Deep dive: Compute, chips & energy demand — what's working, what's not, and what's next

The global data centre industry consumed approximately 460 TWh of electricity in 2025, equivalent to the entire electricity consumption of France. That figure is projected to reach 800 to 1,000 TWh by 2030, driven almost entirely by the explosive growth in AI training and inference workloads. For investors in emerging markets, this trajectory creates both enormous risks and exceptional opportunities: the semiconductor and data centre value chains are restructuring around energy constraints that will determine which companies, technologies, and geographies capture the next decade of growth.

Why It Matters

The relationship between compute and energy has shifted from a secondary operational concern to the primary constraint on industry growth. NVIDIA shipped over 3.8 million H100 GPUs in 2024, each consuming 700 watts at peak load. A single AI training cluster with 10,000 H100 GPUs requires 7 MW of continuous power, equivalent to a small town. Meta's Llama 3 training run consumed an estimated 39 GWh of electricity, roughly equal to the annual consumption of 13,000 US households. These figures represent a fundamental departure from the historical trend where compute efficiency gains outpaced demand growth.

For emerging markets, the implications are particularly acute. India's data centre capacity grew by 35% in 2024, with 1.2 GW under construction across Mumbai, Chennai, and Hyderabad. Southeast Asia added 500 MW of capacity in the same period, with Malaysia, Indonesia, and Thailand competing for hyperscaler investment. Yet grid infrastructure in these regions often cannot reliably supply the power densities that modern AI workloads demand. India experienced over 1,200 grid-level power quality events in 2024 that affected data centre operations, while parts of Malaysia's Johor state face generation adequacy concerns as data centre power demand approaches 2 GW.

The investment landscape reflects these dynamics. Data centre construction spending in emerging markets reached $28 billion in 2025, up from $18 billion in 2023. Semiconductor capital expenditure globally exceeded $180 billion, with TSMC alone committing $32 billion for advanced process nodes. Investors who understand which technologies genuinely reduce the energy intensity of compute, and which geographies can reliably supply that energy, will be best positioned to capture returns from the continued build-out.

Key Concepts

Power Usage Effectiveness (PUE) measures total data centre energy consumption divided by IT equipment energy consumption. A PUE of 1.0 would mean all energy goes directly to computing; real-world facilities add cooling, lighting, power conversion, and infrastructure overhead. The global average PUE in 2025 was approximately 1.55, meaning 55% additional energy was consumed beyond what the IT equipment itself used. Best-in-class hyperscale facilities achieve PUEs of 1.08 to 1.12, while legacy enterprise facilities often operate above 1.8.

Thermal Design Power (TDP) represents the maximum sustained power consumption of a processor under realistic workloads. GPU TDP has increased from 250W (NVIDIA A100, 2020) to 700W (H100, 2023) to 1,000W (B200, 2024), with next-generation chips projected to reach 1,200 to 1,500W. This trajectory drives fundamental changes in cooling infrastructure: air cooling becomes impractical above approximately 40 kW per rack, forcing the transition to liquid cooling that most existing facilities cannot accommodate without major retrofits.

Performance per Watt is the critical efficiency metric, measuring useful computation delivered per unit of energy consumed. NVIDIA's H100 delivers approximately 3.9 PFLOPS per watt for AI inference, compared to 1.3 PFLOPS per watt for the A100. This 3x improvement partially offsets the absolute power increase, but only if workloads scale proportionally. In practice, the industry has responded to efficiency gains by scaling workloads faster than efficiency improves, a dynamic economists call the Jevons paradox applied to compute.

Embodied Carbon accounts for the greenhouse gas emissions generated during chip manufacturing, packaging, and transport. TSMC's fabrication of a single advanced 3nm wafer requires over 10,000 litres of ultrapure water and generates approximately 15 to 20 kg of CO2-equivalent emissions per wafer. When amortised across the useful life of the chips, embodied carbon can represent 20 to 40% of total lifecycle emissions for energy-efficient facilities, a share that grows as operational energy becomes cleaner.

Compute Energy Demand KPIs: Benchmark Ranges

Metric	Below Average	Average	Above Average	Top Quartile
Data Centre PUE	>1.6	1.4-1.6	1.2-1.4	<1.2
GPU Utilisation Rate	<30%	30-50%	50-70%	>70%
Renewable Energy Procurement	<25%	25-50%	50-80%	>80%
Water Usage Effectiveness (WUE)	>2.5 L/kWh	1.5-2.5 L/kWh	0.8-1.5 L/kWh	<0.8 L/kWh
Carbon Intensity (Scope 2)	>400 gCO2/kWh	250-400 gCO2/kWh	100-250 gCO2/kWh	<100 gCO2/kWh
Cooling Energy as % of Total	>35%	25-35%	15-25%	<15%
Inference Efficiency (TOPS/W)	<100	100-200	200-400	>400

What's Working

Liquid Cooling Adoption for High-Density AI Workloads

Direct liquid cooling (DLC) and immersion cooling have transitioned from niche technology to mainstream deployment for AI workloads. Equinix deployed liquid cooling in over 60 facilities globally by end of 2025, with particular concentration in its Singapore, Tokyo, and Mumbai locations. Microsoft's liquid-cooled Azure AI clusters in Southeast Asia achieve PUEs below 1.10, reducing cooling energy by 40 to 50% compared to air-cooled equivalents. For emerging market facilities where ambient temperatures regularly exceed 35 degrees Celsius, liquid cooling eliminates the performance derating that forces air-cooled systems to throttle during peak heat periods. The capital premium for liquid cooling has fallen from 25 to 30% in 2022 to 10 to 15% in 2025, with operational savings recouping the investment within 18 to 24 months in tropical climates.

Custom Silicon for Inference Efficiency

The shift from general-purpose GPUs to application-specific inference accelerators is delivering dramatic efficiency improvements. Google's TPU v5p achieves 2.5x better performance per watt than equivalent GPU configurations for transformer model inference. Amazon's Trainium2 chips, manufactured on a 3nm process, deliver inference at approximately 40% of the energy cost of comparable NVIDIA hardware for specific model architectures. In emerging markets, where electricity costs directly affect data centre economics, these efficiency gains translate to 15 to 25% reductions in total cost of ownership. Startups like Groq and Cerebras have demonstrated that purpose-built architectures can achieve 5 to 10x inference efficiency improvements over general-purpose GPUs for certain workloads, attracting significant investor interest.

Hyperscaler Renewable Energy Procurement in Emerging Markets

Major cloud providers have catalysed renewable energy development in emerging markets through long-term power purchase agreements (PPAs) and direct investment. Google's 500 MW solar PPA in Tamil Nadu, India, represents one of the largest corporate renewable energy deals in Asia. Microsoft's agreement to procure 900 MW of wind and solar capacity across India and Southeast Asia by 2027 has brought institutional-grade project finance to markets where it was previously scarce. These PPAs provide revenue certainty that enables renewable developers to build projects they could not otherwise finance, creating a positive feedback loop between data centre growth and clean energy deployment. In 2025, hyperscaler PPAs accounted for approximately 30% of all new renewable energy capacity contracted in India and 25% in Southeast Asia.

What's Not Working

Grid Infrastructure Gaps in High-Growth Markets

The pace of data centre construction in emerging markets has outstripped grid development in several critical regions. Jakarta's data centre cluster in the Cibitung corridor has experienced repeated power quality issues as total demand approached 800 MW on a transmission network designed for 500 MW of industrial load. In India, the Maharashtra Electricity Regulatory Commission flagged concerns about the strain of 2.5 GW of approved data centre capacity on Mumbai's grid, which already operates near capacity during summer peaks. The fundamental challenge is temporal: data centres can be constructed in 18 to 24 months, while transmission infrastructure upgrades require 4 to 7 years from planning to commissioning. This mismatch creates stranded capacity risk for data centre investors and grid reliability risk for host communities.

The Rebound Effect on Total Energy Consumption

Despite impressive per-chip efficiency gains, total industry energy consumption continues to accelerate. NVIDIA's Blackwell architecture delivers 4x the training performance per watt of its predecessor, but NVIDIA's total GPU power consumption across shipped units grew from approximately 12 TWh in 2023 to an estimated 25 TWh in 2025. The industry has consistently demonstrated that efficiency improvements lower the cost of compute, which stimulates demand that more than offsets the efficiency gain. For investors, this means that semiconductor efficiency improvements, while valuable, do not reduce aggregate energy demand and should not be modelled as doing so. Total data centre electricity consumption in emerging markets grew by approximately 28% annually between 2023 and 2025, far exceeding the 8 to 12% improvement in average performance per watt.

Water Consumption in Water-Stressed Regions

Many of the emerging markets attracting data centre investment face acute water stress. Evaporative cooling towers, which achieve the lowest PUE values, consume 1.5 to 2.5 litres of water per kWh of IT load. A 100 MW data centre using evaporative cooling in central India may consume over 1 million litres of water daily, competing directly with agricultural and municipal needs in regions already experiencing water deficits. Google, Microsoft, and Meta have all faced community opposition to data centre projects in water-stressed regions of India and Southeast Asia. The industry is responding by shifting toward air-cooled and liquid-cooled designs that eliminate evaporative water use, but these alternatives increase energy consumption by 5 to 15%, creating a direct tradeoff between water conservation and energy efficiency.

What's Next

On-Site and Near-Site Power Generation

The limitations of grid infrastructure are driving data centre operators toward dedicated power generation. Microsoft's agreement to purchase nuclear power from Constellation Energy for its data centres signals a broader industry trend toward securing dedicated, 24/7 clean power sources. In emerging markets, this translates to co-located solar-plus-storage installations, small modular reactors (where regulatory frameworks permit), and dedicated natural gas generation with carbon capture. Tata Communications' 100 MW solar-powered data centre campus in Pune, India, demonstrates the emerging model: purpose-built renewable generation sized to match data centre load, with grid connection serving as backup rather than primary supply. This approach sidesteps grid constraints while providing cost certainty and emissions reductions, though it requires significantly more capital and land than grid-connected facilities.

Chiplet Architectures and Advanced Packaging

The semiconductor industry's transition from monolithic chips to chiplet-based designs offers substantial energy efficiency benefits. AMD's MI300X uses chiplets to integrate CPU and GPU compute with high-bandwidth memory in a single package, reducing data movement energy by 30 to 40% compared to discrete multi-chip systems. Intel's Foveros and TSMC's CoWoS advanced packaging technologies enable heterogeneous integration that matches different process nodes to different functions, placing high-performance compute on leading-edge 3nm while running I/O and memory controllers on more energy-efficient 7nm or 12nm processes. For investors, the advanced packaging segment represents one of the highest-growth areas in semiconductors, with the market projected to grow from $44 billion in 2024 to over $85 billion by 2028.

Edge AI and Distributed Inference

Moving inference workloads from centralised data centres to edge devices fundamentally changes the energy equation. Qualcomm's Snapdragon X Elite and Apple's M4 chips demonstrate that many AI inference tasks can run on devices consuming 10 to 30 watts rather than in data centres consuming megawatts. For emerging markets with constrained grid infrastructure, edge AI reduces the need for massive centralised data centres while improving latency for local applications. The Indian government's initiative to deploy AI inference nodes at 500 locations across the country, using locally manufactured hardware, exemplifies how distributed compute can address both energy constraints and digital sovereignty concerns. Edge inference is projected to handle 40 to 50% of all AI inference workloads by 2028, up from approximately 15% in 2025.

Carbon-Aware Computing and Workload Scheduling

Software-level optimization represents the lowest-cost path to reducing compute-related emissions. Carbon-aware scheduling shifts flexible workloads (model training, batch processing, data analytics) to times and locations where the electricity grid has the lowest carbon intensity. Google's carbon-intelligent computing platform demonstrated 30% reductions in carbon intensity for flexible workloads by shifting computation to data centres with high renewable generation. In emerging markets with significant solar penetration, this means scheduling compute-intensive tasks during daylight hours when solar generation peaks. The approach requires no hardware changes and can be implemented through middleware layers that are increasingly available as open-source tools.

Action Checklist

• Assess portfolio exposure to data centre energy constraints across emerging market investments, focusing on grid capacity and reliability in target geographies
• Evaluate semiconductor investments through a performance-per-watt lens, prioritising companies with clear energy efficiency roadmaps over raw performance improvements
• Due-diligence data centre investments for water risk, particularly in India, Southeast Asia, and the Middle East where water stress is intensifying
• Monitor regulatory developments around data centre power consumption caps, which Singapore, Ireland, and the Netherlands have already implemented and other markets may follow
• Track hyperscaler PPA activity as a leading indicator of renewable energy market maturation in emerging economies
• Analyse the advanced packaging supply chain for investment opportunities, including TSMC's CoWoS, Intel's Foveros, and emerging competitors
• Evaluate edge AI hardware companies positioned to capture the shift from centralised to distributed inference
• Consider the embodied carbon implications of semiconductor manufacturing when assessing lifecycle emissions of compute infrastructure

FAQ

Q: How fast is data centre energy consumption growing in emerging markets? A: Data centre electricity consumption in major emerging markets (India, Southeast Asia, Latin America, Middle East) is growing at 25 to 30% annually, roughly double the global average growth rate. India alone is projected to increase data centre power capacity from 1.8 GW in 2024 to over 4.5 GW by 2028. This growth rate creates both significant investment opportunities in power infrastructure and meaningful risks around grid reliability and community acceptance.

Q: Will chip efficiency improvements offset the growth in AI energy demand? A: No. Historical evidence and current trajectories strongly suggest that efficiency improvements stimulate demand growth that exceeds the efficiency gain. NVIDIA's Blackwell architecture delivers 4x better training efficiency per watt, but the company's total shipped GPU power consumption roughly doubled between 2023 and 2025. Investors should model chip efficiency as expanding the addressable market for compute rather than reducing aggregate energy consumption.

Q: What are the biggest risks for data centre investors in emerging markets? A: The three primary risks are grid infrastructure adequacy (power supply cannot meet approved data centre capacity), water availability (particularly for facilities using evaporative cooling in water-stressed regions), and regulatory uncertainty (several jurisdictions are considering power consumption caps or mandatory renewable energy requirements for data centres). Investors should conduct detailed grid capacity assessments, water stress analyses, and regulatory horizon scanning for each target market.

Q: How should investors evaluate the sustainability claims of data centre operators? A: Focus on three verifiable metrics: actual PUE (measured, not designed), renewable energy procurement structure (24/7 carbon-free energy matching versus annual certificate-based claims), and water usage effectiveness. Operators claiming 100% renewable energy through unbundled renewable energy certificates are not reducing emissions if they operate on carbon-intensive grids during periods without renewable generation. The emerging standard of 24/7 carbon-free energy matching provides a more rigorous framework for evaluating genuine emissions reductions.

Sources

• International Energy Agency. (2025). Data Centres and Data Transmission Networks: 2025 Energy Consumption Report. Paris: IEA Publications.
• BloombergNEF. (2025). Data Centre Energy Demand Outlook: Emerging Markets Focus. New York: Bloomberg LP.
• Semiconductor Industry Association. (2025). Global Semiconductor Industry Factsheet: Capital Expenditure and Capacity Trends. Washington, DC: SIA.
• Uptime Institute. (2025). Global Data Centre Survey 2025: Efficiency, Sustainability, and Resilience. New York: Uptime Institute.
• NVIDIA Corporation. (2025). Sustainability Report FY2025: Energy Efficiency and Environmental Impact. Santa Clara, CA: NVIDIA.
• India Brand Equity Foundation. (2025). Indian Data Centre Market: Growth Drivers and Infrastructure Challenges. New Delhi: IBEF.
• Google. (2025). Environmental Report 2025: Carbon-Free Energy and Data Centre Sustainability. Mountain View, CA: Alphabet Inc.