A visit by a Silicon Valley research institute to China's 'Stargate' project highlighted the speed, cost and system-level capabilities of the AI infrastructure buildout. At the Xinghe base, Envision uses solid-state transformers to convert 10 kV AC directly to 800V DC for server racks, eliminating low-voltage transformers, UPS and separate rectifiers; it is described as the world's first commercially operating integrated SST direct-to-800V HVDC supply system in a large AI data center.
Overview: A Field Report From Ulanqab
In September 2026, researchers from SemiAnalysis, a US-based semiconductor and AI infrastructure research firm, traveled to Ulanqab in Inner Mongolia to inspect the Xinghe base, a large computing campus designed and developed by Envision Group. On November 7, the team published a report titled “We Visited China’s Stargate in Inner Mongolia,” framing the site as China’s counterpart to the flagship US Stargate program and arguing that many open questions in the American compute roadmap “can find their answers here.”
The report compares Inner Mongolia, with its abundant land and energy resources, to a Chinese “West Texas,” and benchmarks the Xinghe base against the Texas Stargate project in Abilene. Its conclusions center on three dimensions: construction speed, capital cost, and — most consequentially — system-level engineering capability that extends from wind farms down to GPU racks.
Scale and Delivery Speed at the Xinghe Base
The Xinghe base is being built in four phases. Phase one, a 120MW facility, was delivered in August 2026, less than a year after construction began. It is the first large campus within Envision’s broader plan for roughly 3GW of data center capacity in Ulanqab.
SemiAnalysis estimates that the unit construction cost of the phase-one facility is approximately one-fifth that of a comparable US project. The campus is also paired with planned wind power and energy storage to raise the share of green electricity consumed on site.
Ulanqab is not a new entrant to the data center business. Huawei established a presence there around 2013, followed by Alibaba, Apple and Kuaishou. Those early facilities primarily handled storage, backup and general-purpose cloud workloads; the generative AI boom has since driven a new wave of large-scale compute investment, elevating the value of the region’s wind resources, low average temperatures and proximity to Beijing.
Technical Details: Solid-State Transformers and 800V DC
The most technically significant element of the visit is the base’s power architecture. Envision uses solid-state transformers (SST) to convert 10kV alternating current directly into 800V direct current, which is then delivered to server racks, where onboard DC/DC modules step the voltage down further.
According to the article, this is the world’s first integrated SST direct-to-800V high-voltage DC supply system operating commercially in a large AI data center. SemiAnalysis’s on-site inspection confirmed the system is running in a live machine room, while comparable US architectures remain at an earlier deployment stage.
The design eliminates the low-voltage transformers, uninterruptible power supplies and separate rectifiers that characterize conventional AC distribution. Those intermediate stages introduce conversion losses and occupy floor space — constraints that intensify as rack density climbs. Nvidia’s planned Kyber architecture, for instance, is expected to reach roughly 600kW per rack, pushing power delivery and conversion efficiency to the center of data center design.
From Architecture to Operations: Fault Isolation and Load Dynamics
Proposing an architecture and operating it commercially are distinct challenges. AI server loads fluctuate rapidly with compute tasks, and faults must be isolated quickly so that failures do not propagate to other running servers.
Envision addressed this by co-designing the SST, storage, protection devices and control policies from the outset, predefining fault-isolation boundaries, supply guarantees for critical loads and recovery sequences. A drop in DC bus voltage, for example, may stem either from a sudden GPU load surge — in which case storage must inject power promptly — or from a short circuit on a branch, which requires protection devices to trip and prevent other sources from feeding the fault. Operational data from the Xinghe base feeds back into product design and system validation.
System Integration and the Vertical-Integration Advantage
Beyond the machine room, Envision extends energy management to wind farms, transmission lines and storage. The base plans two wind farms roughly 80 km apart, sited to exploit differing wind conditions and reduce the probability of simultaneous low-output periods. Combined with a four-hour battery storage system, planned green power coverage exceeds 80%.
Envision’s weather forecasting system predicts wind output in advance, while its energy management system dynamically coordinates generation, storage state and compute load. On the construction side, prefabricated substations are assembled and tested in factories before shipment, compressing delivery to under three months versus roughly six months for conventional indoor substations. Cooling leverages Ulanqab’s cold climate to reduce reliance on mechanical refrigeration.
SemiAnalysis described Envision as one of the most vertically integrated data center participants its team has encountered.
Market Context: US Headwinds Versus Chinese Buildout
The report lands amid mounting friction in the United States. Bloomberg reported on November 5 that Dutch group ING is seeking to sell roughly $500 million in data center loan assets, some tied to Stargate financing. In September, Oracle issued a force majeure notice concerning a Stargate data center in New Mexico to address potential delay risk.
Local opposition is also sharpening. Ballotpedia recorded 41 data center-related local ballot measures on November 3, of which 40 leaned restrictive: 27 restrictive measures passed and 13 pro-build proposals were rejected. In Texas, grid expansion and interconnection have become bottlenecks, prompting some projects to add gas turbines, fuel cells and storage for supply security.
Implications for AI Data Centers and Future Outlook
The contrast between the two Stargates reflects divergent buildout paths. The US model leverages capital, chips and cloud ecosystems while power supply increasingly constrains schedules. The Xinghe model integrates renewable generation, power equipment and engineering manufacturing into data center construction, designing from wind farm to GPU rack as one system.
The strategic implication is that competition in AI infrastructure is extending into energy. As compute demand grows, the ability to convert each kilowatt-hour into more compute and more tokens becomes a determinant of cost structure. The Xinghe base demonstrates a commercially operating engineering template for pairing green power with high-density compute — and, in doing so, offers a reference point for building next-generation AI infrastructure faster, cheaper and more energy-efficiently.