Data Centers' Energy Footprint Highlights Climate & Health Costs, Spurs Sustainable Growth & Exascale HPC

TL;DR

• Data Center Energy Use Drives Climate Impact and Public Health Costs.
• Advanced Cooling and Expansion Drive Sustainable Data Center Growth.
• Exascale Supercomputers and High-Performance CPUs Push HPC Frontiers.

Data‑Center Energy Surge: Climate and Health Stakes

Rising Energy and Water Use

• Electricity consumption climbed from 12.5 TWh in 2019 to 25 TWh in 2023 – a 100 % increase, equivalent to powering ~2.4 million U.S. homes.
• Projected 2026‑2028 demand reaches 30 TWh per year (≈ 2.9 million homes), driven by AI‑intensive workloads.
• Carbon emissions rose from 1.1 Mt CO₂ (2019) to 2.2 Mt CO₂ (2023); forecast 3.5 Mt CO₂ by 2028 if current growth persists.
• On‑site water use grew from 1 090 to 2 302 acre‑ft (≈ 113 % increase); 2026‑2028 projection exceeds 3 500 acre‑ft.
• Peak power demand per AI‑driven data center now exceeds 1 GW, with a 12 % annual global rise; 2‑3× current demand anticipated by 2028.

Public‑Health Cost Attribution

• Air‑pollution‑related health expenditures rose from $45 M (2019) to $155 M (2023).
• If emissions continue on the current trajectory, annual health costs could reach $266 M by 2028.
• Diesel backup generators contribute episodic spikes in NOₓ and SOₓ, amplifying acute respiratory risks near high‑density data‑center clusters.

Policy Landscape and Gaps

• California’s veto of mandatory water‑use disclosure limits public reporting; only cost‑increase assessments were approved.
• State‑level consumer‑environmental bills have stalled in several legislatures, reducing enforceable intensity limits.
• FERC’s recent request for utility feasibility data increases reporting obligations for new grid connections.
• Mixed outcomes across states: Texas passed an emergency‑standard water‑use bill; Pennsylvania and New Jersey proposals remain pending.

Emerging Mitigation Strategies

• Waste‑heat recovery pilots (e.g., UK’s HeatHub) convert ~1 kW per module into residential heating, cutting household bills by ~90 % and covering 2.5 % of national electricity demand by 2030.
• Submerged data‑centre concepts (Microsoft’s Project Natick) achieve ~30 % lower cooling energy per compute unit; additional Chinese proposals aim to scale the approach.
• Hybrid renewable procurement contracts (Google‑NextEra, Duane Arnold nuclear restart) target 24/7 carbon‑free electricity, mitigating peak‑demand emissions.

Forecast and Action Priorities

• By 2028, U.S. data‑center electricity demand is likely to exceed 30 TWh annually, requiring 2–3 GW of firm capacity each year.
• Absent systematic waste‑heat capture or baseload renewable contracts, sector emissions could surpass 4 Mt CO₂, raising health costs above $300 M per year.
• Regulatory response is expected to tighten water‑use reporting and enforce emission caps, driven by quantified health externalities.
• Investment focus should accelerate heat‑recovery modules (target 100 MW thermal capacity by 2026) and expand submerged‑facility pilots to validate long‑term operational viability.

Advanced Cooling and Expansion: Technical Foundations for Sustainable Data‑Center Growth

Energy‑Use Baseline and Projections

• 2023 electricity consumption: 25 TWh (≈2.4 M U.S. homes)
• Projected 2026‑2028 consumption: 30 TWh, a 20 % increase
• 2023 carbon emissions: 2.2 Mt CO₂; projected 2.8 Mt CO₂ (+27 %)
• On‑site water use in 2023: 2 302 acre‑ft; projected 5 000 acre‑ft if cooling practices remain unchanged
• Public‑health cost of air pollution in 2023: $155 M; projected $266 M by 2028

These figures reflect the growing energy‑intensity of data‑centers, with a 12 % annual rise in Manitoba and a global demand expected to double by 2030. Cooling accounts for roughly 40 % of Canadian data‑center energy consumption, establishing it as a primary constraint on expansion.

Emerging Cooling Technologies

• Closed‑loop water cooling: Recirculates chilled water, reducing PUE by 0.05‑0.10. Widely deployed in California and Canada.
• Aquifer Thermal Energy Storage (ATES): Utilises 50 °F groundwater, achieving up to 74 % carbon‑emission reduction and 30 % cooling‑energy savings. Pilot projects operating in the United States.
• Heat‑capture for district heating (HeatHub): Extracts waste heat from low‑power servers, cutting household heating costs from £375 to £40 per month. Fifty pilot installations in the UK.
• Submarine/immersive cooling: Submerges sealed servers in seawater, delivering a 10 % PUE improvement with negligible land‑water use. Tested on 850 servers (2018‑2020); commercial rollout planned by Chinese firms.

Expansion Drivers and Grid Implications

• AI workload growth has quadrupled rack power density; Q3 2025 AI‑infra CAPEX reached $100 B.
• Regional power scarcity in Texas and Manitoba creates >2 GW of demand‑response requirements for upcoming AI data‑centers.
• Large capital projects (e.g., a $15 B complex) can triple electricity demand within weeks of commissioning, raising the risk of stranded generation assets.
• On‑site diesel or fuel‑cell backup mitigates peak‑load exposure but adds localized air‑quality impacts.

Policy Landscape

• California: No mandatory water‑use disclosure, limiting operator incentives for low‑water cooling.
• Manitoba: Renewable‑power incentives attract high‑intensity AI workloads yet forecasts 180‑500 Mt CO₂e by 2035.
• UK Power Networks Shield initiative: Grants for heat‑recovery pilots, enabling district‑heating integration.
• U.S. FERC: Requests utility feasibility data for large data‑center projects, improving grid‑planning accuracy while adding compliance overhead.

Forecast and Strategic Recommendations

• Electricity demand will exceed 30 TWh by 2028, representing ~2.5 % of total U.S. electricity consumption.
• Water consumption could surpass 5 000 acre‑ft by 2028 without efficiency gains.
• Implementing ATES or immersive cooling in suitable geologies can reduce net cooling electricity by 15‑30 %, saving 4‑6 TWh annually.
• Heat‑recovery pilots can offset 2‑3 % of residential heating demand in target metros by 2030, delivering up to £335 M in household‑energy cost reductions (UK case).
• Prioritize closed‑loop water and ATES for new builds, mandate real‑time water‑use reporting, embed demand‑response and on‑site renewable generation, and accelerate heat‑capture pilots under utility‑sponsored schemes.

Exascale Is No Longer a Prototype: The New Reality of Heterogeneous, Liquid‑Cooled HPC

From Single‑Site Experiments to Modular Blade Farms

The November 2025 data stream confirms that exascale systems have moved beyond isolated testbeds into production‑grade, modular platforms. HPE’s Cray GX440n and GX350a blades combine four NVIDIA Vera and Rubin GPUs with AMD EPYC Venice CPUs, all housed in liquid‑cooled Slingshot chassis. Aurora’s 80 k Xeon cores and 60 k GPUs, linked by Cray Slingshot‑11 at 400 Gbps, deliver more than 2 EFLOPS peak performance with a 1 TB/s cross‑node bandwidth. The modular ExaPod AI pod, featuring 80‑core Xeon 6781P CPUs, 24 PCIe Gen5 slots and 24 NVMe drives, packs 36 PiB usable storage per rack while maintaining a 900 W / PiB power density.

Network Bandwidth Becomes the Baseline

The 400 Gbps Slingshot fabric, now available in 16‑, 64‑ and 128‑port configurations, is scaling toward 2 k‑port topologies for clusters exceeding 1 EFLOPS. This bandwidth growth is a direct response to the densification of GPUs; any system aiming for exascale AI training now expects 400 Gbps fabrics as a minimum. Expect the next generation to double lane capacity to 800 Gbps, with 256‑port switch modules becoming the norm.

Power and Cooling Define the Scaling Ceiling

Rack power consumption has surged from roughly 200 kW in 2020 to nearly 800 kW in 2025, rivaling the electricity use of a midsize city. Liquid‑cooled blades are no longer optional; they are the de‑facto standard for any heterogeneous deployment beyond 500 PFLOPS. Direct‑liquid and immersion cooling methods are gaining traction as the primary lever to keep thermal envelopes within manageable limits.

CPU and Accelerator Evolution

Intel’s Granite Rapids‑WS pushes workstation core counts to 128 with a 4.80 GHz boost, while AMD’s Zen 6 EPYC Venice chips reinforce the high‑core‑count trend. However, the real performance leap stems from silicon‑level stacking—Intel EMIB/Foveros and TSMC CoWoS enable CPU‑GPU die integration, slashing I/O latency and cutting PCB real‑estate by roughly 30 %. This packaging shift is essential for maintaining power efficiency as GPU density climbs.

Software Gains from Architecture‑Specific Innovations

ARM’s Scalable Matrix Extension (SME) – showcased by the LOOPS framework – delivers up to a 14.4× speed‑up on sparse‑matrix kernels with markedly lower energy per operation. HPC libraries are rapidly adapting to exploit SME, hinting at a broader move toward architecture‑tailored acceleration for graph‑based AI and scientific simulations.

What Lies Ahead

Within the next twelve months, liquid‑cooled heterogeneous blades will dominate any system above 500 PFLOPS, network fabrics will standardize at 800 Gbps per lane, and modular ExaPod‑style reference designs will be commercialized by multiple OEMs. CPU core counts will plateau near 128 for workstation silicon, while server‑class processors focus on higher clock rates and DDR5‑7200 memory to satisfy memory‑bound AI workloads. Advanced packaging will become a prerequisite, delivering >2× inter‑die bandwidth improvements, and ARM SME‑driven sparse kernels will be woven into mainstream HPC stacks. These converging forces reshape the performance‑per‑watt ceiling, positioning exascale HPC for the AI‑driven scientific breakthroughs of the next decade.