It’s no secret that sustained demand for Artificial Intelligence (AI) has had a major impact on cloud infrastructure. The electric grid was not originally designed to accommodate extreme AI and High-Performance Computing (HPC)-based processing.
AI demand is rapidly outpacing grid capacity, leaving data centers and enterprises seeking alternative sources of reliable power. Concepts such as microgrids, nuclear energy, and Orbital Data Centers (ODCs) are now integral aspects of AI infrastructure strategies. At the same time, operators are turning to liquid cooling and AI itself to minimize energy costs.
Table of Contents:
The AI Infrastructure Conundrum
The AI Infrastructure Conundrum
ABI Research forecasts 122 Gigawatts (GW) of active Information Technology (IT) data center capacity to be added globally by 2035. AI workloads are expected to account for 64% of total capacity, underscoring the misalignment between AI demand growth and grid readiness.
Generating clean, reliable energy for AI data centers is a challenge in and of itself. However, a taller order lies in delivering energy to the right locations within a reasonable time frame. Currently, data center interconnection timelines in major data center hubs such as Northern Virginia, Tokyo, and Western Europe take up to 7 to 10 years. Getting substations up and running takes a similar time frame.
Long grid connection timelines are forcing operators to look beyond traditional utility models.
Beyond the Grid
Utility delivery models are proving to be unreliable amid the shift to AI-first infrastructure build-outs. A recent JP Morgan report indicated that more than 60% of data center capacity slated for a 2027 completion hasn’t begun construction; grid constraints are cited as a key reason behind project delays. With the traditional grid being unreliable, on-site energy is emerging as a viable alternative to the traditional grid.
- Microgrids: Small, localized energy storage and distribution systems that can operate either independently or connected to the main grid.
- Small Modular Reactors (SMRs): A compact nuclear reactor that can deliver Megawatt (MW)-scale power for companies on-site. Their modular design means that SMRs can be manufactured quickly.
- Other Solutions: Distributed Energy Resources (DERs), solar panels, battery storage, hydrogen backup, and other solutions are also being leveraged. This reflects the diverse technology stack being embraced by data center operators, rather than the use of any single solution.
Hybrid Cooling
According to ABI Research, annual data center cooling requirements will increase nearly sevenfold by 2035, reaching approximately 195 million tons. Server rack densities are increasing significantly, as IT infrastructure is the biggest cooling requirement at data center locations.
Liquid cooling is gaining traction due to its ability to handle AI/HPC workloads. Despite increased uptake, ABI Research sees the merging of liquid cooling with traditional air-cooled solutions. Direct-to-chip immersion will be essential for high-density AI infrastructure, while air-cooling accommodates low-density workloads (learn more in Research Analyst Paris McKinley’s article).
Data Center Reference Designs
Data center reference designs accelerate AI infrastructure by aligning compute, networking, cooling, and power systems under a single, replicable blueprint.
These pre-validated frameworks provide several benefits for AI data center operators:
- Reduced deployment timelines
- Lower engineering risk
- Improved coordination across infrastructure systems
- Simplified modular expansion
- Designs adaptable to regional climates
By 2030, ABI Research forecasts that 86% of new data center builds will use at least one reference design.
Data Centers in Space
Once considered a sci-fi concept, data centers in space are becoming a very real commercial reality. AI infrastructure build-out requires enormous computing power, something that the traditional grid is not ready to deliver at scale.
As of April 2026, more than US$3 billion has been invested in the ODC sector as technological advancements demonstrate immense potential. Companies like SpaceX and Starcloud lead the way for hyperscale AI/HPC-supported ODCs, which are a future concept. Currently, the earliest ODC deployments are Kilowatt (kW)-scale, with Axiom Space, ADA Space, and Kepler being key players.
Until MW-scale data centers orbit in space, several challenges must be addressed by the broader technology industry:
- Launch Costs: Orbital infrastructure remains costly to deploy.
- Thermal Management: Heat dissipation in space is difficult.
- Hardware Servicing: Repairs and upgrades are challenging.
- Radiation Exposure: Hardware must be radiation hardened.
- Connectivity: Reliable Earth-to-orbit data transfer is essential.
Conclusion
The AI industry is no longer about just model improvements and software development. Innovation cycles are now only as fast as how well AI infrastructure can accommodate compute increases. The implications for data center operators are clear: embrace on-site energy, cooling hybridization, and AI-powered building management tools to circumvent energy challenges. Concurrently, data center reference designs should be used to accelerate infrastructure design, and digital twins should be used to test AI infrastructure changes before real-world implementation.
Download the whitepaper, Overcoming AI Infrastructure’s Biggest Bottlenecks: Power, Cooling, and Data Center Design, for ABI Research’s full analysis. Our study embodies the interconnected nature of the company’s AI & Machine Learning, Cloud, Smart Buildings, Smart Energy, and Space Technologies & Innovation teams.
Paris McKinley