Envision Commissions Galaxy Campus in Ulanqab, Establishing a New Model for Gigawatt-Scale AI Infrastructure

Envision commissioned the Envision Galaxy Campus in Ulanqab, a 120,000-square-meter AI supercomputing facility targeting over 2GW capacity and up to 1 million PFLOPS, designed to support up to 1 million AI accelerators in a single structure. The project is directly powered by renewable electricity and uses Envision’s AI Power System (including dedicated transmission and large-scale energy storage) to deliver “up to 10x” more compute output per square meter than conventional data centers. It also aligns with the Mission Gobi plan to build 5GW of green AI computing capacity by 2030, suggesting a constructive development for green AI infrastructure capacity and energy-efficiency constraints.

Analysis

This is less a near-term revenue event than a signal that the bottleneck in AI is shifting from GPUs to power delivery. The investable winners are the picks-and-shovels around high-density compute: switchgear, thermal management, grid interconnect, storage, and fiber backhaul. That favors names like VRT, ETN, ANET, and CCI more than pure-play AI software, because the spend is tied to physical buildout rather than model adoption.

The contrarian read is that “green AI” does not automatically mean cheaper or faster capacity. A remote, renewable-heavy campus can still be constrained by accelerator supply, interconnection timing, and utilization ramp, so a lot of the announced capacity may stay capitalized long before it becomes cash-generative. If the market extrapolates this into a broad desert-data-center boom, that is overdone; inference workloads still need latency proximity, which protects metro colocation economics in DLR/EQIX.

Time horizon matters: over days, the trade is mostly sentiment around power/AI infrastructure; over 1-3 months, watch for orders into transformers, batteries, cooling, and optical gear; over 6-18 months, the key is whether this becomes a replicable template or a one-off showcase project. The thesis is falsified if utilization lags, financing costs rise, or commissioning slips enough to expose the gap between headline MW and delivered compute.

Key Decisions for Investors

• Long VRT / ETN on any pullback; 1-3 month view. These are cleaner beneficiaries of rising rack density and power-conditioning spend, with better asymmetry than chasing the AI end-user names.
• Pair trade: long ANET, short DLR for 3-6 months. The market may underprice networking capex while overestimating immediate pressure on premium colocation pricing; stop if DLR shows accelerating pricing or occupancy inflects higher.
• Watchlist long SCPAF only if the company is directly exposed to power/grid buildout; otherwise no action. Require proof of signed orders or project finance before owning the story.
• No trade in WWRL for now. The article does not create a clean earnings catalyst; treat it as a monitor unless there is disclosed exposure to transmission, storage, or data-center electrical systems.