China's Galaxy AI Campus Delivers First 120 MW in Under a Year; Cost Claims Rest on Paywalled Analysis

Chinese wind-turbine maker Envision Energy has completed the first 120 MW of its Galaxy AI data center campus in Ulanqab, Inner Mongolia, in less than a year, according to research firm SemiAnalysis, which visited the site in September and published its findings on November 7. The full campus is planned for up to 800 MW of IT capacity across four phases by the end of 2027, putting it on a similar scale to the roughly 820 MW of critical IT capacity at the Stargate campus in Abilene, Texas.

SemiAnalysis says the first phase was delivered in August 2026 at "a fifth of US costs," though the detailed cost breakdown sits behind its paywall and the per-megawatt figure was not published publicly. The firm estimates Chinese data center capex has fallen about 25% over five years to under $3 million per MW, but that figure is its own model output, not an audited benchmark.

What is verified, what is claimed

The primary source for the Galaxy campus is SemiAnalysis's site-visit report, titled "We Visited China's Stargate in Inner Mongolia." The firm toured the campus in September and describes a project that is vertically integrated from wind generation to rack-level power electronics. Envision owns the campus, operates two dedicated wind farms totaling 420 MW of installed capacity, supplies four-hour battery storage through its affiliate AESC, and has designed its own low-voltage power modules, solid-state transformers, and prefabricated substation.

The campus runs under China's "source-grid-load-storage" model on a dedicated transmission line rather than the public grid. Envision targets more than 60% of the first two phases' 240 MW load from its wind farms, which are sited roughly 80 kilometers apart to reduce correlation in low-wind periods. Ulanqab's wind capacity factor is 35–40%. All-in electricity cost in the region is approximately $0.055/kWh, according to SemiAnalysis.

In one 5 MW hall, solid-state transformers convert 10 kV AC directly to 800V DC for server racks, eliminating low-voltage transformers, UPS units, and rectifiers. Envision said on November 2 that the 800V DC system was running live AI workloads and plans a larger deployment in later phases. SemiAnalysis did not report a measured efficiency gain for the pilot.

Conflicting figures and unnamed customers

Several numbers in the public record do not align. SemiAnalysis describes a four-phase plan to 800 MW of IT capacity. Envision's own August commissioning announcement said the campus was designed to scale to more than 2 GW, and a November release gave a planned capacity of 3 GW. The 800 MW figure covers IT load only; the larger numbers may reflect total facility power or different project boundaries. The sources do not reconcile them.

The first-phase milestone also differs: SemiAnalysis says 120 MW was delivered in August 2026; a November 5 Caixin report, citing an Envision executive, described a 200 MW first phase handed over in late October. It is unclear whether the two references use the same power definition or mark different acceptance stages.

All four phases are built-to-suit for two internet companies that SemiAnalysis did not name. Their identities and the lease terms are undisclosed, so the revenue certainty behind the buildout cannot be assessed independently.

Vertical integration: real, but advantage unproven at scale

Envision's integration across wind turbines, batteries, power electronics, and data center construction is extensive. The company has installed more than 100 GW of wind turbines globally and manufactures its own battery cells through AESC. SemiAnalysis notes that factory pre-configuration of the full power chain—high voltage, low voltage, storage, and generation—can compress on-site commissioning from 3–6 months to less than three for the substation.

Whether that integration translates into a durable cost or schedule advantage for AI infrastructure is an open question. Hyperscalers still choose vendors mainly on resources and track record, SemiAnalysis says. The Chinese data center market is highly competitive: rents have roughly halved over five years to about $40/kW/month excluding power, and capex has fallen to under $3 million per MW by SemiAnalysis's count. In that environment, owning more of the value chain is one way to differentiate, but it has not yet been shown to be decisive.

Power reliability managed, not solved

The wind-plus-storage model reduces but does not eliminate the need for firm power. A four-hour battery shifts surplus from windy hours into the next calm period; it cannot cover multi-day lulls. SemiAnalysis's own modeling shows wind share rising from 40% (1:1 wind-to-load) to about 55% with 1.75x oversizing, and higher with storage and better load matching. Envision targets more than 60% for the first two phases. The campus also has diesel generators for emergencies, per a China Securities Journal site report in October. No source provides a measured hourly renewable share for continuous operation.

800V DC pilot is 5 MW, not campus-wide

The solid-state transformer installation that converts 10 kV AC to 800V DC is a 5 MW pilot in one hall. Nvidia's roadmap includes AC-compatible 800V racks in the second half of 2026, row-level power centers in 2027, and a facility-scale medium-voltage-to-800V DC design later. Galaxy's pilot is a working example of the final step, but no independent efficiency or reliability measurements have been published, and the architecture has not been demonstrated at the scale of the full campus.

Cooling and broader context

Galaxy uses a design common in northern China: liquid- and air-cooled servers share one water loop with no central chiller plant. SemiAnalysis places this within a longer-established practice in the region, enabled by local climate and permitted water temperatures. No independently measured annual cooling-efficiency result for Galaxy is available.

Inner Mongolia's broader data center pipeline totals around 25 GW of announced projects, including several gigawatt-scale campuses, according to SemiAnalysis's China Datacenter Model, which tracks over 1,000 facilities across 70+ operators. The Chahar High-tech Zone alone had nearly 30,000 construction workers on site across 22 data center projects at the time of the visit.

What remains unknown

• The per-megawatt construction cost for Galaxy and the basis for the "one-fifth" claim (paywalled).
• The identities and financial commitments of the two anchor tenants.
• Whether the 120 MW August delivery and the 200 MW October figure refer to the same metric.
• Measured PUE, electrical efficiency, and renewable penetration for the operating 120 MW.
• Whether the 800V DC architecture scales economically beyond the 5 MW pilot.
• How the campus handles extended low-wind periods beyond four-hour battery duration.

Envision is presenting a competing model for AI infrastructure: vertically integrated, renewables-anchored, and fast. Whether that model transfers beyond Inner Mongolia's specific regulatory, geographic, and supply-chain conditions is a separate question that the current evidence does not answer.