Meta announced this week that it is expanding its data centre investment in El Paso, Texas, from $1.5 billion to $10 billion. The facility, on a site outside the city, is designed to deliver one gigawatt of computing power by 2028. To put that in context: one gigawatt is roughly the output of a large nuclear power station, and it is being committed to a single data centre by a single company.
El Paso is the third Texas site in Meta's infrastructure network, alongside facilities already under development in Indiana and elsewhere. The company has separately signed a $27 billion infrastructure deal with Nebius, a European cloud provider, and expanded its arrangements with Nvidia to include millions of AI chips. The El Paso announcement is not a standalone decision; it is one piece of a capital expenditure programme that Bloomberg estimates will exceed $60 billion this year alone.
The geography of the choice is worth examining. El Paso sits on the US-Mexico border in a region that faces chronic water stress and summer temperatures regularly above 40 degrees Celsius. Data centres are water-intensive: evaporative cooling systems can consume millions of litres per day. Texas's deregulated electricity grid, which attracted industry with low prices, has also proven vulnerable to demand spikes. The February 2021 grid failure, which killed hundreds and cost the state billions, remains a live cautionary tale.
None of that appears to have slowed the investment. Meta says the site will create more than 300 on-site jobs when completed, which is a modest number for a $10 billion project. The economic case is primarily about infrastructure for training and running the next generation of AI models, not about local employment. Texas has offered significant tax incentives to attract the investment, a pattern repeated across data centre corridors in Virginia, Iowa, and Georgia.
The scale shift in this announcement, from $1.5 billion to $10 billion in a single update, captures something important about the current moment in AI infrastructure. These are not stable, considered plans; they are plans being revised upward in real time as the compute requirements of large model training become clearer. Google, Microsoft, and Amazon have all published similarly upward-revised infrastructure commitments in the past six months. The figures involved now routinely exceed what any single national government spends on science and technology research.
There is a straightforward argument for why this matters beyond the optics of big numbers. AI capability, under current architectures, scales with compute. More compute produces better models, faster. The companies building the most compute today have a structural advantage in the next generation of model releases. Infrastructure commitments made in 2026 shape the competitive landscape for the rest of the decade.
The less straightforward argument concerns what all this capacity actually requires from the rest of society. A gigawatt data centre needs not just electricity but the transmission infrastructure to carry it, the water to cool it, and the political accommodation to permit it. As Meta, Microsoft, Google, and Amazon collectively commit to tens of gigawatts of new capacity, they are placing demands on public infrastructure that were not built with this use case in mind. The bill for that is arriving, through utility rate increases, local water allocations, and grid reliability decisions, in places that had very little say in the investment choices driving it.
The El Paso expansion is a routine announcement in an industry that has become accustomed to scale. That normalisation may be the most interesting thing about it.