AI’s appetite for electricity is no longer a background concern—it’s becoming the main constraint that shapes where data centres can be built, how quickly they can be scaled, and what kind of sustainability claims are even credible. That shift is at the heart of a provocative argument from Zhang Lei, founder of Envision, a company best known for its role in clean-energy technology and power systems. In recent remarks, Zhang urged that if AI data centres are placed in desert regions, they should be powered by off-grid renewable energy rather than depending on conventional grid supply.
On the surface, the idea sounds almost obvious: deserts offer abundant sunlight, large tracts of land, and relatively low population density. But the devil is in the details. Desert sites are also far from existing transmission networks, often face extreme temperature swings, and require significant infrastructure to keep servers cool and operational. The grid, in many cases, is either too distant, too slow to expand, or too carbon-intensive to satisfy the growing demand for cleaner compute. Zhang’s position is essentially that the industry should stop treating energy as an afterthought and start designing AI infrastructure around power generation from day one—especially when the location itself is chosen for its environmental and logistical characteristics.
What makes this argument stand out is not just the preference for renewables, but the insistence on off-grid reliability. Many discussions about “green AI” focus on offsets, renewable energy certificates, or long-term procurement contracts that still rely on the grid. Zhang’s framing is different: he is pointing to a model where the data centre becomes part of a self-contained energy system—one that can generate electricity locally, store it, and deliver it with the stability that high-performance computing demands.
To understand why this matters, it helps to look at what AI workloads actually require. Training and inference are not simply “high power”; they are high power with strict performance expectations. Data centres need stable voltage and frequency, rapid response to load changes, and redundancy that prevents downtime from turning into lost revenue or stalled research. Even when the average consumption is manageable, peak demand and the need for continuous operation push facilities toward robust power architectures. That is precisely where grid dependence can become a bottleneck. If a desert site must wait for new transmission lines, transformers, substations, and grid upgrades, the timeline for scaling AI capacity can stretch far beyond the timeline for buying servers.
Zhang’s proposal implicitly challenges a common assumption in the AI build-out: that the energy system will simply expand to meet compute demand. In reality, energy systems have their own constraints—permitting, construction lead times, interconnection queues, and political negotiations. In many regions, grid expansion is slower than the pace at which investors want to deploy capital into data centres. Off-grid renewables, by contrast, can be planned and built in parallel with the facility itself, potentially reducing the “power gap” that has plagued earlier waves of infrastructure.
The desert angle adds another layer. Desert regions are often marketed as ideal for solar because of high irradiance and large available land. Yet solar alone is not enough for a data centre that runs 24/7. The key question becomes storage and dispatchability: how do you provide power during nights, dust storms, seasonal variations, and periods when solar output drops? Zhang’s argument doesn’t claim that renewables eliminate engineering complexity; rather, it suggests that the complexity should be solved locally, using a combination of generation and storage technologies designed for the site.
In practice, an off-grid renewable-powered data centre would likely involve a portfolio approach. Solar would probably be the anchor, given the geography. Wind could complement it in some desert climates, depending on local wind patterns. Storage—whether battery systems, thermal storage, or other emerging solutions—would smooth out variability and provide resilience during outages. Backup generation might still be part of the design, but the goal would be to minimize reliance on fossil fuels and reduce emissions intensity over the life of the facility. The “off-grid” concept here is not necessarily “no backup ever,” but rather “the primary energy comes from local renewables,” with backup used sparingly and strategically.
There is also the question of cooling. Data centres in deserts face heat management challenges that are different from those in temperate climates. Cooling systems must handle higher ambient temperatures and, in some cases, higher humidity variability. That affects both energy consumption and water use. A well-designed off-grid system could integrate cooling efficiency improvements—such as advanced air handling, liquid cooling strategies, evaporative cooling where feasible, or hybrid approaches—to reduce the total electricity required per unit of compute. In other words, the energy strategy and the thermal strategy are inseparable. If the facility is designed to run on local renewables, it becomes even more important to minimize waste and optimize every watt.
This is where Zhang’s argument becomes more than a slogan about “clean energy.” It points toward a systems engineering mindset: choose a location, then build the entire energy-and-cooling ecosystem around it. That includes not only generation and storage, but also demand management. AI workloads can sometimes be scheduled or shaped to align with energy availability, especially for certain training regimes or non-real-time inference tasks. Even when workloads cannot be shifted freely, data centres can implement power management layers that reduce peaks and improve overall efficiency. Off-grid designs tend to force these optimizations because the energy system is finite and must be balanced against the facility’s needs.
Another reason the off-grid approach is gaining attention is the carbon accounting problem. As AI expands, companies are under pressure to demonstrate that their compute is not merely “powered by renewables somewhere,” but that the electricity used is genuinely cleaner. Grid electricity can be a moving target: the carbon intensity of the grid varies by time and region, and it can change slowly relative to the speed at which new demand arrives. Off-grid systems can, in principle, offer clearer attribution—if the data centre is generating its own renewable electricity, the emissions profile is easier to control and verify.
Still, critics might argue that off-grid renewables are not automatically greener. Manufacturing solar panels and batteries has an environmental footprint, and large-scale deployment requires supply chains for critical materials. There are also land-use considerations and potential ecological impacts. Zhang’s stance does not erase these issues, but it reframes the debate: instead of relying on incremental grid greening that may lag behind AI growth, the industry could build dedicated renewable capacity that scales with compute demand. That could accelerate decarbonization in practice, even if it introduces new upstream challenges.
The financial logic is equally compelling. Data centre economics are sensitive to energy costs, but also to energy certainty. If a facility depends on grid power, its operating costs can fluctuate with electricity prices, demand charges, and policy changes. Moreover, grid dependence can create risk around curtailment or constraints during peak periods. Off-grid renewables, while requiring higher upfront capital for generation and storage, can offer more predictable operating costs once built. For investors, predictability can be as valuable as low cost—especially when AI projects are measured in years and decades, not quarters.
Zhang’s comments also land in a broader context: the AI infrastructure race is increasingly a competition over physical constraints. Chips are one constraint, but power is another. Water is another. Permitting and grid interconnection are another. When companies talk about “scaling AI,” they often focus on procurement and engineering, but the real bottlenecks are frequently external. Desert-based data centres, if they are to be viable, must solve those external constraints through integrated planning. Off-grid renewables are one way to do that.
There is also a strategic dimension. Desert regions are sometimes seen as politically and economically attractive for large infrastructure projects because land is available and development can be concentrated. But governments and communities will ask hard questions: Who benefits? What are the environmental impacts? How is water handled? What happens during extreme weather events? An off-grid renewable model could be positioned as a way to bring clean energy and jobs to remote areas, while reducing strain on existing grids. Yet it also concentrates responsibility on developers to deliver reliable power and manage environmental risks responsibly.
Reliability is the phrase that sits underneath everything in Zhang’s argument. AI data centres cannot afford frequent interruptions. Off-grid systems must therefore be engineered with redundancy and resilience. That means designing for component failures, maintaining storage capacity adequate for worst-case scenarios, and ensuring that the system can ride through transient events without shutting down. It also means planning for maintenance in harsh climates, where dust accumulation can reduce solar output and where equipment must withstand high temperatures and sand exposure.
Dust is not a minor detail. In desert environments, solar panels can lose efficiency due to soiling, and cleaning operations consume water and labor. A serious off-grid plan would need to incorporate cleaning strategies—possibly dry cleaning methods, optimized panel angles, coatings, or robotic maintenance systems—along with water-efficient practices. These operational realities affect both cost and sustainability. They also influence the size of the renewable generation needed to meet the data centre’s baseline demand.
Then there is the question of scalability. AI demand is not static; it grows. If a data centre is built with off-grid renewables, the energy system must scale alongside compute. That implies modular generation and storage expansions, and a design that anticipates future capacity additions. Otherwise, the facility could face a situation where servers can be added but power cannot. Zhang’s vision, if implemented, would likely require a long-term roadmap for energy capacity expansion, not just a one-time build.
One unique take on Zhang’s argument is to view it as a critique of “grid-first” thinking. The grid is a powerful asset, but it was not designed to absorb sudden surges of AI-driven demand concentrated in specific locations. In many places, the grid is already under stress from electrification trends—electric vehicles, heat pumps, industrial electrification. Adding massive data centre loads can tip the balance. Off-grid renewables represent an alternative pathway: instead of forcing the grid to catch up, build a parallel energy system that meets demand directly.
Of course, building parallel systems is not always feasible. Desert