By Ma Shiqing

Edited by Ma Xiaoning

SpaceX’s IPO drew global attention, not only because it set a record for the world’s largest first-round fundraising at $75 billion, but also because the company became the world’s fourth-largest listed company on its third trading day, pushing Elon Musk’s net worth into the trillion-dollar range.

The prospectus from this capital-market super entrant pointed to a stark reality: more than 90% of its future potential market will come from AI. More importantly, AI development is being severely constrained by shortages of electricity on Earth.

Musk’s answer is to build AI data centers in space, using free solar power beyond Earth’s atmosphere to break away from terrestrial power grids entirely.

It is an enormous ambition, but achieving it looks as distant as sending humans to Mars. Capital markets have also started to cool. Two days ago, SpaceX’s market value fell by $400 billion, the second-largest single-day market-cap loss in U.S. stock market history. To solve the AI power shortage already unfolding today, a Chinese company may have moved ahead of SpaceX long ago.

That company is Envision Group, which is now pushing hard to build an AI power system: artificial intelligence infrastructure that integrates energy systems with intelligent systems.

Not long ago, China’s National Energy Administration held an “AI+” energy field promotion meeting. Zhang Lei, chairman of Envision Group, said in his remarks that the power system is becoming the main engineering foundation for AI, and that only by solving energy management across the full chain of intelligent production can the world provide a continuous power supply for AI as a new industrial revolution.

##

AI Calls for a New Type of Power System

AI data centers are, quite literally, power-hungry machines. Data from the International Energy Agency (IEA) show that by 2030, electricity consumption by AI data centers will reach 950 terawatt-hours, roughly double the 485 terawatt-hours expected in 2025 and close to Japan’s total annual power generation.

Image: Electricity consumption by AI data centers from 2020 to 2035, in terawatt-hours

Source: International Energy Agency (IEA)

As electricity demand surges, AI data centers have even begun competing with households for power. In Virginia and Texas, two U.S. states with heavy concentrations of AI data centers, outage durations are rising, and electricity prices in some areas have increased by more than 12% year on year per kilowatt-hour.

At the start of this year, U.S. President Donald Trump convened executives from seven major tech companies, including Microsoft, Google, Amazon, Meta and OpenAI, at the White House to sign related agreements requiring tech giants to bear the power costs of AI data centers themselves and avoid passing them on to ordinary Americans.

Worse still, power grids in Europe and the United States are falling into an aging crisis. Main transmission lines have been operating for an average of 40 to 50 years, and if generation increases, the grid may simply be unable to support it. The Electric Power Research Institute (EPRI) put it bluntly: the grid and earlier policies were not designed for AI infrastructure, and today’s power system has physical bottlenecks.

Rebuilding the grid would not solve the timing problem, either: grid construction cannot keep pace with the expansion of computing power. The construction cycle for AI data centers has already been compressed to several months to about a year, while traditional grid projects, from planning and approval to land acquisition, construction and commissioning, take at least five to 10 years.

The challenge is not limited to Europe and the United States. It also exists in China, despite the country’s strong grid-dispatch capabilities. In the past, AI data centers were densely concentrated in the economically developed but power-constrained east. Although China’s national East Data West Computing strategy has sought to guide workloads westward, long-distance transmission and distribution bring physical losses, while cross-region transport costs also weaken the overall cost-performance equation.

In addition, under the hard constraints of China’s “dual carbon” goals, thermal power generation is no longer sustainable. New national rules make clear that newly built AI data centers at national hub nodes must source no less than 80% of their electricity from green power.

If the grid cannot be relied on, thermal power is unsustainable, and rebuilding grids or transmitting electricity over long distances remains costly, where is the path forward for AI development?

SpaceX’s answer is to rebuild a power system away from the ground: put AI data centers into satellite orbit and let them directly absorb solar energy. Space-based solar generation could exceed that of ground-based arrays of the same area by more than five times, while entirely avoiding competition with residential electricity use.

Image: Design blueprint for the “AI1 satellite”

Source: SpaceX

Yet building AI data centers in space may not be the most rational commercial choice. According to the AI1 satellite design blueprint released by SpaceX, photovoltaic panels used in space face higher requirements for materials, structural reliability and other factors. Beyond the cost of obtaining solar energy, there are also satellite R&D and manufacturing costs, launch costs, space-based cooling costs, chip R&D and manufacturing costs for space use, operating costs and more.

According to SemiAnalysis estimates, in 2026 the cost gap between space-based and ground-based data centers will exceed fourfold, and the two curves will not reach an inflection point until around 2040. Even if Musk’s plan advances aggressively and succeeds, costs would not reach parity until after 2030.

Source: SemiAnalysis, “To Boldly Go: The Case for Space Datacenters”

At Viva Tech, the European technology gathering held in June, French President Emmanuel Macron and technology leaders from around the world discussed AI. Amazon’s Jeff Bezos and Alibaba’s Joe Tsai were also in attendance.

At one forum, addressing the global challenge of AI development and power shortages, Envision Group chairman Zhang Lei announced Envision’s “Mission Gobi” plan. Known in the industry as GobiX, the plan will use the Gobi’s abundant wind and solar resources to generate power and build 5 GW of green AI computing power centers across Gobi and desert regions worldwide by 2030.

Image: Zhang Lei, chairman of Envision Group, at Viva Tech

Zhang Lei said: “As the storm of AI computing power sweeps the world, traditional power grids are no longer able to carry this wave of change. Starting from the Gobi, Envision is taking AI power systems global, offering a Chinese solution to the power bottleneck facing the age of artificial intelligence.”

Unlike SpaceX, which aims to build a new power system in space, how will Envision’s AI power system support its Mission Gobi?

##

A Green Power Bonanza Hidden Deep in the Gobi

In conventional thinking, the Gobi means drought and barrenness. Today, it is being viewed as an energy bonanza.

The Gobi region has intense sunlight, and its surface is mostly gravel and sand with no vegetation blocking the light, making it a natural site for large ground-mounted photovoltaic bases. The region also has strong year-round winds and extremely high wind energy density. According to institutional estimates, if 1% of China’s desertified land were used for new energy generation, the resulting installed capacity would exceed China’s current total installed power generation capacity.

Building AI data centers directly where wind and solar power are produced not only provides local green electricity at extremely low cost, but also avoids, at the source, the risk of competing with urban residents for power.

Still, renewable power generation is inherently intermittent and volatile. Wind power fluctuates with wind strength, while photovoltaic generation depends on sunlight hours and cannot be dispatched at will. How can unstable wind and solar resources be converted into the round-the-clock, highly stable power that AI data centers require?

Envision’s approach is to treat renewable generation, energy storage, power grids, power electronics, computing power and large models as an integrated system and design them systematically.

To do this, Envision has built an “AI power system” that provides a full-process, system-level solution. Through digital mapping of the physical world, it ensures that every kilowatt-hour is generated at the optimal cost and consumed at the most suitable node in the most efficient way. The system uses a three-layer “software-hardware integrated” technology architecture:

The first layer is the brain of the AI power system, solving questions of power generation and cost. Envision uses its “Tianji” meteorological large model to forecast weather, helping wind turbines and photovoltaic panels generate power more precisely, increasing output while avoiding waste. It can also help storage systems charge and discharge at the right moments, earning spreads in spot power markets through peak shaving and valley filling. In addition, the “Tianshu” energy large model can understand demand across every link of power transmission in real time and find the global optimum.

Image: Envision wind turbine

The second layer is the system’s neural network, ensuring real-time coordination across each link. Envision’s self-developed EnOS intelligent IoT operating system has connected hundreds of millions of smart devices, including wind power, photovoltaic, energy storage, transformers, hydrogen electrolyzers and other equipment, ensuring the entire system operates efficiently in real time.

The third layer is the system’s skeleton, supporting the stable operation of computing power centers. Envision relies on a series of self-developed hardware products, including integrated wind-solar-storage controllers, high-voltage direct current (HVDC), solid-state transformers (SST) and smart cabinets, to sharply increase power density while reducing energy losses.

Envision has changed the fragmented structure of traditional power systems. From a higher-level perspective, it is reconstructing how power systems operate, enabling real-time coordination among source, grid, load, storage and computing, and helping AI data centers balance large electricity demand, high stability and low cost.

##

Inner Mongolia’s Prototype Goes Global

Between a vision blueprint and a project on the ground, there are often many uncertainties. Musk has proposed many ideas and plans, from the full rollout of Tesla FSD and mass production of the Optimus humanoid robot to the ramp-up of 4680 battery capacity, but actual progress has often lagged his public commitments by years.

So when Musk says SpaceX will deploy a 1 GW orbital AI data center in space by the end of 2027, even many fans cannot help wondering: is this another case of “Elon Time” that will require a long wait?

Compared with plans in space that remain to be verified, the model being tested by Chinese companies in the Gobi of Inner Mongolia has already proven workable.

In Chifeng, Inner Mongolia, Envision and Tencent have jointly built the world’s first system-level model of “computing-power and electricity coordination,” as well as the world’s first AI data center directly powered by 100% green electricity. Chifeng, a clean energy hub in Inner Mongolia, has exceptional conditions for wind, solar and storage.

Through its AI power system, Envision dynamically integrates each link of power transmission, enabling 100% green power supply and low-carbon production. The project has cut comprehensive energy costs by more than 40% and reduces carbon emissions by 180,000 tons a year.

Image: Chifeng Envision x Tencent green-power direct-supply AI data center

Also in Chifeng, Envision has launched the world’s largest green hydrogen-ammonia project. Using the same AI power system, it achieves intelligent coordination across the full process of wind and solar generation, electrolytic hydrogen production, air separation for nitrogen production and ammonia synthesis. The system can complete wide-range capacity adjustment within five minutes, precisely matching wind and solar output.

In Ulanqab, Inner Mongolia, another GW-scale integrated energy and computing power infrastructure project, the Envision Galaxy Base, has also taken shape. From wind and solar generation forecasting to millisecond-level responses from energy storage systems and task orchestration for computing power clusters, the entire loop is completed within the same AI power system.

Envision packages its AI power system and AI data centers together inside zero-carbon industrial parks. This scenario has clear physical and management boundaries, making software and hardware easier to standardize and modularize, and providing the basis for replicating the Chifeng model across regions.

Zooming out from Inner Mongolia to the global map, it becomes clear that Gobi and desert resources still have room for development. The future picture may look like this: from Africa’s Sahara Desert to the Arabian Desert in the Middle East, from Central Asia’s Kyzylkum Desert to the Taklamakan in northwest China, and from North America’s Great Basin Desert to South America’s Patagonian Desert, Envision’s AI power systems could be seen supporting data center operations.

##

Conclusion

Hundreds of years ago, in the age of the steam engine, coal was converted into physical power and drove the first industrial revolution. In today’s AI era, GPUs have become the steam engines of the present day. Electricity flows into GPUs and, through training and inference, ultimately outputs cognition, decisions and human intelligence. AI’s development is driven not only by models and chips, but also by electricity. The endpoint of AI is electricity, and the process of AI is electricity as well.

If Musk’s SpaceX is racing into deep space in an attempt to push the boundaries of human civilization in the cosmos, then Envision, a Chinese company rooted in the Gobi, is also expanding those boundaries by using AI power systems to give green electricity intelligence.

The difference is that Musk’s space narrative carries some people’s imagination of the universe. But in today’s increasingly intense battle for electricity, what matters more is an answer that can actually be deployed. As Mission Gobi advances, the prototype Envision has proven in Inner Mongolia will be replicated across Gobi regions worldwide, and lands once seen as barren on Earth may become the cradle of the next generation of intelligent civilization.

Click to enter: a global hub for highlights from leading AI academic conferences

After visiting 30-plus Chinese companies at ICRA, we found three major trends in the embodied intelligence industry | ICRA 2026

Reposting this content on websites, forums or communities in any form is strictly prohibited without authorization from AI Technology Review.

For WeChat public account reposts, please first leave a message in the AI Technology Review backend to obtain authorization. Reposts must cite the source and include this account’s card.