U.S. Research Firm's On-Site Visit: Envision Xinghe AI Base Construction Cost Only One-Fifth of Texas Stargate
nashnova research
Semiconductor research firm SemiAnalysis visited Envision Group's Starfield AI base in Inner Mongolia and found its per-unit-capacity construction cost is roughly one-fifth of the Texas Stargate — the gap traces back to Envision's fully integrated power chain from wind turbine to server rack.
What has the Starfield base actually built?
Starfield Phase 1 has an installed capacity of 120 MW, located in Ulanqab, Inner Mongolia, developed by Envision Group.
Ground-breaking to delivery took under one year — roughly matching the Texas Stargate's build pace.
This means → China can now match top-tier US projects on construction speed for large-scale AI compute facilities.
What exactly is the Texas Stargate it's being compared to?
Texas Stargate is developed by AI infrastructure provider Crusoe, leased to Oracle to supply compute for OpenAI. It plans 8 buildings with ~1.2 GW total power capacity.
Once fully built, its target IT capacity is close to Starfield's.
In plain terms = both projects aim for similar total compute, but Starfield Phase 1 is already operational — at one-fifth the unit cost.
A 5× cost gap — where does it come from?
The core difference is power architecture. Envision designed wind farms, dedicated transmission lines, energy storage, and data-center halls as a single integrated system from the planning stage — turbine to rack, end to end.
Texas Stargate relies mainly on the public grid, supplemented by on-site natural-gas turbines — its power system is procured, not self-built.
This means → Envision collapsed "generation → transmission → storage → distribution → consumption" into one line, cutting interface costs and redundancy at every link.
Why does 800V DC power deserve its own section?
Starfield has put an 800-volt DC power system into commercial operation, using solid-state transformers — devices that replace traditional iron-core transformers with semiconductor components.
SemiAnalysis says this is the world's first commercial deployment of this architecture in a large-scale AI data center.
In plain terms = conventional data centers use AC power that must be converted multiple times before reaching the server, losing energy at each step. 800V DC is a "direct expressway" — fewer conversions mean less waste, less equipment, less floor space.
How do Western institutions read this?
SemiAnalysis described Envision as "one of the most vertically integrated data-center players" its research team has ever seen.
Goldman Sachs flagged the rapid rise of the Ulanqab compute cluster in an August 2026 report; SemiAnalysis's on-site visit now adds finer-grained engineering and power-system detail.
This reflects a widening recognition that the gap in power integration and build cost is becoming a key variable in measuring the US-China AI infrastructure race.
市场有风险,内容仅供研究参考,不构成投资建议。
