PaperlayerView exact revision on arXiv

Source-linked AI summary

Toward a Time-Aware Assessment Framework for the Carbon Cost of AI-Enabled Decarbonization

Chenrui Xu, Burcu Akinci, Christopher McComb

arXiv:2609.18029v1cs.CY

TL;DR

Existing AI-for-decarbonization assessments often omit AI-side emissions and overlook the different timing of AI costs and physical-system benefits. This paper introduces a finite-horizon, time-indexed framework and shows that discounting can reverse intervention rankings, while supporting practical screening and timing decisions. Its stylized scenarios and deterministic attributional assumptions limit benchmark and real-world interpretation.

  • Problem

    Existing assessments often omit AI-induced emissions and rarely account for timing differences between AI costs and decarbonization benefits.

  • Method

    The paper models avoided emissions S(t) and AI-induced emissions C(t) as discrete-time streams over a finite horizon under explicit discounting.

  • Results

    Discounting changed preferred rankings across representative interventions and produced ranking reversals at break-even discount rates.

  • Takeaways & Limitations

    NEPV, discounted payback, timing comparisons, and return-on-carbon thresholds support transparent go/no-go and deployment decisions.

  • Takeaways & Limitations

    The scenarios are stylized rather than rigorous benchmark evaluations, and results depend on attributional assumptions, time-varying emissions factors, and realized savings.

Abstract

from arXiv · show

AI is increasingly used to support decarbonization decisions across the built environment, yet the development, training, and use of AI consume energy and induce CO2e emissions. However, existing assessments often report physical-system savings while omitting AI-side emissions. Moreover, they rarely account for the mismatch between when AI costs occur and when decarbonization benefits materialize, which may be substantial for infrastructure-scale projects. To address these issues, we present a time-aware assessment framework that models avoided emissions and AI-induced emissions as discrete-time streams over a finite time horizon. In demonstrating this process, we seek to show that time-aware assessment can support temporal decision-making, identify cases in which accounting for time value of carbon can change preferred rankings relative to time-invariant totals, and explore how decisions may vary with slightly different governance priorities. Using four representative interventions with intentionally different temporal profiles (multi-project low-carbon concrete design support, AI-assisted construction logistics, agentic HVAC control, and predictive maintenance), we demonstrate how discounting can change preferred rankings relative to time-invariant totals and supports ranking sensitivity analysis, discounted payback screening, and break-even discount-rate analysis. We also provide decision guidelines that support go/no-go screening, timing decisions, and minimum "bang-for-your-buck" thresholds. Ultimately, this work contributes a lightweight framework for deciding whether and when to deploy AI-enabled interventions for decarbonization under explicit time preference.

INTRODUCTION

AI can reduce built-environment emissions but also creates emissions from development, training, inference, and supporting compute. The paper addresses inconsistent AI-side accounting and the timing mismatch between AI costs and decarbonization benefits with a time-aware framework.

  • AI deployment creates a carbon-accounting tension because physical-system savings may be accompanied by emissions from AI development, training, inference, and compute infrastructure.
  • The paper targets two gaps: inconsistent omission of AI-side emissions and inadequate treatment of different cost and benefit timelines.These gaps complicate comparison and can make options difficult to rank transparently.
  • The framework represents avoided emissions and AI-induced emissions as time-indexed streams over a finite horizon for comparison under explicit time preference.It distinguishes avoided emissions S(t) from AI-induced emissions C(t).
  • Across stylized cases, discounting can change preferred rankings relative to time-invariant totals and identify break-even discount rates.The cases intentionally span different temporal profiles.
  • The paper provides discounted carbon metrics and heuristics for go/no-go screening, deployment timing, and minimum bang-for-your-buck thresholds.The stated metrics include discounted return on carbon, NEPV, discounted carbon payback, and break-even discount rates.

METHODS

The assessment process specifies how avoided and AI-induced emissions are represented as time-indexed streams and demonstrates the framework with representative scenarios.

  • The methods section defines the assessment process around time-indexed avoided and AI-induced emissions streams and representative demonstration scenarios.

Time-indexed emissions streams and systems boundary

The framework models annual emissions impacts over a finite horizon, separates avoided emissions from AI-induced emissions, and applies a consistent environmental discount rate within an ISO-aligned boundary.

  • The framework uses discrete annual bins from t = 0 through H, with deployment activities at t = 0 and subsequent years at t = 1, . . . , H.
  • For each intervention, S_j(t) captures avoided emissions and C_j(t) captures AI-induced emissions from training and inference, both measured in tCO2e.
  • The system boundary includes processes expected to change because of AI deployment and excludes components identical with and without AI, consistent with ISO 14040/14044.
  • A single environmental discount rate r is applied consistently across calculations, with r = 3% used as the illustrative reference case.

Assumptions

The framework assumes additive discrete-time impacts over a finite horizon, uniform discounting, exogenous emissions factors, and an attributional boundary relative to a no-AI baseline.

  • Impacts are assumed to form additive discrete-time streams over a finite horizon, with annual bins used here but finer temporal resolution permitted.
  • Portfolio impacts are treated as additive across interventions, allowing streams to be summed for portfolios.
  • The framework assumes a constant discount rate applied uniformly and treats emissions factors as exogenous inputs.
  • All quantities are measured relative to a no-AI baseline and include only processes expected to change because of AI deployment.

Time-aware decision metrics

The framework converts avoided and AI-induced emissions into discounted present-value metrics that support efficiency, net-benefit, and payback decisions over a finite horizon.

  • Present-value accounting: Net present value is calculated separately for emissions avoided through AI use and emissions induced by development, training, and inference.
  • Return on carbon: Discounted return on carbon measures present-value tons of CO2e avoided per present-value ton induced by AI over horizon H.The undiscounted ratio is R_simple(j) = R(j; r = 0; H).
  • Net benefit: Discounted net benefit supports evaluation of whether time-weighted avoided emissions exceed time-weighted AI-induced emissions.
  • Payback: Discounted carbon payback is the earliest time at which cumulative discounted net benefit becomes nonnegative.

Representative scenarios with time-sensitive profiles

Four stylized interventions represent distinct build-phase and use-phase timing profiles over a 10-year horizon, with one-time training costs and recurring inference costs.

  • Scenario set: The four cases span low-carbon concrete design, construction logistics, agentic HVAC control, and predictive maintenance across build and use phases.The evaluation horizon is H = 10 years and represents a multi-project program or pipeline rather than one construction project.
  • B1: Concrete pipeline: B1 delays savings until year 3, then provides 15 tCO2e annually, while its AI costs are 6 tCO2e at t = 0 and 0.2 annually thereafter.
  • B2: Logistics pipeline: B2 provides 12 tCO2e annually during years 1–4 and 6 tCO2e during years 5–10, with 2 tCO2e upfront and 0.6 annually in AI costs.
  • U1: HVAC control: U1 delivers 7.5 tCO2e annually from year 1, against 5 tCO2e upfront and 0.5 annually in AI costs.
  • U2: Predictive maintenance: U2 delays 12.5 tCO2e annual savings until year 5, with 5 tCO2e upfront and 0.3 annually in AI costs.B1 and U2 are back-loaded, whereas B2 has tapering benefits and U1 has early, steady savings.
  • Temporal representation: Figure 1 plots annual avoided-emissions and AI-induced-emissions streams, treating plotted negative C(t) values as positive costs for calculations.For B1 and B2, savings recur for new-project cohorts under an approximately constant firm pipeline without accumulating across cohorts.

RESULTS

Across four interventions, time-aware metrics reveal that temporally mismatched avoided and AI-induced emissions can alter rankings relative to static totals. Portfolio preferences and reversal points depend on the governance objective and discount rate.

  • Delayed benefits, tapering deployments, and persistent inference make all four intervention cases time-sensitive, with construction cases especially dependent on multi-year modeling.Treating construction savings as a single t = 0 event can hide temporal effects that influence time-weighted rankings.
  • A static ratio favors high-total, back-loaded portfolio pairings, whereas discounted net impact favors pairings with earlier impacts and shorter payback.The comparison uses the portfolio outcomes in Table 2 with discounted metrics at r = 3%.
  • r⋆R = 0.0821 marks the build-phase efficiency reversal: below it B1 dominates through larger delayed savings, while above it earlier-impact B2 is more efficient.The corresponding NEPV reversal occurs at r⋆NEPV = 0.1655, where B1 remains preferred across a wide range of values.
  • Even modest discounting may favor earlier-benefit HVAC control U1 over back-loaded predictive maintenance U2 under NEPV.This use-phase comparison shows that preferred interventions depend on the selected time-weighted decision metric.

Worked example for facility management

The facility-management example applies the framework to a campus lab-building proxy with recurring energy use and cloud-inference emissions. With an upfront AI cost of 100 tCO2e, discounted payback is reached within the first annual interval.

  • For the campus lab-building proxy, the example combines annual electricity use of 15 GWh, thermal use of 100,000 MMBtu, and PJM-average grid emissions of 0.40 kgCO2e/kWh.The setup uses location-based annual emissions factors and a thermal proxy for facility emissions.
  • Continuous cloud inference is represented as a recurring AI-side emissions rate based on 6,000 kWh/yr of compute use.
  • 100 tCO2e of upfront AI-induced carbon cost yields tpayback = 1 year at annual resolution, likely satisfying the go/no-go gate with substantial margin.Sub-year payback is reported as 1 year; finer intervals such as monthly bins can be used when needed.

DISCUSSION

The discussion translates time-aware metrics into governance heuristics for deployment, timing, and efficiency decisions. It emphasizes that construction impacts can unfold across project pipelines, while the framework assumes constant discount rates and deterministic emissions factors.

  • DISCUSSION: Programmatic construction deployment distributes embodied and transport savings across years, so discount rates can change alternative rankings and discounted payback can screen deployment.AI costs may be front-loaded or persistent, while project savings materialize over multiple years.
  • DISCUSSION: The framework operationalizes three governance heuristics covering go/no-go screening, deployment timing, and minimum acceptable carbon efficiency.
  • DISCUSSION: Deploy when NEPV(j; r; H) > 0 and tpayback ≤ Tmax, where Tmax is a governance-specified maximum acceptable payback.
  • DISCUSSION: Timing decisions evaluate shifted avoided-emissions and AI-cost streams for candidate deployment delays d ∈ {0, 1, . . . , H}.
  • DISCUSSION: Adoption also requires R(j; r; H) ≥ Rmin, making minimum carbon efficiency explicit under the decision-maker’s chosen time preference.
  • DISCUSSION: The simplified framework assumes constant r and deterministic emissions factors; fuller analysis would use hour-resolved marginal factors and uncertain inputs.

CONCLUSION

The framework supports deciding whether and when to deploy AI for decarbonization by accounting for avoided and AI-induced emissions over time. Across four representative interventions, it demonstrates transparent deployment heuristics while cautioning that stylized scenarios are not rigorous benchmark evaluations.

  • CONCLUSION: Time-indexed avoided and AI-induced emissions support deciding whether and when AI should be deployed for decarbonization.The framework represents physical-system benefits and AI-side costs as streams over time, capturing commonly front-loaded AI costs and later-arriving benefits.
  • CONCLUSION: Discount-rate choices can reverse intervention rankings, making temporal preference material to comparative deployment decisions.Across four representative interventions, the framework demonstrates ranking sensitivity and supports evaluating alternative timing choices.
  • CONCLUSION: NEPV and payback thresholds enable transparent go/no-go screening, while delayed-deployment comparisons support explicit timing decisions.A minimum acceptable return-on-carbon threshold provides an additional “bang-for-your-buck” screening rule.
  • CONCLUSION: The framework applies across AI modalities when S(t), C(t), and r can be estimated and monitored.Its deployment guidance is designed to produce transparent and checkable decisions from a minimal set of time-aware inputs.
  • CONCLUSION: Stylized scenarios illustrate framework behavior but should not be treated as rigorous benchmark evaluations.Results depend on attributional assumptions, time-varying emissions factors, and realized savings; future work targets observed-data validation and greater uncertainty and emissions-factor resolution.
Loading 2609.18029v1…