OpenAI construction on Stargate Phase 2 begins in Arizona, featuring first-of-its-kind liquid-immersion cooling.
What Stargate Phase 2 Signals
OpenAI has broken ground on Stargate Phase 2 in Arizona, a construction push that pairs multi-gigawatt power planning with liquid-immersion cooling. The project’s stated scale—on the order of five gigawatts of liquid cooling capacity—puts thermal design at the center of the build, not as a retrofit. For teams that plan AI capacity, that framing matters: the limiting factor is no longer only how many GPUs you can order, but whether power delivery and heat rejection can keep those chips in a usable operating range for long training and inference runs.
Phase 2 construction also implies a multi-stage campus model. Early phases establish site, interconnect, and core plant; later phases densify compute. Liquid immersion is being treated as a first-class facility choice rather than an experimental bay. That shifts risk from “will this rack cool?” to “can the site’s electrical, water, and maintenance systems support immersion at campus scale?”
Why Liquid Immersion Changes Facility Design
Air cooling hits practical limits as rack power density rises. Immersion places boards or modules in a dielectric fluid so heat moves by convection through liquid instead of forced air. That can cut fan load, reduce the need for massive cold-aisle airflow, and allow denser packing of accelerators—if the rest of the plant is designed for it. The fluid loop still has to reject heat to outdoors (dry coolers, towers, or other heat exchangers), so immersion does not remove cooling infrastructure; it relocates and reshapes it.
First-of-its-kind immersion at this scale means many details will be learned on site: fluid handling, leak containment, service procedures for hot-swapping hardware, and how to drain and reseal enclosures without long downtime. Operators should assume longer mean-time-to-repair early on, and plan spare capacity and clear isolation zones until procedures stabilize.
- Power path: multi-gigawatt load needs staged interconnect, redundancy, and clear failure domains so a cooling or electrical fault does not take the whole campus offline.
- Thermal path: immersion tanks, secondary loops, and outdoor heat rejection must be sized together; a mismatch at any stage becomes a hard capacity ceiling.
- Ops path: technicians need fluid-safe tooling, PPE, and runbooks that air-cooled data centers never required.
Practical Takeaways for Builders and Buyers
If you are planning GPU clusters, treat cooling and power as co-equal design drivers with model architecture. Ask vendors and colocation partners how heat leaves the rack, how service works under immersion, and what happens when a loop segment fails. Prefer designs with clear N+1 (or better) on pumps, heat exchangers, and power feeds—not just on servers.
If you consume OpenAI-scale capacity rather than build it, Phase 2 construction is a capacity and reliability signal: more training and serving headroom depends on plant delivery as much as on chip shipments. Watch for when cooled megawatts come online, not only for marketing milestones. For internal roadmaps, model scenarios where compute is available but thermally constrained; that mismatch is common when hardware arrives before facility systems are fully commissioned.
What to Watch as Construction Proceeds
Arizona construction will surface concrete constraints: grid interconnection timelines, water and heat-rejection strategy in a hot climate, supply chain for tanks and dielectric fluid, and workforce training for immersion maintenance. None of those are side issues; each can delay usable capacity even after buildings stand.
The useful lens is systems engineering, not hype. Stargate Phase 2 with five-gigawatt-class liquid cooling is a bet that immersion can be operated safely and continuously at industrial scale. Teams elsewhere can borrow the same checklist—power, heat, serviceability, and staged bring-up—whether or not they ever deploy immersion themselves.