Microsoft water-efficiency drive stalls 7–11 GW of US AI data center capacity

The Sina analysis crystallizes a growing bottleneck in the AGI race: it’s no longer just about GPUs, but about power, water and local politics. According to the piece, the US has about 5 GW of AI data center capacity under construction and another 7–11 GW in planned projects that are delayed or frozen, even as analysts project up to 300 GW of AI‑driven data center power demand by 2030. Microsoft is leaning into aggressive water‑efficiency goals—driving average usage down from 2.3 to 0.27 liters per kWh and rolling out low‑ or zero‑water cooling—to make projects more palatable to regulators and communities. ([finance.sina.com.cn](https://finance.sina.com.cn/stock/usstock/summary/2026-06-28/doc-iniewssk1122523.shtml))

For the race to AGI, this is a reminder that scaling frontier models isn’t purely a chip problem. If local authorities won’t approve substations, transmission upgrades and water‑intensive cooling, the practical ceiling on deployable compute will sit well below what NVIDIA roadmaps imply. In the near term, that could slow the rollout of massive training clusters and, more subtly, tilt investment toward regions and technologies that can satisfy environmental and social constraints—hyperscale campuses near abundant hydro, nuclear‑adjacent sites, or more efficient accelerators.

Longer term, infrastructure friction might push labs to squeeze more capability out of each joule and liter: better algorithms, quantization, and model architectures may be forced not just by cost, but by permitting reality. If Microsoft and peers can’t turn planned capacity into energized racks, the AGI timeline becomes as much a story about grid modernization and water politics as about scaling laws.