Trend watch: Compute, chips & energy demand in 2026 — signals, winners, and red flags

Global data centre electricity consumption reached 460 TWh in 2025, roughly 1.8% of total global electricity demand, and the International Energy Agency projects it could exceed 800 TWh by 2028. For EU-based founders navigating this landscape, the tension between compute expansion and energy constraints is no longer a background concern. It is the central strategic variable shaping product roadmaps, infrastructure decisions, and competitive positioning.

Why It Matters

The compute-energy nexus sits at the intersection of three converging forces: explosive AI demand, semiconductor supply constraints, and tightening EU energy and climate regulation.

On the demand side, AI training compute has grown by roughly 4x annually since 2020, according to Epoch AI. A single GPT-4 scale training run consumed an estimated 50 GWh of electricity, equivalent to powering 4,500 EU households for a year. Inference workloads, which initially received less attention, now consume more total energy than training across the industry. Meta reported that inference accounted for approximately 70% of its AI compute energy in 2025, a ratio that grows as deployed models serve billions of daily queries.

On the supply side, semiconductor manufacturing itself has become an energy bottleneck. TSMC's fabrication facilities in Taiwan consumed 23.8 TWh in 2024, representing approximately 8% of Taiwan's total electricity consumption. The company's planned Arizona and Kumamoto facilities will add significant demand in their respective grids. Intel's planned Magdeburg, Germany fab is projected to require 1 GW of continuous power supply, comparable to a medium-sized city.

Regulation is accelerating the urgency. The EU's Energy Efficiency Directive requires data centres above 500 kW to report energy performance indicators starting in 2024, with binding efficiency targets expected by 2027. The Corporate Sustainability Reporting Directive (CSRD) mandates disclosure of digital infrastructure emissions for companies meeting reporting thresholds. The EU AI Act's transparency requirements include energy consumption reporting for high-risk AI systems. Collectively, these regulations create both compliance obligations and competitive differentiation opportunities for founders building energy-efficient compute solutions.

Key Concepts

Power Usage Effectiveness and Its Limitations

Power Usage Effectiveness (PUE), the ratio of total facility energy to IT equipment energy, has been the dominant data centre efficiency metric since its introduction by The Green Grid in 2006. The global average PUE improved from 2.0 in 2010 to approximately 1.58 in 2025, with hyperscale operators achieving 1.08-1.15. However, PUE has significant blind spots. It measures facility overhead efficiency but tells nothing about computational efficiency. A data centre with a PUE of 1.10 running workloads on legacy hardware at 15% utilisation is wasting far more energy than one with a PUE of 1.40 running optimised workloads at 80% utilisation. Emerging metrics including Carbon Usage Effectiveness (CUE) and Water Usage Effectiveness (WUE) provide complementary perspectives.

Chiplet Architecture and Energy Efficiency

The shift from monolithic chip designs to chiplet-based architectures represents a structural change in compute energy efficiency. AMD's EPYC processors use chiplet designs that achieve 40-60% better performance per watt than equivalent monolithic competitors. Intel's disaggregated approach with Foveros 3D packaging and EMIB interconnects enables mixing of process nodes, placing high-performance logic on leading-edge nodes while moving I/O and memory controllers to more efficient mature nodes. For founders, chiplet architectures mean that compute energy efficiency improvements are likely to continue at 25-35% per generation even as Moore's Law transistor scaling slows.

Specialised Silicon for AI Workloads

General-purpose GPUs are increasingly inefficient for production AI inference. Google's TPU v5p achieves approximately 2.5x better inference performance per watt than equivalent NVIDIA H100 configurations for transformer workloads. Startups including Groq, Cerebras, and SambaNova have developed purpose-built architectures that trade general flexibility for 3-10x efficiency gains on specific model architectures. The EU's European Processor Initiative (EPI) is developing RISC-V based accelerators targeting energy-efficient AI inference, with initial silicon expected in late 2026.

Liquid Cooling and Thermal Management

Air cooling, which accounts for 30-40% of data centre energy overhead, is reaching physical limits as rack power densities exceed 30 kW for AI-optimised configurations. Direct liquid cooling (DLC) and immersion cooling reduce cooling energy by 40-60% while enabling rack densities of 60-100+ kW. The EU's new data centre energy reporting requirements explicitly track cooling efficiency, creating regulatory tailwinds for liquid cooling adoption.

Sector-Specific KPI Benchmarks

Metric	Laggard	Average	Leader	Notes
Data Centre PUE	>1.6	1.3-1.6	<1.2	Hyperscale targets <1.10
AI Training Energy (per PFLOP-day)	>15 MWh	8-15 MWh	<8 MWh	Highly architecture-dependent
Inference Efficiency (queries/kWh)	<500	500-2,000	>2,000	Varies by model size
Renewable Energy Matching	<50% annual	50-80% annual	>90% 24/7 hourly	EU taxonomy alignment requires hourly
Water Usage (L/kWh)	>2.5	1.0-2.5	<0.5	Adiabatic/liquid cooling advantage
Server Utilisation Rate	<25%	25-50%	>65%	Key driver of true efficiency
Chip Performance/Watt Improvement	<15% YoY	15-30% YoY	>30% YoY	Architecture transitions drive jumps

Signals to Watch

Signal 1: Nuclear Power Agreements for Data Centres

Microsoft's 2024 agreement with Constellation Energy to restart the Three Mile Island Unit 1 reactor (835 MW) specifically for data centre power signalled a structural shift. Amazon Web Services followed with agreements for nuclear-powered data centres in Pennsylvania and Virginia. In the EU, France's EDF is reportedly in discussions with multiple hyperscalers for dedicated nuclear power supply. For founders, the signal is clear: large-scale AI compute is driving a new category of power procurement that bypasses traditional grid constraints entirely. Startups building nuclear-compute integration infrastructure, small modular reactor deployment services, or power purchase agreement platforms for dedicated generation have emerging tailwinds.

Signal 2: Custom Silicon Replacing General-Purpose GPUs

Amazon's Trainium2 chips, deployed in late 2025, deliver 4x better energy efficiency than previous generation for training workloads. Google's TPU v5p powers Gemini model training at approximately 60% of the energy cost of equivalent GPU clusters. Microsoft's Maia 100 accelerator, designed specifically for Azure AI workloads, began volume deployment in early 2026. The trend is unmistakable: every major cloud provider is investing billions in custom silicon to reduce compute energy costs. For EU founders, this creates opportunities in the supporting ecosystem: chip design tools, energy-aware workload orchestration, and performance benchmarking services.

Signal 3: EU Regulatory Framework Crystallising

The European Commission published its data centre energy efficiency benchmarking methodology in Q4 2025, establishing standardised reporting for PUE, WUE, renewable energy fraction, and server utilisation. Member states have until 2027 to transpose binding targets. Early signals suggest that PUE requirements of 1.3 or below for new facilities and minimum renewable energy fractions of 75% will be mandated. For founders, this regulatory clarity creates a compliance services market estimated at 2-4 billion euros annually by Gartner.

Winners and Losers

Emerging Winners

Liquid cooling providers: Vertiv, CoolIT Systems, and GRC are scaling rapidly as AI rack densities make air cooling impractical. European startup Submer has raised over 50 million euros and is deploying immersion cooling systems across EU data centres. The addressable market for data centre liquid cooling is projected to reach $15 billion by 2030, growing at 25% CAGR.

Energy-efficient chip designers: Companies that deliver superior performance per watt are winning design wins. Ampere Computing's Arm-based cloud processors achieve 2-3x better performance per watt than x86 alternatives for cloud-native workloads. Tenstorrent, led by chip legend Jim Keller, is developing RISC-V AI processors targeting 5x efficiency improvements.

Workload optimisation platforms: Run:ai (acquired by NVIDIA), Weights & Biases, and European startup Jina AI are building platforms that improve GPU utilisation from typical 25-35% to 60-80%, effectively doubling compute capacity without additional energy. This category addresses the single largest source of compute energy waste.

Red Flags

Unbounded energy demand projections: Some hyperscalers are projecting 2-3x increases in energy consumption over the next five years without credible efficiency offset plans. Founders building on these platforms should assess whether their cloud providers can deliver stable pricing and capacity, or whether energy constraints will create supply bottlenecks.

Greenwashing through unbundled RECs: Several data centre operators claim 100% renewable energy through purchased Renewable Energy Certificates that do not correspond to physical electricity delivery. The EU taxonomy requires temporal and geographic matching, meaning that annual REC purchases will not satisfy reporting requirements. Founders relying on provider sustainability claims should verify whether renewable matching is annual (weak) or hourly and locational (credible).

Stranded investment in air-cooled facilities: Data centres designed for 10-15 kW per rack air cooling cannot accommodate 40-100+ kW AI workloads without fundamental retrofit. Facilities built or financed in 2020-2023 without liquid cooling readiness face potential stranded asset risk as AI workloads grow. Colocation providers unable to offer high-density cooling may lose enterprise AI customers.

What's Working and What Isn't

What's Working

Google's 24/7 Carbon-Free Energy (CFE) programme achieved 64% hourly carbon-free matching across its global data centre fleet in 2024, with its Finland facility reaching 97%. The programme demonstrates that hourly renewable matching is technically feasible at scale, providing a template for EU regulatory compliance. The approach combines direct power purchase agreements, battery storage co-location, and real-time workload shifting to match compute demand with renewable availability.

Equinix's liquid cooling deployment across its EU facilities reduced cooling energy by 45% in retrofitted halls while enabling rack densities of 50 kW, a 3x increase over air-cooled configurations. The company reported that total cost of ownership decreased by 18% for AI workloads despite higher upfront capital costs for cooling infrastructure.

What Isn't Working

The global semiconductor industry's reliance on a single company (TSMC) for leading-edge fabrication creates systemic risk for compute energy efficiency. The most energy-efficient chip architectures require 3nm or 5nm process nodes available only from TSMC. Intel's foundry services remain 12-18 months behind on process technology, and Samsung's yields at advanced nodes continue to lag. EU efforts through the European Chips Act have allocated 43 billion euros but new fabrication capacity will not come online before 2028-2029.

Action Checklist

• Audit your cloud infrastructure's actual energy consumption per workload using provider carbon dashboards (AWS, Azure, GCP all offer these)
• Evaluate inference workloads for custom silicon migration that could reduce energy costs by 40-60%
• Assess data centre providers' liquid cooling roadmaps before signing multi-year colocation agreements
• Verify renewable energy claims against EU taxonomy requirements for hourly and locational matching
• Monitor EU data centre energy efficiency regulation timelines and prepare for reporting obligations
• Benchmark server utilisation rates and implement workload orchestration to target 60%+ utilisation
• Evaluate edge compute deployment for latency-tolerant workloads to reduce centralised data centre energy demand
• Include Scope 3 compute emissions in climate disclosures as CSRD requirements take effect

Sources

• International Energy Agency. (2025). Electricity 2025: Analysis and Forecast to 2028. Paris: IEA Publications.
• Epoch AI. (2025). Trends in Machine Learning Compute, 2020-2025. San Francisco: Epoch.
• European Commission. (2025). Data Centre Energy Efficiency Benchmarking Methodology: Technical Guidance. Brussels: EC.
• Uptime Institute. (2025). Global Data Center Survey 2025. New York: Uptime Institute.
• TSMC. (2025). 2024 Corporate Social Responsibility Report. Hsinchu: TSMC.
• Google. (2025). 2024 Environmental Report: 24/7 Carbon-Free Energy Progress. Mountain View: Google LLC.
• Gartner. (2025). Market Guide for Data Centre Sustainability and Energy Management. Stamford: Gartner Research.
• The Green Grid. (2025). Beyond PUE: Comprehensive Data Centre Sustainability Metrics Framework. Beaverton: The Green Grid.