New AI Tool Cuts Data Center CO2 by 45% and Extends Server Life by 1.6 Years

AI-generated Image (Credit: Jacky Lee)

A new AI-driven orchestration system that jointly tracks server “health” and electricity grid emissions could cut carbon dioxide from AI data centres by up to 45% over five years, according to simulations by researchers at the University of California, Riverside (UCR).

The framework, dubbed Federated Carbon Intelligence (FCI), models how AI workloads are routed across large fleets of accelerators. It combines real-time telemetry on hardware wear and tear with dynamic data on the carbon intensity of local power grids, then uses reinforcement learning to decide where and when to run inference jobs. The study appears in MRS Energy & Sustainability and was highlighted by UCR on 20 November 2025.

In simulations, FCI reduced cumulative CO₂ emissions by up to 45% versus a baseline scheduler and extended the average operational life of a server fleet by around 1.6 years, while maintaining performance constraints.

AI’s Growing Energy and Water Footprint

The work lands amid mounting concern over the energy and water use of AI infrastructure. The International Energy Agency (IEA) estimates that global data centres used about 240–340 TWh of electricity in 2022, roughly 1–1.3% of global final electricity demand, excluding cryptocurrency mining.

A newer IEA analysis focused on AI reports that data centres accounted for around 1.5% of global electricity consumption in 2024 (about 415 TWh), with usage growing several times faster than total electricity demand.

A separate study led by Cornell University, published in Nature Sustainability in November 2025, projects that the rollout of AI servers in the United States alone could add 24–44 million metric tons of CO₂-equivalent per year by 2030, alongside 731–1,125 million cubic metres of annual water use for cooling. That is roughly comparable to adding 5–10 million cars to U.S. roads and the household water use of 6–10 million Americans.

The Cornell team also finds that smarter siting, faster grid decarbonisation and operational efficiency measures could cut those carbon and water impacts by around 73% and 86%, respectively, under certain scenarios – underscoring the importance of software-driven optimisation alongside hardware upgrades and clean energy investments.

What FCI Actually Does

In the UCR work, Professors Mihri Ozkan (electrical and computer engineering) and Cengiz Ozkan (mechanical engineering) propose FCI as a “lifecycle-aware” orchestration layer for heterogeneous AI fleets.

According to their paper and university release, FCI has three main ingredients:

1. Hardware health modelling (SoH-AI):
   Telemetry from accelerators – for example temperatures, utilisation, and age-related degradation indicators – feeds a State-of-Health for AI hardware (SoH-AI) model. This estimates how quickly each device is wearing out and how that affects its energy efficiency over time.

2. Carbon-aware emissions profiling:
   The system ingests grid carbon-intensity and cooling data for different regions, using sources such as Electricity Maps and WattTime to reflect when and where electricity is cleaner or dirtier, and how cooling efficiency varies by location.

3. Reinforcement learning-based scheduler:
   A reinforcement learning (RL) agent uses these signals, plus workload metadata (latency sensitivity, throughput needs), to route inference jobs across accelerators and regions. The objective is to jointly minimise cumulative emissions and hardware degradation while respecting performance targets.

The authors simulate a global fleet of accelerators, including NVIDIA GPUs, Google TPUs and Cerebras WSE-2 chips, operating over a five-year deployment window, with realistic diurnal load patterns and region-specific grid data. Rather than turning hardware off, FCI modulates utilisation and shifts jobs in time and space to spread thermal and mechanical stress more evenly and exploit lower-carbon power windows.

How Big Are the Gains?

Within that simulated environment, FCI delivers several headline benefits versus a conventional, performance-only scheduler:

• Up to 45% reduction in cumulative CO₂ emissions over five years, driven by shifting workloads toward lower-carbon grid periods and regions, and by limiting efficiency losses from heavily worn hardware.

• Average fleet life extension of about 1.6 years, as the algorithm avoids repeatedly overloading ageing or thermally stressed devices.

• Lower embodied emissions, because slower hardware turnover means fewer new accelerators need to be manufactured and installed during the simulated period.

The results are still based on modelling rather than live deployments. The UCR team emphasises that real-world validations with cloud providers are a necessary next step and that performance will depend on the accuracy of telemetry and grid forecasts in production settings.

How FCI Compares With Other “Carbon-Aware” Systems

FCI enters a growing ecosystem of tools aimed at making AI infrastructure less carbon intensive. Several prominent efforts include:

• Google’s carbon-intelligent computing
  Google has described a system that shifts flexible workloads in time and across regions based on real-time and forecast grid carbon intensity. Early results published by the company suggest this can increase the share of low-carbon electricity used by its data centres, for example by scheduling some batch workloads for times when wind or solar output is higher.

• Clover
  Clover: Toward Sustainable AI with Carbon-Aware Machine Learning Inference Service proposes a runtime that reduces emissions by dynamically selecting between mixed-quality models and partitioning GPU resources, trading small amounts of accuracy or performance for lower energy use while meeting service-level agreements. The authors report that Clover can “substantially” reduce carbon emissions for ML inference workloads under their experimental setups.

• EcoServe
  EcoServe: Designing Carbon-Aware AI Inference Systems focuses on large-language-model serving and combines resource provisioning with carbon-aware scheduling guided by a four-step “Reduce, Reuse, Rightsize, Recycle (4R)” principle. According to its authors, EcoServe can lower total (operational plus embodied) carbon emissions by up to 47% relative to several performance- and cost-optimised baselines, largely by better utilising host systems and delaying hardware refreshes.

• CO2 AI and related reporting platforms
  Platforms such as CO2 AI, developed with Boston Consulting Group, help large companies quantify and track emissions across supply chains and operations, including IT. Unilever, for example, has used CO2 AI to map emissions and identify decarbonisation levers. These tools, however, focus on footprint accounting and scenario analysis rather than live workload orchestration.

The FCI authors position their framework as complementary to this landscape. In their analysis, many existing schedulers emphasise grid-aware timing or provisioning, but treat hardware performance as static. FCI explicitly couples real-time degradation models with grid signals, aiming to minimise both operational and embodied emissions in a single optimisation loop.

From Lab Simulations to Policy and Practice

The UCR work aligns with a broader shift in research and policy thinking: that AI’s environmental footprint must be tackled at multiple layers at once – location, grid mix, hardware design, and scheduling software. Recent IEA and academic reports stress the need for both efficiency improvements and demand-flexibility from data centres if electricity systems are to cope with AI growth while staying on track for climate targets.

The Cornell Nature Sustainability “roadmap” suggests that combining smart siting, grid decarbonisation and operational optimisations can take a large bite out of projected AI server emissions and water use, but that no single approach is sufficient. In that context, FCI is one more piece of the puzzle: a software-only, telemetry-driven approach that could, in principle, be integrated into existing orchestration stacks like Kubernetes or cloud provider schedulers without new hardware.

For now, the results remain simulation-based, and key questions around data quality, implementation cost, and incentives for cloud operators and tenants are unresolved. But as energy agencies, regulators and cloud companies grapple with how to keep AI growth compatible with climate commitments, tools that merge carbon awareness with hardware longevity are likely to feature more prominently in both research agendas and future best-practice guidelines.

License This Article

Source: UCR, Springer Nature Link, IEA, Nature, Cornell University, Google Blog, arXiV, Economist Group

$5.00

$10.00

$20.00

$30.00

Custom Amount

Please enter an amount

$

One-Time Donation Monthly Donation Quarterly Donation Annual Donation

3% Cover the Fee

Support us to keep delivering insightful content on AI development. Your contribution fuels our mission to inspire and inform! Support us to keep delivering insightful content on AI development. Your contribution fuels our mission to inspire and inform!