Projected Trends and Capacity Allocations for Cloud, AI Training, and AI Inference Workloads (2025–2035)
The next decade will be defined by more transformation in data center infrastructure, driven by the explosive growth of cloud computing, artificial intelligence (AI) training, and AI inference workloads. As digitalization accelerates across every sector, the demand for compute, storage, and network resources is surging, with AI workloads fundamentally reshaping the physical and operational fabric of data centers worldwide.
Project DC capacity growth to 2035
AE training and inferencing workloads
Workload distribution shift
Infrastructure updates for power density, cooling and network design
Cost per Inference, Training TCO, and Pricing Dynamics
Vendor and Technology Trends
Operational Implications
Scenarios and Sensitivity Analysis Through 2035
1. Projected Global Data Center Capacity Growth (2025–2035)
1.1 Capacity Growth Trajectories
The global data center industry is entering an era of unprecedented expansion. According to multiple leading analyses, global data center power capacity is projected to more than triple between 2024 and 2035, rising from approximately 40–60 GW today to over 106 GW by 2035. This growth is being driven primarily by the proliferation of AI workloads and the continued migration of enterprise IT to the cloud.
BloombergNEF forecasts that by 2035, data centers will draw 106 GW, up from 40 GW in 2024—a nearly 300% increase.
McKinsey estimates a compound annual growth rate (CAGR) of 19–22% for global data center capacity from 2023 to 2030, reaching 171–219 GW by 2030, with a high-end scenario of 298 GW.
Goldman Sachs projects global power demand from data centers to increase by 165% by 2030, with AI workloads accounting for a growing share.
Bain & Company expects global data center capacity demand to reach 163 GW by 2030, double today’s demand, with the U.S. alone consuming 409 TWh of electricity—about 9% of the country’s total.
This expansion is not just about more facilities; it is also about larger, denser, and more specialized data centers. The average new facility is expected to draw well over 100 MW, with some mega-sites exceeding 500 MW or even 1 GW.
1.2 Regional Patterns and Constraints
While North America—especially the United States—remains the largest data center market, growth is accelerating in Asia-Pacific (notably India, China, Singapore, and Japan), Europe (with a focus on sustainability and regulatory compliance), and the Middle East (Saudi Arabia, UAE). However, regional capacity constraints are emerging:
Primary U.S. markets (e.g., Northern Virginia, Silicon Valley) are experiencing record-low vacancy rates (as low as 1.9% in 2024), pushing expansion into secondary and rural markets with available power and land.
Asia-Pacific is witnessing exponential growth, with India’s capacity projected to rise from 1.4 GW in 2024 to 9 GW by 2030.
Europe is enforcing strict sustainability and water usage regulations, with some regions (e.g., Dublin, Amsterdam) imposing moratoriums or quotas on new builds until grid and environmental upgrades are completed.
1.3 Infrastructure Supply Chain and Construction Costs
The cost of building data centers is rising, driven by higher energy prices, interest rates, and supply chain constraints. Modern builds average $7–12 million per MW for hyperscale facilities, with compute (servers) as the largest line item, followed by networking, power, and cooling. Construction timelines have extended to 2–4 years, and in some cases up to 6 years, due to permitting and grid connection delays.
2. AI Training vs. Inference Workload Trajectories (2025–2035)
2.1 Diverging Compute Demands
AI workloads are bifurcating into two distinct categories: training and inference. Each imposes unique demands on data center infrastructure:
Training involves running massive datasets through large models (e.g., LLMs) over days or weeks, requiring high numerical precision, dense GPU clusters, and ultra-fast interconnects. It is throughput-driven and capital-intensive, but typically a one-time or periodic event per model.
Inference is the process of deploying trained models to serve real-time user queries, often at massive scale. It is latency-sensitive, continuous, and increasingly dominates the operational cost of AI systems.
2.2 The Shift to Inference
The industry is witnessing a seismic shift: inference workloads are rapidly overtaking training in both hardware demand and operational expenditure. Key trends include:
By 2025, inference compute demand is expected to outstrip training by a factor of 3–10× in aggregate hardware usage, with some forecasts suggesting inference will consume 75% of all AI compute by 2030.
The global AI inference market is projected to grow from $97.24 billion in 2024 to $253.75 billion by 2030 (CAGR 17.5%), with North America leading and Asia-Pacific as the fastest-growing region.
Inference is becoming the main growth driver for AI infrastructure spending, with ongoing operational costs (per query, per token) far exceeding the one-time cost of training. For example, OpenAI’s 2024 inference spend was estimated at $2.3 billion—15 times the cost to train GPT-4.
2.3 Economic and Technical Implications
Training is increasingly concentrated in a few hyperscale “AI factories” located near abundant, low-cost power, often in rural or secondary markets. These facilities are designed for extreme density (up to 120 kW/rack and beyond) and leverage advanced liquid cooling and high-voltage DC power architectures.
Inference is distributed across cloud, colocation, enterprise, and edge environments, with a growing emphasis on latency, cost efficiency, and sustainability. Inference hardware is becoming more heterogeneous, with a mix of GPUs, ASICs (e.g., TPUs, Inferentia), FPGAs, and emerging accelerators optimized for low-power, high-throughput serving.
3. Workload Distribution Shifts: Cloud, Colocation, Enterprise, and Edge
3.1 Hybrid IT and Workload Placement
The era of “cloud-only” is over. Nearly all enterprises (98%) are adopting hybrid IT models, blending public and private cloud, on-premises, and colocation services to optimize workload placement based on cost, performance, compliance, and connectivity.
Cloud providers (hyperscale and neo-cloud) are capturing the majority of new AI training and inference workloads, especially for large-scale, multi-tenant applications.
Colocation is experiencing a renaissance, with enterprises repatriating workloads from public cloud to colocation for cost, control, and direct interconnection to cloud and network providers. AI and high-density workloads are a primary driver of this shift.
Enterprise data centers are evolving toward hybrid and edge deployments, focusing on latency-sensitive, regulated, or legacy workloads that cannot easily migrate to cloud or colocation.
Edge data centers are proliferating to support real-time AI inference, 5G, IoT, and other latency-critical applications, often in micro or modular form factors.
3.2 Table: Workload Distribution Trends (2025–2035)
Category
2025 Share
2030 Share (Est.)
2035 Share (Est.)
Key Workloads
Growth Drivers
Hyperscale
~55%
~60%
~65%
AI training, cloud, inference
AI/GenAI, SaaS, cloud migration
Neo-Cloud
~5%
~10%
~15%
GPUaaS, specialized AI
Cost, performance, flexibility
Colocation
~25%
~20%
~15%
Hybrid IT, AI inference
Repatriation, interconnection, cost
Enterprise
~10%
~7%
~3%
Legacy, regulated, hybrid
Compliance, latency, legacy systems
Edge
~5%
~8%
~12%
AI inference, IoT, 5G
Latency, distributed AI, 5G/IoT
*Estimates synthesized from multiple sources including McKinsey, Goldman Sachs, CoreSite, and industry surveys.
3.3 Analysis
The next decade will see a continued migration of workloads to hyperscale and neo-cloud platforms, especially for AI training and large-scale inference. However, colocation and edge will play increasingly strategic roles for latency-sensitive, regulated, or cost-optimized workloads. Enterprises will continue to optimize their hybrid IT mix, leveraging colocation for high-density AI, direct cloud interconnection, and predictable cost structures, while deploying edge nodes for real-time inference and data sovereignty.
4. Infrastructure Implications: Power Density, Cooling, and Network Design
4.1 Power Density and Rack-Level Trends
The rise of AI workloads is driving a dramatic increase in rack power density:
Average rack density has increased fivefold since 2011, from 2.4 kW to 12 kW in 2024, and is projected to reach 30–120 kW/rack for AI-centric deployments by 2027–2030.
AI training clusters (e.g., NVIDIA GB200, Blackwell) can require 80–120 kW per rack, with future designs targeting 250–600 kW/rack and even 1 MW/rack in experimental deployments.
Power usage effectiveness (PUE) has plateaued at 1.55–1.6 on average, but leading-edge hyperscale sites achieve PUEs as low as 1.10–1.20 through advanced cooling and power management.
4.2 Cooling Technologies and Adoption
Traditional air cooling is reaching its limits in high-density environments. The industry is rapidly adopting advanced cooling solutions:
Direct-to-chip liquid cooling is now deployed in 20–35% of new AI-centric data centers, with adoption expected to exceed 50% by 2030.
Immersion cooling is gaining traction for the most demanding workloads, enabling rack densities of 100–250 kW and beyond.
Hybrid cooling (air + liquid) is common in colocation and retrofit scenarios, balancing cost and performance.
Rear-door heat exchangers and cold plate cooling are widely used for intermediate densities (40–80 kW/rack).
Table: Cooling Technology Adoption and Capabilities
Cooling Type
2025 Adoption
Max Density Supported
Typical Use Cases
Air Cooling
~60%
Up to 20–30 kW/rack
Legacy, enterprise, cloud
Direct-to-Chip Liquid
~20–35%
60–120 kW/rack
AI, HPC, new builds
Immersion Cooling
~5–10%
100–250+ kW/rack
AI training, HPC, edge
Hybrid Cooling
~20%
40–80 kW/rack
Colocation, retrofits
*Sources: Uptime Institute, Arizton, Grand View Research, Delta Electronics.
Analysis
Liquid cooling is becoming essential for AI training and high-density inference workloads. Operators are investing in modular, scalable cooling architectures that can be retrofitted into existing facilities or deployed in new builds. Innovations in dielectric fluids, cold plates, and closed-loop systems are reducing water usage and improving energy efficiency. Patent activity and litigation in cooling and power systems are intensifying, reflecting the strategic importance of these technologies.
4.3 Network Architecture Evolution for AI Workloads
AI workloads are fundamentally reshaping data center network design:
Backend networks for AI training require ultra-high bandwidth (400–800 Gbps per link, moving to 1.6 Tbps), ultra-low latency, and lossless transport (InfiniBand, RoCE, Ultra Ethernet).
Rack-scale integration is emerging as the new unit of compute, with platforms like NVIDIA GB200 NVL72 and proprietary interconnects (NVLink, UALink) delivering exascale performance within a single rack.
Data center interconnects (DCI) are evolving to 400ZR/ZR+ and 800ZR/ZR+ modules, enabling high-speed, low-latency links between geographically dispersed campuses for distributed training and inference.
SmartNICs/DPUs are increasingly used for security, micro-segmentation, and offloading network functions, especially in multi-tenant and hybrid environments.
Table: Network Architecture Requirements by Workload
The shift to AI-centric workloads is driving a re-architecture of data center networks, with a focus on scale, bandwidth, and programmability. Open standards (e.g., UET, SONiC) are gaining traction, but proprietary solutions maintain performance advantages. Network-aware workload scheduling and distributed, federated training are emerging as key trends.
4.4 Power Infrastructure: HVDC, Microgrids, On-Site Generation, and Storage
High-voltage DC (HVDC) architectures (800 VDC, ±400 VDC) are being adopted to improve efficiency, reduce copper usage, and support megawatt-scale racks.
Microgrids and on-site generation (solar, wind, fuel cells, small modular reactors) are increasingly used to ensure reliability and meet sustainability goals.
Battery energy storage systems (BESS) provide fast-response backup and grid stabilization, enabling participation in demand response and grid services.
Power delivery innovations (solid-state transformers, vertical voltage regulators, e-Fuse modules) are improving efficiency and reliability at every stage of the power train.
Analysis
Power is now the primary constraint and differentiator in data center design. Operators are investing in resilient, scalable, and sustainable power architectures, often integrating renewables, storage, and advanced distribution systems. Regulatory and permitting delays, grid interconnection bottlenecks, and community opposition are significant challenges, especially in power-constrained regions.
4.5 Thermal Management and Water Usage Constraints
Water usage is a growing concern, with large data centers consuming millions of gallons per day for cooling and humidification. Water usage for cooling may increase by 870% in the coming years as more facilities come online.
Closed-loop and air-based cooling systems are being adopted to minimize water consumption, especially in arid regions and areas with water stress.
Heat reuse and waste heat recovery are being piloted in Europe and North America, providing district heating and reducing environmental impact.
Analysis
Sustainable water management is becoming a critical factor in site selection, permitting, and community relations. Operators are investing in water-efficient cooling technologies, closed-loop systems, and heat reuse to meet regulatory and ESG requirements.
5. Competitive Landscape and Industry Dynamics
5.1 Categorization of Players
The data center industry is increasingly stratified into five major categories, each with distinct strategic roles and infrastructure priorities:
Table: Data Center Category Comparison
Category
Typical Size
Workload Focus
Geographic Strategy
Infrastructure Approach
Key Differentiators
Hyperscale
50–1000+ MW
AI training, cloud, inference
Global, power-rich, rural/secondary
Custom, high-density, liquid cooling, HVDC
Scale, efficiency, innovation
Neo-Cloud
1–100 MW
GPUaaS, AI inference, HPC
Urban, edge, global
Modular, GPU-optimized, flexible
Speed, cost, specialization
Colocation
5–100+ MW
Hybrid IT, AI inference
Metro, interconnect hubs
Multi-tenant, high-density zones
Interconnection, flexibility
Enterprise
0.5–10 MW
Legacy, hybrid, regulated
On-prem, regional
Traditional, hybrid, edge nodes
Control, compliance, latency
Edge
<1–5 MW
AI inference, IoT, 5G
Distributed, near users
Modular, containerized, liquid/air
Latency, proximity, resilience
*Synthesized from industry reports, CoreSite, Azura, and market analyses.
5.2 Hyperscale Operators
Workload Focus: AI training, large-scale inference, cloud services, SaaS, and storage.
Geographic Expansion: Moving from saturated primary markets to secondary/rural regions with abundant power and land (e.g., Midwest U.S., India, Nordics).
Infrastructure Strategy: Custom-built, high-density campuses with advanced liquid cooling, HVDC, on-site renewables, and modular construction. Emphasis on energy efficiency (PUE 1.05–1.25), sustainability, and rapid scaling.
Evolution: Increasing focus on AI “factories,” sovereign AI mandates, and distributed, federated training across multiple campuses. Investing in proprietary silicon, open standards, and supply chain resilience.
5.3 Neo-Cloud Providers
Positioning: Specialized, GPU-centric clouds offering GPUaaS, AI inference, and HPC as a service. Compete on speed, cost, and flexibility, often serving workloads hyperscalers cannot accommodate quickly or cost-effectively.
Workload Focus: AI inference, model fine-tuning, GPU leasing, and edge deployments.
Infrastructure Choices: Modular, high-density, liquid-cooled racks; rapid deployment; direct interconnect to cloud and edge; focus on developer experience and transparent pricing.
Evolution: Expanding global footprint, partnering with hyperscalers, and integrating with edge and enterprise environments. Driving innovation in rack design, cooling, and workload orchestration.
5.4 Colocation Providers
AI-Ready Offerings: Upgrading facilities to support 30–70+ kW/rack, liquid/hybrid cooling, and direct cloud interconnection. Offering predictable cost structures, carbon reporting, and compliance certifications (ISO 50001, SOC 2, etc.).
Interconnection: Serving as the nexus for hybrid IT, enabling seamless workload migration between on-prem, cloud, and edge. Direct connections to cloud, network, and ecosystem partners are a key differentiator.
Pricing Power: Tight supply and high demand for high-density, AI-ready space are driving up lease rates and occupancy. Operators with available capacity and advanced infrastructure are capturing disproportionate value.
Evolution: Moving from “landlord” to strategic partner, offering managed services, AI hardware, and sustainability dashboards. Expanding into secondary markets and edge locations.
5.5 Enterprise Data Centers
On-Prem Trends: Enterprises are consolidating legacy data centers, repatriating select workloads from public cloud to colocation or on-prem for cost, control, and compliance reasons.
Hybrid IT: Blending on-prem, colocation, and cloud to optimize for performance, security, and regulatory requirements. Investing in edge nodes for latency-sensitive applications.
Repatriation: Motivated by cost, data sovereignty, and performance, but constrained by talent shortages and capital costs.
Evolution: Focus on automation, DCIM, and integration with cloud and edge. Some are building private AI clusters for proprietary workloads.
5.6 Edge Data Centers
Latency-Sensitive Inference: Supporting real-time AI inference, 5G, IoT, and AR/VR applications. Deployed in urban, industrial, and remote locations for ultra-low latency (sub-10 ms).
Distributed Patterns: Modular, containerized, and micro data centers with liquid cooling, renewable integration, and AI-optimized hardware.
Evolution: Integration with telecom networks, 5G central offices, and local microgrids. Emphasis on sustainability, resilience, and autonomous operation.
6. Infrastructure Supply Chain, Construction Costs, and Regulatory Constraints
6.1 Supply Chain and Construction Costs
Build Costs: $7–12 million per MW for hyperscale; $39 million per MW for advanced AI facilities (including $25 million/MW for compute, $4.3 million/MW for networking).
Lead Times: Equipment (transformers, switchgear) lead times of 8–24 months; construction timelines of 2–4 years, sometimes up to 6 years due to permitting and grid delays.
Land and Power: Land costs are rising, especially for large parcels (224 acres average for new campuses). Power availability is the primary gating factor for new builds.
6.2 Regulatory, Permitting, and Community Constraints
Permitting Delays: Multi-year environmental reviews, public hearings, and interconnection studies are common, especially in the U.S. and Europe.
Grid Constraints: Interconnection queues are backlogged, with over 2,600 GW of new generation and storage waiting in the U.S. alone. Some regions (e.g., Dublin, Amsterdam) have imposed moratoriums or quotas on new data center connections.
Community Pushback: Concerns over land use, noise, water, and power infrastructure are leading to local opposition and stricter zoning controls.
Sustainability Mandates: Increasing requirements for renewable energy sourcing, water efficiency, heat reuse, and carbon accounting. Operators must provide real-time ESG reporting and comply with regional and global standards (EU Green Deal, ISO 30134, etc.).
7. Sustainability Strategies: Renewables, PUE, Heat Reuse, and Carbon Accounting
7.1 Renewable Energy and Grid Integration
Renewable PPAs: Hyperscalers are the largest buyers of renewable energy, accounting for over 40% of all corporate PPAs (110+ GW contracted).
On-Site Generation: Increasing use of solar, wind, fuel cells, and small modular reactors (SMRs) for grid independence and carbon neutrality.
Grid-Friendly Operations: Participation in demand response, flexible load programs, and microgrid integration to support grid stability and resilience.
7.2 Power Usage Effectiveness (PUE) and Efficiency
PUE Trends: Industry average PUE is 1.55–1.6; leading-edge sites achieve 1.10–1.20 through advanced cooling, power management, and AI-based optimization.
AI-Driven Optimization: Predictive analytics and automation are reducing energy costs by 15–20% in data centers embracing AI-based energy management.
7.3 Heat Reuse and Water Management
Heat Reuse: Pilots in Europe and North America are redirecting waste heat to district heating, reducing carbon footprint and improving community relations.
Water Efficiency: Closed-loop cooling, air-based systems, and water reuse are minimizing water consumption, especially in arid and water-stressed regions.
7.4 Carbon Accounting and ESG Reporting
Real-Time Dashboards: Operators are providing tenant-level carbon reporting, Scope 1–3 emissions tracking, and compliance with ESG frameworks (ISO 50001, EU taxonomy).
Regulatory Compliance: Data centers must align with regional mandates (e.g., EU climate-neutral data centers by 2030, U.S. Energy Star, Asia data sovereignty laws).
8. Economic Models: Cost per Inference, Training TCO, and Pricing Dynamics
8.1 Cost per Inference and Training TCO
Inference Economics: Inference costs now dominate AI operational budgets, with per-query and per-token costs driving infrastructure decisions. Optimization techniques (quantization, pruning, knowledge distillation) are reducing compute requirements and costs.
Training TCO: Training remains a major capital expense, but efficiency gains (hardware, software, model design) are reducing per-model costs. However, the scale and frequency of new model training continue to drive demand for high-density, specialized infrastructure.
8.2 Pricing Dynamics
Colocation Pricing: Lease rates for high-density, AI-ready space are rising due to tight supply and high demand. Operators with available capacity and advanced infrastructure command premium pricing.
Cloud and Neo-Cloud: Transparent, flat-rate pricing and GPUaaS models are attracting enterprises seeking cost predictability and flexibility.
Edge and Enterprise: Pricing is driven by latency, compliance, and integration with cloud and colocation ecosystems.
9. Vendor and Technology Trends: GPUs, Accelerators, DPUs, and Open Standards
9.1 Hardware Evolution
GPUs: NVIDIA’s H100, H200, and Blackwell B200 are setting new benchmarks for AI training and inference, with power consumption rising from 700W (H100) to 1000W (B200) and memory bandwidth up to 8 TB/s.
ASICs and TPUs: Google’s TPUs, Amazon’s Inferentia and Trainium, and other custom accelerators are gaining share in inference workloads, offering 40–65% lower power consumption and 4× better cost-performance than GPUs for specific tasks.
DPUs and SmartNICs: Offloading network, security, and storage functions to DPUs is improving performance and security in multi-tenant and AI-centric environments.
9.2 Open Standards and Interoperability
Ultra Ethernet Consortium (UEC): Developing open transport protocols (UET) for AI and HPC workloads, targeting 1.6 Tbps and 1 million endpoints.
SONiC and Open Compute Project (OCP): Promoting open hardware and software standards for data center networking and power architectures.
Patent Landscape: Litigation and IP activity are intensifying in cooling, power, and control systems, with recent settlements and ongoing disputes shaping the competitive landscape.
10. Operational Implications: Staffing, Automation, and Site Reliability Engineering
10.1 Talent Shortage and Workforce Evolution
Staffing Crisis: The industry faces a severe shortage of skilled trade workers (electricians, HVAC, technicians), with 400,000 unfilled jobs in the U.S. in 2025, projected to reach 2 million by 2033.
Automation and Robotics: Adoption of AI-powered robotics, predictive maintenance, and DCIM tools is mitigating labor shortages and improving operational efficiency. The data center robotics market is projected to grow at a 21.6% CAGR, reaching $44.2 billion by 2030.
Site Reliability Engineering (SRE): SRE practices are being adopted to ensure uptime, automate incident response, and optimize resource utilization.
10.2 Automation and Predictive Maintenance
AI-Driven Operations: Predictive analytics, robotic process automation (RPA), and autonomous facility management are reducing downtime, improving efficiency, and enabling lights-out operations.
Security and Compliance: Enhanced physical and cyber security, micro-segmentation, and compliance automation are becoming standard in hyperscale and colocation environments.
11. Scenarios and Sensitivity Analysis Through 2035
11.1 Scenario Modeling: Low, Base, and High AI Adoption
AI dominates workloads, megawatt racks, modular/edge proliferation, grid constraints
*Synthesized from McKinsey, Goldman Sachs, Bain, and industry forecasts.
11.2 Sensitivity Factors
AI Model Efficiency: Advances in model design, quantization, and hardware efficiency could moderate infrastructure growth, but overall demand is expected to remain robust.
Regulatory and Grid Constraints: Permitting delays, grid bottlenecks, and community opposition could slow expansion, especially in power-constrained regions.
Sustainability and ESG: Stricter sustainability mandates and carbon accounting could drive innovation in renewables, heat reuse, and water management, but may also increase costs and complexity.
Economic and Geopolitical Factors: Interest rates, supply chain disruptions, and geopolitical tensions could impact capital flows, construction costs, and technology adoption.
The decade ahead will be defined by the convergence of cloud computing, AI training, and AI inference workloads, driving a fundamental transformation in data center infrastructure and strategy. Global capacity is set to more than triple by 2035, with AI workloads accounting for the majority of new demand. The industry is rapidly adopting high-density, liquid-cooled, and HVDC-powered architectures, while re-architecting networks for exascale AI and distributed inference. The competitive landscape is stratifying into hyperscale, neo-cloud, colocation, enterprise, and edge, each evolving to meet the unique demands of AI and hybrid IT. Sustainability, regulatory compliance, and operational efficiency are becoming strategic imperatives, as operators navigate power, water, and talent constraints. The winners in this new era will be those who secure power, innovate in cooling and network design, execute builds efficiently, and align with the evolving needs of AI-driven digital infrastructure. As the backbone of the information age, data centers are poised to offer both stability and growth, but only for those who can adapt to the relentless pace of technological and market change.
