Platform in 2011, and the big three carved up the market. By 2015 cloud-native companies like Netflix, Airbnb, Slack, Spotify and Reddit had never owned a data centre. Traditional institutions — banks, retail giants, governments — moved more than half their workloads to the cloud between 2018 and 2024. This was not just a technical migration: IT budgets shifted from capital expenditure (CapEx) to operating expense (OpEx); new companies could launch services with millions of users on a few thousand dollars.

From the 2020s the cloud also became the substrate of artificial intelligence: the 25,000 Nvidia H100 GPUs that trained ChatGPT ran on Microsoft Azure; AWS Trainium and Google TPU grew up as competitors. Together in 2024 the three hyperscalers were supplying around 300 billion dollars of annual infrastructure capacity — a "digital continent" consuming roughly 2 percent of the planet's electricity and belonging to no state. Owning your own server dropped to about 20 percent of computing; the remaining 80 percent is now either rented from a handful of companies' data centres or consumed through products built on top of them.