Stargate is a Delaware based joint venture created by OpenAI, SoftBank, Oracle and MGX to fund and operate a new generation of AI focused data centers primarily for OpenAI’s models. The partners talk about up to 500 billion dollars in total spending and around 10 gigawatts of power capacity over just a few years, with early sites centered around Abilene in Texas and additional campuses announced across several US states and in regions like the UK, UAE, Norway and Argentina. Oracle in particular has committed to several gigawatts of hosting, with facilities expected to hold more than two million accelerator chips when fully built out.
Epoch AI’s research provides the missing physical context. Its Frontier Data Centers Hub uses high resolution satellite images, building permits and other open data to estimate the size, power draw and compute of the largest AI campuses in the US, including projects linked to OpenAI and its competitors. For existing sites, Epoch AI estimates that just thirteen of the biggest AI data centers already represent roughly 15 percent of global AI compute shipped by chipmakers by late 2025, equivalent to around 2.5 million GPUs of the NVIDIA H100 class. Some single campuses are approaching or exceeding one gigawatt of electrical capacity, enough to power around a million US homes, and several more 1+ gigawatt sites are expected to come online in 2026 alone.
Putting this together, a fully realized Stargate network at 10 gigawatts looks less like a classic “cloud project” and more like a small national grid dedicated to machine learning. The energy appetite, land use and cooling demands are at a scale that invites comparison to Cold War era industrial or military programs rather than typical commercial IT build outs. At the same time, the infrastructure is nominally framed as enabling breakthroughs in areas such as drug discovery or climate modeling, even though control over the stack is concentrated in a small set of corporations and most of the capacity will be exclusive to OpenAI. That tension between public benefit narratives and tight corporate ownership echoes the broader evolution of OpenAI itself, which shifted from a charity like mission to a heavily capitalized, closed ecosystem around proprietary models. Epoch AI’s mapping work implicitly underlines this by showing how opaque many of these facilities still are, even when their rooftops are clearly visible from orbit.
