The underlying mechanism is structural. GPU-intensive AI training workloads—the largest application segment—consume materially more power per rack than legacy compute

Decision Lens

A market sizing exercise from Precedence Research positions AI-driven power consumption as one of the fastest-scaling infrastructure cost categories in the industry. The headline projection—roughly $12.5 billion in 2025 growing to $70.6 billion by 2035—is a commercial analyst estimate, not a verified industry consensus. The structural logic underneath it, however, matters: hyperscale operators reportedly hold 45% of this market, AI training workloads account for 35% of application demand, and the compound growth rate of 18.9% annually spans the exact horizon of most long-dated PPAs being signed today. For energy heads managing 10–15 year offtake commitments and multi-year interconnection queues, this trajectory is a planning variable, not background noise.

90-Second Brief

In recent days, precedence Research projects AI data center power consumption growing from $12.5 billion in 2025 to $70.6 billion by 2035, at a CAGR of 18.9%. North America holds an estimated 42% share today; Asia Pacific is projected to grow fastest at 21.5% annually through 2035. Cooling systems represent the largest component segment, reflecting the thermal pressure of high-density GPU workloads. These figures are commercial analyst projections and carry the uncertainty inherent in decade-long demand modeling.

What’s Actually Happening

The underlying mechanism is structural. GPU-intensive AI training workloads—the largest application segment—consume materially more power per rack than legacy compute. As generative AI and large language model deployments scale, average rack power density rises, simultaneously stressing cooling capacity, UPS headroom, and grid interconnection limits.

Hyperscale operators hold the dominant share, reflecting the concentration of AI training inside large-scale purpose-built facilities. Colocation providers at roughly 25% of market share are absorbing enterprise overflow from organizations unwilling to take on the capital intensity of dedicated AI infrastructure. The cooling technology split illustrates a transitional tension: air cooling holds a 55% share by cost, but liquid cooling at 30% is projected to grow fastest—a gradual, capex-heavy shift not yet fully absorbed into long-range infrastructure budget cycles.

Geographically, North America’s 42% concentration reflects entrenched hyperscaler density. Asia Pacific’s projected 21.5% growth rate signals new procurement markets forming faster than interconnection infrastructure in some of those jurisdictions can comfortably accommodate.

Why It Matters for Global Heads of Data Center Energy?

If this growth trajectory is directionally correct—even discounting for analyst optimism—the implications for procurement strategy are material. A market scaling at roughly 19% annually does not permit incremental PPA adjustments or reactive queue management. Long-dated energy contracts being negotiated today will govern assets operating under significantly higher load densities than currently modeled.

The hyperscale concentration reinforces this pressure. Competition for firm, long-duration clean power offtake agreements becomes structurally tighter when the majority of market spend is concentrated among a small number of operators. AI training workloads compound the challenge: unlike inference, they are not easily dispatchable, which limits demand-response optionality and raises the premium on baseload-adjacent renewable supply.

For colocation operators, the calculus is different but equally sharp. Energy heads pricing power for enterprise customers with AI workloads do not fully control those customers’ load profiles—introducing basis risk into procurement decisions that were previously more predictable. Asia Pacific growth adds jurisdictional complexity requiring utility relationships and interconnection strategies built under regulatory and grid conditions that diverge substantially from the North American baseline.

The Forward View

The composition of demand growth—weighted toward training workloads, generative AI, and high-density GPU clusters—suggests that power density per square foot will keep rising faster than overall facility count. That reframes the strategic problem from megawatt volume to megawatt quality: reliability margin, delivery density, and proximity to load.

Cooling infrastructure investment follows directly from that reframing. The projected shift toward liquid-dominated thermal management carries lead times that now interact with transformer and switchgear supply constraints already extended to two or three years in many markets. Energy procurement timelines and cooling capital cycles need to be synchronized more tightly than most planning processes currently allow.

Asia Pacific will likely move from a growth footnote to a board-level strategy priority within the forecast window. The cited drivers—China and India—present distinct grid access conditions, renewable supply curves, and regulatory frameworks that do not map onto North American procurement playbooks. Teams without regional expertise embedded in those markets face meaningful execution risk.

What We’re Uncertain About?

  • Forecast reliability over a 10-year horizon. Precedence Research is a commercial intelligence firm; this release promotes a paid product with no publicly disclosed methodology or confidence interval. The $70.6 billion figure is one scenario, not a range. Cross-referencing against Lawrence Berkeley National Laboratory load growth models or Wood Mackenzie grid forecasts would provide useful calibration.

  • Efficiency gains offsetting gross demand. Advances in model compression, inference optimization, and next-generation chip performance could materially reduce per-token energy costs. The forecast does not appear to model efficiency-adjusted demand trajectories. Scenario analysis distinguishing compute intensity per workload type from raw deployment volume would resolve this gap.

  • Asia Pacific grid readiness. The 21.5% regional CAGR assumes demand materializes but does not address whether grid infrastructure and interconnection capacity in high-growth markets can absorb it within the projected window. Market-specific assessments from regional grid operators or national energy authorities are needed.

  • Hyperscale share stability. The 45% concentration assumes current ownership patterns persist. A structural shift toward distributed edge inference or third-party AI compute platforms could disperse the load profile considerably, altering both the volume and geographic distribution of procurement demand.

One Question to Bring to Your Team

If AI power demand grows at even half the projected rate through 2030, which of our current PPA portfolios and interconnection positions are structurally undersized for the load densities we will actually be operating—and what is the cost of renegotiating or supplementing them now versus in three years?

Sources

  • Openpr — AI Data Center Power Consumption Market Size to Surge from USD 12.50 Billion in 2025 to USD 70.59 Billion by (Link)

Tags