Artificial intelligence is surging from pilot to production, and the physical footprint to run it is expanding just as quickly. That shift is not only about racks and GPUs. It is about where to find land, how to secure electricity, and which technologies can keep power and cooling reliable as workloads grow.
Analysts now estimate the generative AI economy could reach $4 trillion by 2030. Meeting that demand requires a step change in infrastructure: today’s global data center load is roughly 70 gigawatts, and within about five years it could approach 220 gigawatts. Around 75% of the expansion is linked to AI workloads. Build cycles of 18–24 months are common and often stretch longer, while total capital needs are discussed in the vicinity of one trillion dollars.
Power availability is the new site selector
Historically, the industry’s primary hubs were developed based on an expectation of steady growth. Today, many of these major markets are constrained by the limited capacity of their power grids in the near term.
As a result, large-scale cloud providers and data center developers are now establishing campuses in locations where electricity can be delivered more quickly, even if those areas are not traditional technology centers and have modest local demand. This is why regions not previously considered major players are now being shortlisted; the strategy has shifted to “power first, everything else second.”
Two realities are driving this change. First, the power consumption of server racks continues to rise, leading to single campuses that may require hundreds of megawatts. Second, lengthy queues for grid interconnection and bottlenecks in power transmission limit how fast that electricity can actually be supplied.
Architectures are shifting to handle heat and scale
As the physical footprint of AI expands, the underlying data center architectures are evolving to meet new demands for power density and thermal management.
Cooling pivots to liquids
Compute-dense AI clusters generate more heat than legacy cooling strategies can efficiently remove. Operators are moving toward liquid solutions to keep thermals in range while preserving performance.
From training heavy to inference everywhere
Right now, large training clusters dominate capacity planning. Over the next several years, the balance tilts toward inferencing at scale, supported by a mesh of edge sites. Expect a dual track: very large campuses for model training and a broader constellation of smaller facilities to serve low-latency inference.
Building fast is hard: The grid is the gating factor
Even with shovel-ready land, construction, commissioning, and interconnection commonly take 18–24 months. In many regions the critical path is upstream of the meter. Developers need transmission upgrades, new substations, and firm generation commitments. In markets that have not seen material net load growth for years, AI demand is now a primary driver of new electricity planning.
A pragmatic toolkit: Near term, mid term, long term
Addressing these infrastructure challenges requires a multi-horizon approach, with distinct strategies for the immediate future, the medium term, and the long run.
Near term: Squeeze existing assets
- Deploy battery storage to smooth peaks and increase utilization on constrained lines.
- Add modular, on-site options such as fuel cells, generator sets, or small turbines to bridge interconnection delays.
- Standardize high-density designs and liquid cooling to raise watts per rack without runaway PUE.
Mid term: Build for reliable, cleaner supply
- Advance large central generation where feasible, including gas plants that can support firm capacity needs.
- Accelerate utility-scale wind and solar coupled with storage, anchored by long-term contracts and clear interconnection plans.
- Pilot the next wave of cleaner technologies at commercial scale to prove cost curves and operating models.
Long term: Commercialize the next generation
- Scale options such as geothermal, carbon capture on thermal units, and advanced nuclear designs as they clear demonstration milestones.
- Modernize transmission to connect resource-rich regions with demand centers and shorten future interconnection queues.
Strategy by role: How to invest with fewer regrets
Navigating this complex landscape requires tailored strategies for different stakeholders, from enterprise leaders to investors.
Enterprises (CIOs and CTOs)
- Plan for rapid adoption but confront structural blockers: data readiness, governance, and budget ownership across business units.
- Tie AI programs to measurable outcomes, not tool counts. Track cost per inference, service levels, risk controls, and revenue impact.
Suppliers, developers, and model providers
- Work backward from the customer: hyperscalers, platform partners, or enterprise buyers. Clarify who benefits and how usage grows.
- Design for diverse cooling and power envelopes. Offer reference architectures that de-risk high-density deployments.
Investors
- Favor durable business models that survive efficiency leaps and hardware cycles.
- Build flexibility into capital plans so allocations can scale up or down as supply and demand shift.
- Account for geopolitical and regulatory risk across siting, equipment sourcing, and power procurement.
Partnerships decide speed
No single organization can solve siting, generation, transmission, and technology evolution alone. Utilities, grid operators, hyperscalers, developers, equipment makers, and governments all have a piece. The fastest projects align on standard designs, transparent interconnection roadmaps, and clear risk-sharing. Remember the grid is shared. There is no separate “AI power.” Data centers must fit within regional systems that keep hospitals, factories, and homes running at the same frequency.
What to watch over the next 12–24 months
- Adoption of liquid and hybrid cooling across new builds.
- Shifts in site selection toward power-rich regions and cross-state transmission agreements.
- Growth in edge facilities to support inference latency targets.
- Interconnection queue timelines and policies that unlock stranded capacity.
- Corporate power contracts that pair renewables, storage, and firming resources.
- Demonstrations of geothermal, carbon capture, or advanced nuclear moving from pilots to commercial commitments.
The bottom line
The long-standing checklist for site selection—once topped by network latency and land availability—has been completely upended. Today, the first and most critical question is not “Is there fiber?” but “Can we get power?” Access to a robust, scalable, and readily available energy source has become the ultimate gating factor. This new reality is turning previously overlooked regions with ample grid capacity into prime real estate, as developers follow the megawatts.
This new era of high-density computing is also generating unprecedented levels of heat, rendering traditional air-cooling methods insufficient. Consequently, cooling technology is in the midst of a critical evolution, with a rapid pivot towards advanced solutions like direct-to-chip liquid cooling and full immersion systems. These aren’t just upgrades; they are essential innovations to manage the thermal dynamics of powerful AI processors packed tightly together.
However, securing a power source is only half the battle. The most significant bottleneck is often the grid itself. Project timelines are no longer dictated solely by construction speed but are increasingly stretched by the lengthy and complex processes of securing grid interconnections and waiting for transmission infrastructure to be upgraded. A project can be shovel-ready, but if it has to wait years in a queue for a substation to be built, progress grinds to a halt.
