AI Data Centers Are Hitting the Power Wall: SpaceX, SoftBank, and Mission Gobi Offer Three Answers

Introduction

The next limit on artificial intelligence may not be model architecture or semiconductor performance. It may be the physical infrastructure required to keep millions of accelerators powered, cooled, connected, and operating around the clock.

Several events in 2026 made that constraint unusually visible.

A major heatwave disrupted access to the University of Cambridge's Dawn supercomputer after problems at the facility that housed it. In the United States, PJM—the country's largest regional electricity market—faced record demand, sharp balancing-cost spikes, and capacity prices that had risen by more than tenfold in two years. Manufacturers in the same region reported electricity charges climbing as data-center demand expanded.

At the same time, QTS abandoned its part of the long-disputed Digital Gateway project in Virginia. The campus had been described as one of the largest data-center developments ever proposed, yet years of legal challenges, community opposition, land-use disputes, and environmental concerns prevented it from moving forward.

These cases point to the same conclusion: AI infrastructure is no longer a simple matter of buying GPUs and finding warehouse space. Electricity, transmission, cooling, water, land, permits, and social acceptance are becoming part of the AI stack.

Three major strategies now illustrate how companies are trying to solve that problem:

1. Move compute toward continuous solar energy in orbit.
2. Build dedicated power generation beside giant terrestrial data-center campuses.
3. Move compute toward renewable-rich desert regions and coordinate energy, storage, and workloads as one system.

This article examines those approaches through SpaceX's orbital data-center filing, SoftBank and SB Energy's Ohio project, and Envision Energy's Mission Gobi initiative.

The Power Constraint Is Becoming an AI Constraint

AI systems consume electricity at multiple layers.

Accelerators need power for computation. Networking equipment moves data between machines. CPUs, storage, memory, and cooling systems add further demand. Power conversion and distribution introduce losses, while backup systems and redundancy increase the amount of infrastructure that must be installed.

The result is a facility whose energy requirements can resemble those of a city or an industrial complex.

The physical constraints are also tightly coupled:

power-generation processes |
| Land | Provides space for buildings, substations, cooling, generation, and transmission |
| Permitting | Controls how quickly a project can move from plan to construction |
| Community acceptance | Influences legal risk, project timelines, and political support |
| Hardware supply | Limits how quickly the facility can be equipped |
| Grid stability | Determines whether the campus can operate without worsening regional reliability |

A data center can secure GPUs and still fail if the transmission connection is delayed. It can secure power and still face opposition over water use, noise, land conversion, or proximity to communities. It can build the physical campus and still suffer downtime if cooling was designed around historical temperatures rather than more extreme weather.

This is why the AI infrastructure debate is shifting from “Who has the best chip?” to “Who can organize the complete system?”

Heat Exposed the Fragility of High-Performance Computing

In late June 2026, the University of Cambridge’s Dawn AI supercomputer was taken offline during a heatwave after technical problems affected the data center’s cooling infrastructure.

Dawn supports projects involving climate modeling, medical research, and other computationally intensive work. Reports said no research data was lost, but the outage interrupted access for multiple teams and delayed ongoing projects.

The incident matters for two reasons.

First, it shows that even advanced research infrastructure can be vulnerable to environmental conditions outside its expected operating envelope.

Second, it demonstrates a feedback loop that will become increasingly important: AI is being used to model climate change, optimize energy systems, and develop new technologies, yet the computing infrastructure itself must survive the physical effects of rising temperatures.

Resilience planning for AI facilities therefore needs to consider:

• Higher ambient temperatures
• Longer heatwaves
• Reduced cooling efficiency
• Water restrictions
• Grid stress during peak demand
• Simultaneous failures across power and cooling systems
• Recovery times after protective shutdowns

A facility designed only for average historical weather may not be sufficient for the next decade of AI deployment.

PJM Shows How Data-Center Demand Can Affect the Wider Grid

PJM Interconnection coordinates electricity markets across a region serving roughly 67 million people in the United States.

The rapid addition of data-center demand has collided with slow generation development, delayed transmission, plant retirements, and the long queue for connecting new resources.

PJM’s 2026/2027 capacity auction cleared at $329.17 per megawatt-day. That compares with $28.92 per megawatt-day for the 2024/2025 delivery year, an increase of more than tenfold.

Capacity payments are intended to ensure enough generation is available during periods of peak demand. When projected demand rises faster than available supply, the cost of securing that capacity can increase sharply.

The effect is not limited to technology companies.

Reuters reported that Belden Brick, a 141-year-old manufacturer in

Ohio, saw its monthly capacity charge rise from about $1,600 to about $12,000. The company operates in the same broad market where hyperscale data centers are creating large new electricity loads.

This does not mean every increase in electricity prices can be attributed only to AI. Fuel costs, transmission congestion, generation retirements, market rules, weather, and broader demand all matter.

However, very large data-center forecasts are now affecting how utilities, grid operators, regulators, and industrial customers plan for future supply.

Why Existing Grids Struggle With AI Loads

Traditional electricity systems were built around relatively predictable growth.

AI campuses can request gigawatts of new capacity on compressed timelines. A single project may be larger than the entire load of a regional city.

The grid must then solve several problems:

1. Build or retain enough generation.
2. Upgrade transmission and substations.
3. Maintain reliability during peaks.
4. Decide who pays for the new infrastructure.
5. Prevent speculative or duplicated load requests from distorting planning.
6. Protect ordinary consumers from costs created by unusually large users.

These questions are becoming central to AI policy.

Digital Gateway Demonstrated That Land and Permission Matter

The proposed Digital Gateway project in Prince William County, Virginia, became one of the most prominent examples of community resistance to data-center expansion.

The broader development area covered more than 2,100 acres and was planned to support tens of millions of square feet of data-center construction. QTS and Compass Datacenters were among the major developers involved.

Supporters argued that the project would create investment, tax revenue, and jobs. Opponents raised concerns about land-use procedures, historic landscapes, power demand, transmission lines, noise, environmental effects, and the industrialization of rural areas near Manassas National Battlefield Park.

A court invalidated the rezoning approval after finding problems with the public-notice process. Other parties ended their appeals, and QTS withdrew its remaining legal challenge in July 2026.

The project’s collapse illustrates a fundamental infrastructure rule:

Technical feasibility does not equal project feasibility.

A data center may be financeable, buildable, and commercially attractive but still fail because the surrounding community, legal process, and physical landscape do not support it.

For AI developers, site selection now requires more than access to fiber and electricity. It requires early work on:

• Community engagement
• Environmental review
• Water planning
• Historic preservation
• Noise mitigation
• Transmission routing
• Local employment
• Tax arrangements
• Emergency services
• Public transparency

The cheapest land on a spreadsheet may become extremely expensive after years of litigation.

Strategy One: SpaceX Moves Compute Toward Orbital Solar Energy

In January 2026, SpaceX filed an application with the U.S. Federal Communications Commission for a new non-geostationary satellite system of up to one million satellites.

The filing described the

The proposed network is referred to as the SpaceX Orbital Data Center system.

According to the FCC public notice, the satellites would operate across orbital shells between approximately 500 and 2,000 kilometers. They would use high-bandwidth optical inter-satellite links and connect with parts of the Starlink network.

The underlying idea is easy to understand.

Space has access to abundant solar energy. Orbital systems avoid competition for terrestrial land and may reduce dependence on local electricity grids. Heat can be rejected through radiation rather than conventional cooling towers.

SpaceX’s approach represents an engineering-first response: if power, land, and permitting are difficult on Earth, move the infrastructure outside the terrestrial system.

Potential Advantages of Orbital Compute

An orbital data-center architecture could offer:

• Long periods of solar exposure
• Reduced competition for terrestrial land
• No direct use of local freshwater for cooling
• Global connectivity through satellite networks
• Proximity to space-based sensors and applications
• The ability to process some data before transmitting it to Earth

Space-native workloads may be the most practical starting point.

For example, satellites could process Earth-observation imagery, weather data, navigation information, or communications traffic before sending a smaller result to the ground. This reduces the amount of data that must be transmitted.

The Engineering Problems Are Severe

Moving compute to orbit does not eliminate infrastructure constraints. It changes them.

An orbital data center must solve:

• Launch mass and launch cost
• Radiation damage
• Thermal management in vacuum
• Solar-array sizing
• Energy storage during eclipse
• Space debris
• Hardware replacement
• Communications bandwidth
• Latency
• Maintenance
• Satellite lifetime
• End-of-life disposal

Heat remains a major challenge.

On Earth, data centers can transfer heat to air or water. In space, there is no surrounding fluid to carry heat away. The system must use radiators, and those radiators can become very large relative to the amount of computing power deployed.

A 2026 technical analysis of orbital data-center economics concluded that general-purpose terrestrial workloads would require extremely low combined launch and spacecraft costs, high utilization, long operating life, and favorable communication requirements to compete with ground infrastructure.

This suggests that orbital AI compute may develop gradually rather than replacing terrestrial data centers quickly.

A Filing Is Not a Deployed System

The FCC accepted SpaceX’s application for filing and invited public comment. That is an early regulatory step, not permission to deploy one million satellites immediately.

Large questions remain about spectrum, orbital safety, environmental effects, astronomy, debris, and the cumulative impact of launching and replacing such a large constellation.

The orbital strategy is ambitious, but its commercial timeline depends on breakthroughs in spacecraft manufacturing, launch frequency, thermal design, networking, and regulation.

Strategy Two: SoftBank Builds Power Beside Compute

SoftBank and SB Energy are pursuing a very different approach in Ohio.

The U.S. Department of Energy announced plans for a technology campus at the former Portsmouth Gaseous Diffusion Plant in Pike County. The project is intended to support approximately 10 gigawatts of data-center development.

The associated energy plan includes up to 10 gigawatts of new generation, with at least 9.2 gigawatts expected to come from natural gas.

SB Energy and AEP Ohio also plan to invest in major transmission infrastructure. The Department of Energy said the project would pay for the required power-delivery upgrades and use a dedicated rate structure intended to prevent ordinary customers from absorbing those costs.

This approach is built around speed and scale.

Instead of waiting for the existing grid to produce enough spare capacity, the project would construct a large supply of generation specifically alongside the planned compute campus.

Why Gas Is Attractive for Rapid Development

Natural gas power plants can provide dispatchable electricity.

Unlike wind and solar generation, they can operate when required rather than only when weather conditions are favorable. That makes them easier to pair with data centers that expect continuous service.

The Ohio model offers several practical advantages:

• Dedicated generation for a giant new load
• A site already controlled by the federal government
• Large-scale transmission investment
• Faster access to dispatchable power
• A clear connection between the data-center customer and infrastructure costs
• The possibility of sending excess power to the wider grid

For a company focused on deploying compute quickly, this can look more realistic than waiting for a slow regional grid expansion.

The Climate Trade-Off Is Significant

The main weakness is carbon intensity.

A 9.2-gigawatt natural-gas generation portfolio would produce substantial emissions if operated heavily. It may also require pipelines, cooling, water, and long-lived infrastructure that could continue operating for decades.

The project therefore creates a tension between two policy goals:

1. Build AI infrastructure rapidly.
2. Reduce emissions from the energy system.

Natural gas can provide reliable power during a period of fast demand growth, but it risks locking future AI capacity into fossil-fuel dependence.

The actual environmental outcome will depend on plant efficiency, utilization, methane leakage, carbon-management measures, grid interaction, and how quickly lower-carbon resources become available.

Strategy Three: Mission Gobi Moves Compute Toward Renewable Energy

Envision Energy announced Mission Gobi at VivaTech 2026.

The initiative targets 5 gigawatts of green AI data-center capacity in desert and arid regions by 2030.

Its central principle is that compute should follow energy.

Rather than build a massive AI campus near an existing city and then ask the grid to supply it, Mission Gobi proposes placing data centers where wind, solar, land, and energy-storage resources are already abundant.

Envision describes the approach as an AI Power System that connects:

• Wind generation
• Solar generation
• Battery

storage

• Grid infrastructure
• Computing workloads
• Green hydrogen production
• AI-based forecasting and dispatch

The design treats the data center and the energy system as one coordinated platform.

The Chifeng Demonstration

Envision says it has already deployed this model in Chifeng, Inner Mongolia.

The company’s official reporting describes a 2-gigawatt renewable power system that uses EnOS and an energy foundation model to coordinate wind, solar, storage, computing workloads, and green hydrogen production.

Envision also says it is working with Tencent to match AI workloads dynamically with renewable-energy availability.

This is different from a conventional renewable purchase agreement.

A traditional data center may consume electricity continuously and buy certificates or contracts that represent renewable generation elsewhere. In an AI-native power system, the facility attempts to coordinate real power flows, storage, generation forecasts, and flexible computing demand in near real time.

The Ulanqab Galaxy Campus

Envision says the larger Galaxy Campus in Ulanqab is being developed as a gigawatt-scale AI data center connected directly to renewable energy.

The company’s press materials describe it as a flagship implementation of the Mission Gobi model.

The basic site-selection logic is strong:

• Desert regions offer extensive land.
• Wind and solar resources can be abundant.
• Population density is lower.
• Community conflict may be reduced.
• Cooler and drier conditions can improve some cooling strategies.
• Direct power connections can reduce dependence on congested urban grids.

However, each desert site has different water availability, transmission options, environmental conditions, and distance from users. "Build in the desert" is not a complete solution on its own.

How Mission Gobi Handles Variable Renewable Power

The largest technical challenge is intermittency.

Data centers are designed for continuous operation, while wind and solar output changes over seconds, hours, days, and seasons.

Mission Gobi’s architecture combines several forms of flexibility.

Battery Storage for Fast Changes

Batteries can respond within milliseconds.

They are useful for:

• Frequency support
• Short-term balancing
• Smoothing rapid changes in wind and solar
• Bridging short outages
• Supporting power quality
• Shifting energy across several hours

Batteries are less suitable for storing enough energy to cover very long periods of low renewable output at gigawatt scale.

Wind and Solar Complementarity

Wind and solar often produce at different times.

Solar generation is strongest during daylight hours. Wind patterns may be stronger at night or during different seasons.

Combining both resources can reduce—but not eliminate—the amount of storage required.

Hydrogen and Ammonia for Longer-Duration Storage

Excess renewable electricity can be used to produce hydrogen through electrolysis.

Hydrogen can then be stored directly or converted into ammonia, which is easier to transport and store in some applications. When renewable output is low, these energy carriers can support power generation or industrial

processes.

This approach offers longer-duration storage but introduces conversion losses and additional cost.

The full cycle involves:

1. Generating renewable electricity.
2. Producing hydrogen.
3. Potentially converting hydrogen to ammonia.
4. Storing the fuel.
5. Converting it back into electricity or another useful output.

Each step reduces efficiency. The system must therefore decide when long-duration storage provides more value than additional generation, transmission, batteries, or flexible workloads.

Flexible AI Workloads

Not every computing task needs to run at the same moment.

Some workloads can be shifted:

• Data preprocessing
• Synthetic data generation
• Batch inference
• Model evaluation
• Checkpoint conversion
• Nonurgent training
• Rendering
• Simulation
• Offline analytics

An AI-aware energy-management system can schedule more flexible work when renewable energy is abundant while reserving stable capacity for latency-sensitive inference and critical services.

This is one of Mission Gobi’s most important ideas.

Instead of forcing the energy system to behave as if every computing task is inflexible, it allows some computing demand to adapt to energy availability.

Comparing the Three Strategies

The three approaches solve the same basic problem in very different ways.

None is a universal answer.

Orbital compute may be attractive for space-native processing before it is competitive for general cloud workloads.

Gas-backed campuses may scale fastest but face carbon and environmental pressure.

Renewable desert campuses may offer strong long-term economics, but they require sophisticated control, storage, flexible workloads, and reliable connectivity.

The likely future is a mixture.

Why “Compute Follows Power” Is Becoming a Site-Selection Rule

For most of the cloud era, data centers were often placed near network hubs, customers, skilled labor, favorable tax regimes, and reliable grids.

AI changes the balance.

Training and many batch workloads are less sensitive to physical distance than real-time consumer applications. A model-training cluster can be placed far from a major city if it has power, fiber, and operational support.

This makes energy-rich regions more attractive.

The site-selection process may increasingly begin with:

1. Where is large-scale power available?
2. Can it be delivered continuously?
3. What is the marginal carbon

intensity?
4. Can the load support the regional grid rather than destabilize it?
5. Is enough land available for generation, storage, cooling, and compute?
6. Can the project secure permits and community acceptance?
7. Is the fiber network sufficient?
8. Which workloads can tolerate the location?

The data center is becoming part of the power system rather than simply a customer at the end of the wire.

Water Remains a Constraint Even in Green Projects

Renewable electricity does not automatically eliminate water use.

Water can be consumed by:

• Evaporative cooling
• Cooling towers
• Humidification
• Power-generation processes
• Semiconductor manufacturing
• Hydrogen production
• Construction and maintenance

Desert regions can offer low humidity and cooler nights, but they may also have scarce water resources.

A credible green AI project therefore needs a complete water strategy.

That may include:

• Direct-to-chip liquid cooling
• Closed-loop systems
• Dry coolers
• Air cooling where practical
• Reclaimed water
• Heat reuse
• Water-use reporting
• Site-specific limits
• Designs that avoid competition with local communities and agriculture

Water efficiency should be evaluated alongside carbon intensity and electricity cost.

Green Hydrogen Is Useful, but It Is Not a Free Battery

The original Mission Gobi concept gives green hydrogen and green ammonia an important role in long-duration balancing.

That role is technically plausible, especially where the system also serves industrial demand for hydrogen or ammonia.

However, using hydrogen only to regenerate electricity is less efficient than direct use of renewable power or short-duration battery storage.

Hydrogen becomes more attractive when it can provide several forms of value:

• Seasonal energy storage
• Industrial feedstock
• Export fuel
• Backup generation
• Grid balancing
• Revenue from chemical products
• Use of renewable energy that would otherwise be curtailed

The economics depend on electrolyzer utilization, storage infrastructure, conversion efficiency, fuel demand, and the cost of excess renewable electricity.

Mission Gobi should therefore be understood as an integrated energy-and-compute system, not simply a data center with a hydrogen backup generator.

The AI Layer Could Become the Main Differentiator

Wind turbines, solar panels, batteries, data centers, and hydrogen equipment already exist.

The more distinctive part of Mission Gobi is the coordination layer.

A system like EnOS can combine:

• Weather forecasts
• Electricity production forecasts
• Battery state of charge
• Hydrogen storage levels
• Data-center load
• Model-training schedules
• Grid prices
• Network constraints
• Maintenance requirements

It can then decide when to:

• Run flexible AI jobs
• Charge batteries
• Produce hydrogen
• Reduce nonessential computing
• Export power
• Import power
• Reserve capacity for critical inference
• Shift work between locations

This is where AI infrastructure begins to resemble a cyber-physical operating system.

The value comes not only from owning energy assets, but from coordinating them better than a collection of separate operators could.

What

Data-Center Developers Should Learn

The events around Dawn, PJM, Digital Gateway, Ohio, orbital compute, and Mission Gobi point toward a new set of design principles.

1. Energy Must Be Planned Before Compute

A project should not assume that the grid will produce gigawatts of spare electricity on demand.

Generation, transmission, storage, backup, and demand flexibility must be part of the initial design.

2. Community Acceptance Is Infrastructure

Permits and public support are not public-relations details. They are critical-path engineering dependencies.

A legally contested site can delay or destroy an otherwise viable project.

3. Resilience Must Include Climate Extremes

Cooling and power systems should be tested against heatwaves, drought, storms, fires, and other conditions outside historical averages.

4. Flexible Workloads Have Economic Value

The ability to move nonurgent computing across time or geography can reduce power costs and make renewable energy easier to use.

5. Carbon, Water, and Land Must Be Evaluated Together

A low-carbon design can still create water or land conflicts. A water-efficient site can still depend on high-emission electricity.

The full environmental system matters.

6. Dedicated Power Does Not Remove Public Responsibility

A privately funded power plant can protect ratepayers from some direct costs, but emissions, pipelines, water, air pollution, and regional reliability still affect the public.

7. The Best Architecture May Be Hybrid

Future AI systems may combine:

• Urban inference centers
• Remote training campuses
• Renewable desert compute
• Gas or nuclear backup
• Distributed edge systems
• Limited orbital processing
• Workload movement across regions

Different workloads have different requirements.

Frequently Asked Questions

Why are AI data centers causing power-grid concerns?

AI clusters can require hundreds of megawatts or several gigawatts of electricity. When projects arrive faster than generation and transmission can be built, they can increase capacity costs, delay other connections, and create reliability concerns.

What happened to the Digital Gateway data-center project?

QTS withdrew its remaining legal challenge in July 2026, effectively ending its portion of the proposed Virginia project. The development had faced years of litigation and community opposition, including disputes over rezoning procedures and environmental impact.

What is SpaceX’s orbital data-center plan?

SpaceX filed an FCC application for a non-geostationary system of up to one million satellites described as an orbital data-center network. The concept would use solar-powered satellites and optical links, but it remains at an early regulatory and engineering stage.

What is the SoftBank Ohio AI campus?

SoftBank, SB Energy, and partners plan a 10-gigawatt data-center campus at the former Portsmouth Gaseous Diffusion Plant in Ohio. The associated plan includes up to 10 gigawatts of new generation, at least 9.2 gigawatts of which is expected to be natural gas.

What is Mission Gobi?

Mission Gobi is Envision Energy’s initiative to develop 5 gigawatts of green AI data-center capacity

in desert and arid regions by 2030. It combines renewable generation, storage, grid infrastructure, computing, and AI-based energy management.

Is GobiX the official project name?

Envision’s official announcement uses Mission Gobi. “GobiX” is an informal label used in some commentary to contrast desert-based AI infrastructure with SpaceX’s orbital approach.

Can renewable energy power an AI data center continuously?

Yes, but continuous service generally requires a portfolio that may include overbuilt wind and solar, batteries, long-duration storage, grid support, backup generation, and workload flexibility. The exact design depends on the site and reliability target.

Why place AI data centers in deserts?

Desert and arid regions can offer large areas of land, strong wind and solar resources, lower population density, and reduced competition with urban grids. Developers still need to solve water, connectivity, environmental, storage, and maintenance challenges.

Related Tools

• Envision EnOS: An AI-powered operating system for coordinating renewable energy, storage, assets, and industrial workloads.
• PJM Data Viewer: A public interface for monitoring electricity demand, generation, pricing, and grid conditions in the PJM region.
• NVIDIA Mission Control: Software for operating and managing large-scale AI infrastructure.
• Open Compute Project: An industry community developing open designs for efficient data centers, servers, cooling, and power systems.
• Green Software Foundation: A nonprofit organization developing standards and practices for reducing the environmental impact of software.

Related Links

• Envision Launches Mission Gobi: The official press release describing the 5GW desert AI data-center initiative.
• Envision 2026 Net Zero Action Report Announcement: Details on Chifeng, Tencent collaboration, EnOS, and Mission Gobi.
• FCC Notice on SpaceX Orbital Data Centers: The official public notice covering SpaceX’s application for up to one million satellites.
• U.S. Department of Energy Ohio AI Campus Fact Sheet: Official details on the 10GW data center, gas generation, and transmission investment.
• PJM 2026/2027 Capacity Auction Report: Official auction results and capacity-clearing prices.
• Reuters on QTS Ending Digital Gateway: Reporting on the termination of the Virginia

project.

Summary

AI infrastructure is moving beyond a chip-supply problem. Power generation, transmission, cooling, water, land, permits, and public acceptance are becoming the practical limits on how quickly new compute can be deployed.

SpaceX, SoftBank, and Envision represent three different responses. SpaceX proposes moving compute toward solar energy in orbit. SoftBank is pairing a giant Ohio campus with dedicated gas generation. Mission Gobi aims to move compute toward desert renewable resources and coordinate energy, storage, hydrogen, and flexible workloads through an AI power system.

Each approach solves one set of constraints while introducing another. Orbital systems face launch and thermal challenges. Gas-backed campuses create carbon risk. Renewable desert systems need storage, water planning, connectivity, and sophisticated coordination.

The next AI infrastructure winner may not be the company with the most accelerators, but the one that can organize energy, land, cooling, networks, and compute into the most reliable complete system.