The computing power cost war has reached the power distribution room: Envision's 800V DC system has achieved commercial operation.
Envision's world's first commercial 800V DC power supply system for AI data centers has been officially launched.
Envision Energy learned that Envision recently announced in Ulanqab that the 800V high-voltage DC power supply system at Envision Ulanqab Xinghe Base has achieved end-to-end connectivity: 10kV power enters the solid-state transformer (SST), and the output 800V DC is directly fed into the cabinets, eliminating the power frequency transformer, AC UPS, and the AC-to-DC conversion links in server power supplies all at once. This is the world's first commercial system of its kind deployed in a large-scale AI data center.
The industry has not yet reached a consensus on the 800V route: Nvidia advocates adopting 800V in one step, but its supporting cabinet power supplies are scheduled to go into mass production only in 2027. Google, Microsoft and Meta are promoting ±400V in the Open Compute Project (OCP), keeping the voltage at the 400V level first. The two factions seem to be arguing over voltage levels, but in terms of the park's financial statement, it is essentially the same core concern.
The cost focus of AI parks is shifting from the procurement price of GPUs to the cost at which the power supply system can deliver a certain amount of electricity into the cabinets. The successful power connection in the grassland is an early submission of a satisfactory answer.
01
Power Saving, Copper Saving, Space Saving
The Three Goals Are Hard to Achieve Simultaneously
When this calculation is applied to equipment, three indicators come into play: conversion efficiency, copper consumption, and land occupation. Low efficiency means electricity costs eat up all profits; excessive copper consumption leads to high cable costs and difficult construction; since computer rooms are priced by square meter, if power supply equipment takes up one extra room, dozens of cabinets will have to be excluded.
The difficulty of the three indicators lies in that they cannot be optimized all at the same time. To achieve higher efficiency, more conversion stages or larger heat dissipation space are usually required; to save copper, the voltage has to be raised and the conductor cross-section has to be reduced, which puts higher requirements on the voltage resistance of components; to reduce land occupation, the equipment has to be compact, which squeezes the space for heat dissipation and maintenance. Most solutions in the industry can only optimize two of the three indicators.
The solution Envision provided for the first phase of Xinghe Base is to replace the iron core and copper coils in traditional transformers with silicon carbide components and high-frequency magnetic parts. 10kV AC is input and 800V DC is output, without passing through power frequency transformers and AC UPS in the middle.
The conversion efficiency of this solid-state transformer stage remains above 98.5%, and the land occupation of power supply equipment is reduced by more than half. The simultaneous improvement of efficiency and land occupation comes half from the elimination of two conversion stages, and the other half from the fact that the magnetic parts no longer need to be designed according to power frequency standards.
When these indicators are put into the local electricity price in Ulanqab, the differences become more tangible. The on-grid electricity price for local computing power projects is stably between 0.32 yuan/kWh and 0.36 yuan/kWh, and the comprehensive operating cost is 30% to 50% lower than that of first-tier eastern cities. For a million-kilowatt-level park, every one percentage point increase in efficiency can save tens of millions of yuan in electricity bills a year.
In terms of copper cost, to deliver the same 1kW of power, the lower the voltage, the thicker the wire has to be. According to the data provided by Nvidia, if 54V DC power distribution is used for 1000kW cabinets, the busbar inside the cabinet alone requires 200kg of copper; when the voltage is raised to 800V, conductors of the same cross-sectional area can deliver about 85% more power.
Existing computer rooms cannot benefit from this advantage, as the original power frequency transformers and AC UPS cannot be removed, and they can only add a set of rectifiers next to them to connect the 800V busbar, which reduces the efficiency gain. The first batch of 800V deployments can only be in newly built parks, which explains why the world's first such system appears in a newly commissioned grassland base.
02
The Power Grid Is Not Afraid of High Power Consumption
It Fears Sharp Fluctuations in Power Load
The power consumption of AI computer rooms is manageable for the power grid, but the real problem lies in its load fluctuation.
Thousands of GPUs draw power at the same time, causing the busbar voltage to drop, and the load suddenly drops to zero once the training is finished, which is equivalent to all air conditioners in a building turning on and off at the same time, repeating dozens of times a day.
This set of load is also connected to wind and solar power, whose supply side also fluctuates, leading to instability on both ends. The North American Electric Reliability Corporation issued its highest-level reliability alert so far this May, triggered by this type of computing load.
It recorded that a computing power center experienced an unexpected off-grid of more than 1000 MW, whose impact on grid frequency is comparable to the instantaneous tripping of a large generator set, and the off-grid speed is often faster than the response speed of dispatchers. The power grid has a high tolerance for total power consumption, but a very low tolerance for sharp fluctuations, which explains why it is increasingly difficult for new AI parks to obtain grid connection permits.
To stabilize this type of load, the power supply system needs to achieve several things at the same time: support the millisecond-level voltage drop, fill the second-level power gap, provide minute-level backup, and avoid passing the GPU's peak power directly to the power grid.
The trouble is that neither of the two existing solutions can do the whole job: energy storage can fill the gap at the millisecond level, but its battery capacity is limited and runs out in a few minutes; the solid-state transformer can continuously draw power from the grid, but its adjustment speed cannot keep up with the GPU's millisecond-level jumps. The controller needs to make real-time judgments on when to act, how much to adjust, and when to return the load, and wrong judgments will make even excessive energy storage configuration useless.
The verification completed at Xinghe covers exactly these points: the 10kV to 800V system remains stable under all working conditions, the energy storage can support both second-level power gaps and minute-level backup scenarios, and the solid-state transformer and energy storage cooperate at the millisecond level to absorb the GPU's fluctuations locally.
Envision sets the power supply reliability of the park at 99.995%, which translates to an allowable downtime of less than half an hour per year. Training tasks that run for dozens of days do not need to restart from scratch due to power supply interruptions.
This set of technologies was tested in Chifeng, hundreds of kilometers away, where a 2GW off-grid system runs entirely on wind and solar power, all equipment is power electronic devices, no synchronous generator is used to stabilize the system, and the voltage and frequency are maintained entirely by control strategies.
The conventional practice in the industry is to deploy excessive batteries to resist fluctuations, while Envision replaces part of the hardware with algorithms, cutting the required energy storage capacity by half compared with similar solutions, and increasing green ammonia production capacity by 15%. The Xinghe scenario is different, where the load is replaced by GPU training tasks that can hardly be interrupted, and grid power must be used as backup.
03
Relevant Policies Have Been Implemented
The Difficulty Lies in Aligning the Interests of Three Parties
"Collaboration between computing power and power" was included in the Government Work Report for the first time this year, and incorporated into the 15th Five-Year Plan Outline. In May, four departments including the National Development and Reform Commission and the National Energy Administration issued the Action Plan for Promoting the Two-Way Empowerment of Artificial Intelligence and Energy, which clearly encourages computing power facilities to configure grid-forming energy storage and improves the direct connection policy for green power.
The policy orientation is already very clear: to enable computing power to use green power, and turn computing power load into a regulable resource for the power grid. In actual implementation, the difficulty lies in coordination: power generation assets belong to power generation enterprises, energy storage is owned by equipment vendors, and the load is controlled by cloud vendors, each of the three parties has its own financial considerations.
Envision covers two of the three parts: wind turbines, photovoltaics, energy storage and control systems are all part of its product line, and the Xinghe Park is constructed and operated by Envision. The data from the power generation side and control side are integrated into the same system, and dispatching instructions do not need to be negotiated across companies, but even Envision can only adjust the load within the scope authorized by its customers.
Envision's approach is to, within the scope of customer authorization and delivery time limit, try to shift adjustable training tasks to periods with sufficient green power. This is the most realistic implementation standard for the "collaboration between computing power and power" policy.
The planned green power direct connection proportion of Xinghe Park exceeds 80%, and the green power proportion of the whole Ulanqab city reaches two thirds. This is the premise for Envision to test this system here. If the same solution is applied to a place with poor wind and solar conditions and high electricity prices, the financial calculation will not add up.
North America faces different bottlenecks: the lead time for ordering large transformers above 100MW is 2.5 to 3.5 years, and the average time from applying for power connection to grid connection for projects in the PJM region exceeds 7 years. The capacity price has risen from 28.92 USD/MW-day in the 2024/2025 period to 329.17 USD/MW-day in the 2026/2027 period.
When they cannot afford to wait, cloud vendors prefer to adopt solutions such as ±400V that can leverage the mature electric vehicle supply chain. 400V-level components have been mass-produced in the electric vehicle industry for 10 years, and the difficulty of insulation and arc protection is relatively lower.
The interface standards will most likely be set by organizations such as OCP, IEC and UL in the end. Chinese vendors can compete for equipment market share, where competition is fierce and profit margins are not high. What is really hard to replicate is the scheduling mechanism among green power, energy storage and computing power load, for which no standard parts are available, and each park needs to be designed separately according to its own wind and solar conditions and electricity price structure.
Envision sets its next milestone in 2030, planning to build 50GW of green computing power centers in global gobi regions. China's solid-state transformer track has already attracted dozens of vendors including TBEA, China XD Group, Sifang Co., Ltd. and Sungrow Power. 2026 is regarded as the first year of its industrialization, and most insiders judge that large-scale capacity expansion will not come until 2028.
The power consumption of computing power centers will rise from 170 billion kWh in 2025 to 800 billion kWh in 2030, an increase of more than 600 billion kWh in five years. This volume cannot be supported simply by installing more traditional transformers.
This system has not been in operation for very long, and whether it can be replicated to places with different wind and solar conditions and electricity price structures depends on subsequent operation records. The first set solves the problem of "whether it exists", while the 100th set will determine the actual value of this solution.
As Zhang Lei, Chairman of Envision Group, once said, the competition for artificial intelligence in the future is not only the competition for models and chips, but also the competition for power systems, and the competition for how much effective intelligence each megawatt of power can generate.
This article is from the WeChat Official Account "Foresee Energy", written by ZHAO Jianan, and published with authorization from 36Kr.