A live 5MW hall at Envision’s Galaxy AI campus in Ulanqab, Inner Mongolia, already runs 800V DC power all the way to the rack, according to a site-visit report SemiAnalysis published on November 7. That is the architecture Nvidia says will reach full-scale production alongside its Kyber rack systems in 2027.
By our arithmetic, Galaxy’s 800V is 3.3 times the 240V DC that China Telecom piloted in 2007, and the 5MW pilot is about 4% of the campus’s 120MW first phase.
An evergreen Business Engineer explainer on vertical integration. It teaches the general idea only and does not discuss Envision or Galaxy.
The key insight: SemiAnalysis says a live 5MW hall at Envision’s Galaxy AI campus in Ulanqab already runs 800V DC to the rack, with solid-state transformers taking 10kV AC straight to 800V DC. Nvidia says full-scale 800 VDC data centres will arrive with its Kyber racks in 2027. By our arithmetic, Galaxy’s 800V is 3.3 times China Telecom’s 2007 DC pilots, and the pilot is about 4% of the 120MW first phase.
The report says that in one 5MW hall, “Solid-state transformers (SSTs) take 10kV AC from the grid and convert it straight to 800V DC for the racks”. It says this “removes the low-voltage transformers, UPS and rectifier stages in between”.
SemiAnalysis writes that “The same architecture is still on the US roadmap”. It says Galaxy’s pilot “already runs 800V DC all the way to the rack at a live hyperscaler site”, and that a larger version is planned for the next phase. Our first piece on the same report covered Galaxy’s 120MW build in under a year at a fifth of US costs.
The conventional AC power chain as Nvidia’s 800 VDC blog post draws it (transformer to 415V AC, UPS, rack power supplies, 54V/12V DC) against Galaxy’s live 5MW hall, where solid-state transformers take 10kV AC straight to 800V DC for the racks. Simplified, not to scale. Data: SemiAnalysis (7 Nov 2026); Nvidia (20 May 2025). Our arithmetic: 5MW / 120MW = about 4%.
Nvidia’s Own 800V Plan
In a technical blog post dated May 20, 2025, Nvidia says it is “leading the transition to 800 VDC data center power infrastructure to support 1 MW IT racks and beyond, starting in 2027”. It says today’s racks rely on 54 VDC in-rack distribution, which “begins to hit physical limits” as racks exceed 200 kilowatts.
Nvidia says switching from 415 VAC to 800 VDC “enables 85% more power to be transmitted through the same conductor size” and cuts copper requirements by 45%. Its design converts 13.8 kV grid power to 800 VDC at the data centre perimeter, and the post says Nvidia and partners are studying “both traditional transformer-based and solid-state transformer (SST) approaches”.
The same post says “Full-scale production of 800 VDC data centers will coincide with NVIDIA Kyber rack-scale systems in 2027”. SemiAnalysis says Kyber racks at about 600kW “cannot be served efficiently at today’s voltages”. Our earlier piece covered the Kyber timeline and Rubin Ultra.
Two Decades of DC Power in China
The report places Galaxy on a long path. It says Chinese datacenters “have been running on DC power for almost two decades, from China Telecom’s 240V DC pilots in 2007 to Alibaba’s Panama power system in 2018”, and calls an SST “a natural next step on that path”.
No Central Chiller Plant
On cooling, the report says liquid- and air-cooled servers at Galaxy share one water loop, “with no central chiller plant”. It notes that Nvidia says Vera Rubin can run on 45°C water without chillers, “yet most datacenters in the US keep one for air-cooled equipment and for the flexibility to shift between the two”.
SemiAnalysis says Galaxy drops the central chiller while keeping much of that flexibility, “a design that has been running at scale in northern China for years”.
The Integrator Behind the Pilot
The report says Envision, a renewable systems integrator, designed and integrated the low-voltage power modules and the 800V SST system for Galaxy. Its affiliate AESC supplied the battery storage, and Envision co-designed a prefabricated substation that the report says takes less than 3 months to deliver, versus about 6 months for a conventional indoor substation.
It adds that most vendors cover only one part of the power system, so on-site commissioning can take 3 to 6 months. It names a similar tie-up between VNET, CATL and Zhongheng Electric, which it calls China’s largest supplier of HVDC power systems for datacenters.
The Business Lever: The Power Chain as a Product
As we read the report, the competition here is over the power chain from the grid connection to the rack, not only over chips. SemiAnalysis says that power architecture “is being redesigned” and that the shift gives integrated players room to compete differently from traditional developers. Power has run through our recent coverage, from Anthropic’s 1.7GW of gas generators to the 2.2GW that 1.2M Blackwells would need.
The Structural Read
As we read the report, China’s head start on DC power is a path, not a single pilot: the report traces it from 240V DC in 2007 to Alibaba’s Panama system in 2018 to an SST at Galaxy.
As we read Nvidia’s post, Nvidia is still studying both transformer-based and solid-state transformer approaches for 800 VDC, while Galaxy already runs an SST version in a live hall.
As we read both sources, the scale is very different: Galaxy’s pilot is 5MW, while Nvidia’s plan is written for racks of 1 MW and beyond.
SemiAnalysis, We Visited China’s Stargate in Inner Mongolia
“Many questions still open on the US roadmap find a reference point here.”
Three Implications
FOR US POWER SUPPLIERS Nvidia names Delta, Eaton, Schneider Electric and Vertiv among the companies in its 800 VDC electrical ecosystem.
FOR DATA CENTRE BUILDERS SemiAnalysis says Envision’s prefabricated substation takes less than 3 months, versus about 6 months for a conventional indoor substation.
FOR THE NUMBERS TO WATCH SemiAnalysis says a larger version of the 800V pilot is planned for Galaxy’s next phase.
The essay says “The structural bottleneck isn’t computing, it’s electrons.”
As we read it, Galaxy’s 800V pilot is a bet on delivering those electrons with fewer stages between the grid and the rack. The essay does not discuss Galaxy or Envision.
What Is Not Established
The 5MW pilot is a small share of Galaxy, and the free section of the report gives no efficiency, cost or uptime figures for it; those details are behind SemiAnalysis’s paywall, which we did not read. The 2027 date is Nvidia’s own plan. The free section of the report does not say which servers or chips run in the 800V hall.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
SemiAnalysis says a 5MW hall at Envision’s Galaxy campus already runs 800V DC to the rack through solid-state transformers, with no central chiller plant on the campus. Nvidia says full-scale 800 VDC data centres will arrive with its Kyber racks in 2027. By our arithmetic, Galaxy’s voltage is 3.3 times China Telecom’s 2007 DC pilots.
A note on sourcing. Details of the Galaxy campus, its 5MW 800V DC hall and its cooling are from the free section of SemiAnalysis’s report We Visited China’s Stargate in Inner Mongolia (7 November 2026), read on 8 November 2026; we did not read the paywalled section. Nvidia’s plan and figures are from its technical blog post of 20 May 2025. The 3.3 times and about 4% figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
