Envision Galaxy Campus: Wind-Linked AI Data Center in Inner Mongolia

Short answer: Envision Group says it commissioned Galaxy Campus in Ulanqab, Inner Mongolia, in August 2026 — a large AI supercomputing hall fed by dedicated renewable power and designed to scale to multi-gigawatt capacity. Trade reporting adds wind farms, battery storage, a live 800V DC distribution path, and tenant internet giants leasing chip-filled halls. Headlines that say "800MW wind-powered data center" usually mix SemiAnalysis's ~800MW IT campus target with Envision's longer 2–3GW build-out story; all figures are real, but they are not interchangeable.

Inner Mongolia is already the default landing zone for Chinese hyperscale power in explainx.ai's SemiAnalysis capacity piece. Galaxy is a single flagship project inside that region: Envision arrives as a wind-turbine and storage integrator that now builds, powers and outfits AI halls itself.

Update (November 8, 2026): SemiAnalysis site visit adds regional scale and cost context

On November 7, 2026, SemiAnalysis published We Visited China's Stargate in Inner Mongolia, a paywalled field report whose public section adds numbers our Envision-first summary had not stressed.

SemiAnalysis frames Ulanqab's Chahar High-tech Zone as a corridor with nearly 30,000 construction workers across 22 data-center projects on site during the visit, not only Galaxy. That undercuts the meme of empty "shell" halls in the region and matches explainx.ai's earlier national GW census caveat: land and power can scale faster than chip fill rates, but labor and civil work are visibly active.

On economics, SemiAnalysis quotes a competitive China colocation market with rent near $40 per kW per month excluding power, capex under $3 million per MW, and all-in Inner Mongolia power near $0.055 per kWh. It also claims Galaxy's first 120 MW phase landed at roughly a fifth of US costs for comparable scale — a ratio we already flagged as contested; the visit adds a reporter eyewitness, not independent audited cost data.

SemiAnalysis directly compares Galaxy's ~800 MW IT campus plan to Crusoe's Abilene core, which it describes as about 820 MW of critical IT on 1.2 GW of power, with a later behind-the-meter expansion taking Abilene toward ~1.5 GW IT. That makes Galaxy one of China's largest single campuses by power envelope, not only by press release language.

On cooling, the newsletter stresses a northern China pattern Galaxy uses: liquid- and air-cooled servers on one water loop with no central chiller plant, while US sites often keep chillers for air-cooled gear. NVIDIA's Vera Rubin guidance on 45°C water without chillers gets cited as converging with designs already running at scale in Inner Mongolia.

Timeline note: SemiAnalysis says the full Galaxy campus could complete by end of 2027, while Bloomberg relayed end of 2028 in October. Treat those as different reporting passes, not a confirmed slip.

TL;DR: confirmed facts vs mixed headline numbers

What Envision says it built

Envision's August 6, 2026 announcement describes Galaxy as a next-generation AI infrastructure campus meant to accelerate large-scale AI deployment. It highlights:

• A 120,000-square-meter AI supercomputing facility, described as roughly twenty football pitches.
• Power drawn from renewable generation with Envision's AI Power System integrating generation, dedicated transmission and large storage.
• A designed end state of one million PFLOPS and support for up to one million AI accelerators in one structure.
• A quote from Ricky Zheng, general manager of Envision's AIDC business, that supporting a million accelerators requires purpose-built green power delivering far more compute per square meter than conventional sites.

The same release ties Galaxy to Mission Gobi, unveiled at VivaTech Paris in June 2026, targeting 5GW of green AI computing in arid regions by 2030.

That framing matters for readers who only follow chip exports: this operator's moat is vertical integration from turbine blades near the fence line to rack power — closer to power-first entrants in the West than to a pure colocation landlord.

Wind, storage and why Ulanqab

A bridge between two regions suggests linking remote wind generation to a distant AI campus load

SemiAnalysis's November 2026 visit post (paywalled in full, but the opening is public) situates Ulanqab as a windy "West Texas of China" without the same anti-datacenter backlash seen in some US states. For Galaxy, the newsletter reports Envision aiming to supply more than 60% of early-phase power from its own wind farms, using:

1. Siting on high-capacity-factor sites even when they sit hours from campus.
2. Oversized wind installs — about 420MW turbine capacity against roughly 240MW datacenter load in the example given.
3. Four-hour battery storage shifting windy hours into calm periods, on software Envision also builds.

Inner Mongolia's share of China's usable wind resources shows up in both SemiAnalysis national charts and this campus story. It connects to wider grid-stress reporting such as PJM wholesale price spikes when data centers load the grid, except here the operator owns both sides of the meter.

DatacenterDynamics adds that Envision claims about 80% renewable electricity at Galaxy, consistent with the corporate wind-turbine heritage (the firm cites 100GW of turbines installed lifetime, 20GW globally in 2025 on that article).

800V DC and the Dubhe controller

US hyperscalers discuss 800V DC as a future efficiency step; Envision says it is running workloads now. DatacenterDynamics reports solid-state transformers stepping 10kV AC down to an 800V DC bus, with batteries and voltage conditioning gear for AI racks.

Envision names the orchestration stack Dubhe, backed by a foundation model dubbed Dubhe Energy, coordinating transformers and batteries against sudden AI load swings. SemiAnalysis's tour notes a 5MW hall demonstrating the architecture, with fewer conversion stages than traditional UPS/rectifier chains — the same theme as Open Compute Project white papers on 800V distribution, but with SSTs on the input.

For builders, the lesson is operational: power electronics software is now part of cluster stability, not only job schedulers. That parallels NVIDIA's push on data-center management planes covered in NVKE preview guides, except here the vendor is the utility integrator.

Built-to-suit tenants, not Envision training models

Bloomberg's October 2026 interview coverage, relayed by Energy Connects, stresses a split role: Envision builds power and shells; technology firms install accelerators under lease. Ricky Zheng told Bloomberg TV the company sees strong domestic demand with offtake lined up through 2029–2030, and that the full Galaxy campus could finish around end of 2028 at up to 3GW.

Caixin's November 5, 2026 depth piece narrows the near term: 200MW phase one handed to a Chinese internet giant in late October 2026, and 200MW phase two expected in Q2 2027. General manager Zheng Zihao (Caixin romanization) leads the AIDC unit in that story.

That model differs from ByteDance-or Alibaba-owned halls in the national GW census, but it still adds AI shell and power in the same province SemiAnalysis already mapped.

How to read "800MW" next to "3GW"

Readers hit three numbers in one headline chain:

1. ~800MW IT — SemiAnalysis campus planning figure, with ~120MW delivered quickly.
2. 2GW+ — Envision's August press language for scalable design.
3. 3GW — Bloomberg/Caixin full-campus completion framing.

None contradict the others if you separate commissioned IT load today, designed campus IT ceiling, and total power infrastructure nameplate Envision discusses for Mission Gobi. The mistake is comparing Galaxy's 3GW press cap directly to useful Claude-class tokens nationwide; our China GW explainer still applies — chips and utilization gate effective AI compute.

A tilted solar panel beside green compute imagery reflects mixed wind and solar siting for desert AI campuses

Mission Gobi and the green-AI branding wave

Mission Gobi is Envision's umbrella for desert and arid-region green AI. Galaxy is the first flagship. The pitch blends:

• Cheap renewable energy geography (Inner Mongolia wind).
• Land near existing digital infrastructure hubs (Caixin and Envision both cite Ulanqab's network role).
• On-site manufacturing symmetries (SemiAnalysis photographed blade stacks beside the campus).

It sits in the same conversation as Let's Encrypt's confidential-computing goals for GPU trust and renewable siting fights in the US, but with a state-backed build corridor that keeps expanding.

What builders and policymakers should watch

For infrastructure planners: vertically integrated wind-to-rack operators change who signs the PPA and who owns the SST vendor relationship — procurement templates written for Equinix-style landlords do not transfer verbatim.

For AI buyers: leased halls still mean you bring GPUs; Envision's one-million-accelerator headline is a design envelope, not a product SKU.

For energy analysts: pairing 420MW wind with 240MW hall load plus storage is a concrete case study for matching variable generation to bursting training jobs, more tangible than lease spreads in Abilene postmortems.

For China-US comparisons: SemiAnalysis claims Galaxy's early phase landed at roughly a fifth of US costs for comparable scale — a contested but widely quoted ratio that will draw teardown pieces as more phases light up.

Limits on what we can verify from outside

• Envision's English press page is thin in automated fetches; rely on dated press copy plus corroborating trade outlets.
• SemiAnalysis tenant names for the "two internet giants" are partly paywalled.
• Accelerator types running in the 800V hall are undisclosed in public DCD copy.
• Effective FLOPS and token throughput are design marketing until third-party benchmarks land.

Related reading

• China 24 GW vs US 56 GW SemiAnalysis
• China underwater data center fact check
• Data center environmental impact
• PJM data centers and wholesale power spikes
• SpaceX Starbase compute siting
• NVIDIA NVKE management plane

Comparison with other renewable AI siting stories

Galaxy is not China's only experiment in pairing renewables with compute-dense halls, but it is unusual in how many layers one company owns. SemiAnalysis compares the first 120MW delivery timeline to US mega-campus schedules and cost baselines; whether that ratio holds on phase two is an open question once accelerator densities rise.

The underwater Shanghai project explainx.ai fact-checked in October leaned on seawater cooling and offshore wind narratives. Galaxy is land-based in Ulanqab, emphasizes SST-native 800V distribution, and foregrounds tenant internet firms rather than a single cloud operator brand. Both stories feed the same national GW charts, but the operational playbooks differ.

In the US, siting fights show up in zoning dockets and bond referenda — for example the public notice trail around SpaceX Starbase supercompute expansion. Inner Mongolia's posture, as described in SemiAnalysis's visit framing, is closer to acceleration: blades stacked within sight of the data hall, dedicated lines, and political economy that treats AI infrastructure as industrial policy.

For readers planning workshops on where inference will be cheap in 2028, Galaxy is a concrete anchor: renewable megawatts are necessary but not sufficient; storage software, DC distribution, and who owns the shell determine whether cheap wind turns into cheap tokens.

Update watchlist

Worth revisiting when new primary sources land:

• Named tenants and accelerator generations in each commissioned phase.
• Independent power and utilization metering once halls are fully loaded.
• Mission Gobi announcements outside China — Envision pitched a global program at VivaTech; second sites would test portability of the model.
• Regulatory or grid rules if export-class chips fill a larger fraction of racks than domestic accelerators.

Figures and dates follow Envision, DatacenterDynamics, Caixin Global, Energy Connects and the public portion of SemiAnalysis as accessed November 8, 2026.