2026-04-16·12 min read·

EU AI Act GPAI Art.53(1)(d) Energy Consumption Reporting: What Systemic Risk Providers Must Measure and Report — Developer Guide (2026)

EU AI Act Article 53(1)(d) requires every provider of a General-Purpose AI model with systemic risk to report energy consumption to the European Commission and the AI Office. The obligation applies to both training-phase energy and inference-phase energy per unit of output — and the methodology that determines exactly what you must measure is being finalised right now, through a delegated act consultation process closing 15 May 2026.

If you train, fine-tune at scale, or operate a GPAI model that crosses the Art.51 10^25 FLOPs systemic risk threshold, Art.53(1)(d) is legally binding as of 2 August 2025. The measurement methodology will be codified in a delegated act under Art.97(2) and reflected in the GPAI Code of Practice energy efficiency chapter. What you measure during the consultation window shapes the evidence base for that delegated act.

For SaaS developers and infrastructure providers who integrate GPAI APIs rather than train models themselves, Art.53(1)(d) determines what energy transparency obligations your GPAI providers must maintain — and what Art.55 downstream disclosure requirements you should expect to receive from them.

This guide covers everything a developer needs to understand Art.53(1)(d) before the consultation deadline: scope, measurement methodology, reporting process, the delegated act pipeline, Python tooling, and a 25-item implementation checklist.

Why Art.53(1)(d) Matters Now: The Consultation Deadline

The EU AI Office opened a public consultation on the delegated act for GPAI energy consumption measurement methodology under Art.97(2) of the EU AI Act. The consultation window closes 15 May 2026.

This matters for three reasons:

1. The methodology determines what you must implement. Art.53(1)(d) says GPAI systemic risk providers must report energy consumption, but does not specify the measurement standard, granularity, reporting format, or update frequency. All of that will be established by the delegated act. Once finalised, providers have limited time to align their measurement infrastructure.

2. Your GPAI CoP commitments are upstream inputs. The GPAI Code of Practice (Art.56) includes an energy efficiency chapter (commonly referred to as Chapter 4 in working drafts). Commitments made in the CoP — which are due before the formal enforcement cycle — must be consistent with the delegated act measurement methodology. Getting the methodology right in the consultation phase protects you from having to retrofit CoP commitments.

3. Audit and enforcement trails start now. Art.53(1)(d) has been applicable since 2 August 2025. Market surveillance authorities and the AI Office can request energy consumption records for that period. If you wait until the delegated act is finalised to start measuring, you will have a gap in your compliance documentation for the 2025–2026 period.

Art.53(1)(d) in Context: The GPAI Obligation Cascade

Art.53 is the enhanced obligation tier of the GPAI chapter, sitting above the Art.52 baseline that applies to all GPAI providers. The obligation cascade is:

Article	Obligation tier	Trigger
Art.51(1)(a)	GPAI classification — standard tier	Any GPAI provider (by definition)
Art.51(1)(b)	Systemic risk classification	Training compute ≥ 10^25 FLOPs OR AI Office designation
Art.52(1)–(4)	Baseline obligations	All GPAI providers
Art.53(1)(a)–(d)	Enhanced obligations	Systemic risk tier only
Art.56	Code of Practice	Systemic risk tier — compliance pathway

The four Art.53 enhanced obligations are:

Sub-article	Obligation
Art.53(1)(a)	Adversarial testing (red-teaming) before market placement and throughout lifecycle
Art.53(1)(b)	Serious incident reporting to the Commission without undue delay
Art.53(1)(c)	Cybersecurity measures for model weights and inference infrastructure
Art.53(1)(d)	Energy consumption reporting — training and inference

Art.53(1)(d) is the only Art.53 obligation that has a direct environmental dimension. It connects EU AI Act compliance to the European Commission's AI and Energy policy agenda, the European Green Deal digital transition pillar, and — for frontier model providers — to scope 2 and scope 3 emissions reporting under the CSRD (Corporate Sustainability Reporting Directive).

Exact Text of Art.53(1)(d)

Article 53(1)(d) provides:

"Providers of GPAI models with systemic risk shall: … (d) report to the Commission and the AI Office, upon request, the results of the evaluations referred to in point (a), and information on energy consumption of the GPAI model with systemic risk, including as regards the training of the GPAI model and its inference."

Three elements define the scope:

"Upon request" — Unlike Art.53(1)(b) serious incident reporting, which is triggered automatically, Art.53(1)(d) energy reporting is formally triggered by a Commission or AI Office request. However, the GPAI Code of Practice and the delegated act methodology are expected to establish proactive reporting obligations and minimum documentation requirements that must be maintained at all times to respond promptly to such requests.

"Energy consumption of the GPAI model with systemic risk" — This is not general corporate energy use. It is specifically the energy attributable to the model — including training compute and inference compute — not the data centre, cooling, or overhead not directly attributable to model operations.

"Including as regards the training … and its inference" — Two separate measurement components are required: (1) training-phase energy (the total energy consumed during the initial training run, including any publicly disclosed fine-tuning at systemic risk compute scale), and (2) inference-phase energy expressed per query, per token, or per unit of meaningful output as the delegated act methodology specifies.

What Energy Consumption Means Under Art.53(1)(d)

Current EU AI Office and GPAI CoP working documents distinguish three energy measurement components:

1. Training Energy

Training energy is the total electrical energy consumed by all compute infrastructure (GPUs, TPUs, networking, storage I/O) during the primary training run of the model, measured from the start of the training job to model convergence. The standard unit is MWh (megawatt-hours) or GWh (gigawatt-hours) for frontier-scale models.

Key scoping questions under the current consultation:

Does training energy include pre-training data pipeline processing (tokenisation, deduplication)?
Are evaluation runs during training included or only the forward+backward passes?
Is fine-tuning at sub-systemic-risk compute included if the base model crosses the threshold?
How are distributed training checkpoints and restarts accounted for?

The working methodology in GPAI CoP Chapter 4 drafts refers to IEA measurement standards and the Green Software Foundation's Software Carbon Intensity (SCI) specification as reference frameworks, though neither is binding until the delegated act.

2. Inference Energy

Inference energy is measured differently from training: rather than a single total figure, inference energy must be expressed as a per-query or per-token rate to be comparable across models with different deployment scales.

Proposed units in consultation documents:

Wh per 1,000 tokens (input + output combined)
Wh per meaningful output unit (per document summary, per code completion, per image) where token-level measurement is not granular enough

The distinction matters for SaaS providers: if you integrate a GPAI API, your energy footprint per API call can be estimated from the provider's published per-token inference energy figure × your token consumption.

3. Disaggregation Requirements

Current consultation proposals suggest that energy reporting should be disaggregated by:

Training phase vs inference phase
Hardware generation (older training cluster vs newer inference cluster can differ by 3–5× in energy efficiency)
Geographic deployment (data centres with different energy mixes affect carbon intensity, even if the watt-hour figure is identical)

The disaggregation requirement is significant for providers with multi-region deployments: you cannot report a single blended average figure if the delegated act requires geographic disaggregation.

The Delegated Act Under Art.97(2)

Article 97(2) empowers the Commission to adopt delegated acts specifying the measurement methodology for energy consumption reporting under Art.53(1)(d):

"The Commission is empowered to adopt delegated acts … in order to specify the technical conditions for [energy consumption] reporting … including measurement methods …"

The delegated act pipeline:

Public consultation (open now, closing 15 May 2026)
Commission drafting following consultation (estimated Q3 2026)
European Parliament and Council scrutiny period (3 months)
Entry into force and application date (estimated Q1 2027)

Between now and the delegated act entry into force, providers must maintain energy records using best available methodology. The GPAI CoP energy efficiency chapter (Art.56) provides an interim compliance pathway: CoP commitments that align with the consultation-phase methodology serve as a defensible good-faith record even if the final delegated act methodology differs in detail.

GPAI Code of Practice: Energy Efficiency Obligations

The GPAI Code of Practice (Art.56) includes energy efficiency as a dedicated chapter. While the CoP is the primary compliance pathway for systemic risk obligations, its energy provisions are specifically designed to be consistent with Art.53(1)(d).

Key CoP energy commitments (based on final CoP drafts circulated in Q1 2026):

Measurement infrastructure commitment — Providers must demonstrate they have instrumented training and inference workloads to capture energy consumption at the hardware level (GPU/TPU power draw), not estimated from billing records alone.

Annual energy report commitment — Providers must publish an annual energy consumption summary covering training energy for any new model released in the reporting period and aggregate inference energy for production deployments.

Efficiency improvement commitment — Providers must demonstrate year-over-year improvement in inference energy per token through hardware, batching, quantization, or other optimization measures. The CoP does not specify a numeric target; it requires a documented improvement trajectory.

Third-party verification commitment — For systemic risk tier providers, energy figures must be verifiable by the AI Office on request. This means measurement logs, hardware telemetry records, and energy billing data must be retained and accessible.

The CoP commitments are not a substitute for the Art.53(1)(d) upon-request reporting obligation, but they establish the documentation infrastructure needed to respond to an AI Office request promptly.

Practical Implementation: What You Need to Measure

For GPAI Model Providers (Training Energy)

The minimum measurement infrastructure for training energy:

import dataclasses
from datetime import datetime
from typing import Optional

@dataclasses.dataclass
class TrainingRunEnergyRecord:
    """
    Art.53(1)(d) training energy record for a GPAI model.
    One record per distinct training run (initial training or 
    continued pre-training at systemic risk compute scale).
    """
    model_version: str
    training_start_utc: datetime
    training_end_utc: datetime
    
    # Hardware inventory
    accelerator_type: str          # e.g., "NVIDIA H100 SXM5"
    accelerator_count: int
    data_centre_location: str      # ISO country code
    data_centre_pue: float         # Power Usage Effectiveness
    
    # Measured energy (not estimated)
    gpu_energy_mwh: float          # Direct GPU/TPU power draw
    total_facility_energy_mwh: float  # GPU energy * PUE
    
    # Compute parameters (for cross-validation with FLOPs estimate)
    training_tokens: int
    training_flops: float          # Scientific notation, e.g., 3.14e25
    
    # Measurement method
    measurement_method: str        # "hardware_telemetry" | "pdu_metering" | "billing_allocation"
    measurement_uncertainty_pct: float  # ±percentage
    
    # Documentation
    prepared_by: str
    last_updated: datetime
    
    def is_systemic_risk_training(self) -> bool:
        """Art.51(1)(b): training compute ≥ 10^25 FLOPs triggers Art.53"""
        return self.training_flops >= 1e25
    
    def energy_per_billion_flops(self) -> float:
        """Normalised efficiency metric: MWh per 10^9 FLOPs"""
        if self.training_flops == 0:
            return 0.0
        return self.total_facility_energy_mwh / (self.training_flops / 1e9)


@dataclasses.dataclass  
class InferenceEnergyRecord:
    """
    Art.53(1)(d) inference energy record — rolling measurement period.
    Expressed per-token for comparability.
    """
    model_version: str
    measurement_period_start: datetime
    measurement_period_end: datetime
    
    # Production deployment snapshot
    serving_cluster_location: str
    accelerator_type: str
    batch_size_p50: float          # Median batch size during period
    quantization_bits: Optional[int]  # e.g., 8 (INT8), 4 (INT4), None (FP16/BF16)
    
    # Measured inference energy
    total_tokens_processed: int    # Input + output tokens
    total_inference_energy_wh: float
    
    # Derived metrics
    @property
    def wh_per_1k_tokens(self) -> float:
        """Primary reporting metric: Wh per 1,000 tokens"""
        if self.total_tokens_processed == 0:
            return 0.0
        return (self.total_inference_energy_wh / self.total_tokens_processed) * 1000
    
    @property
    def co2_kg_per_1k_tokens(self) -> float:
        """Carbon intensity — requires grid emission factor"""
        # This field requires the data centre grid carbon intensity (gCO2/kWh)
        # Placeholder: implementors must integrate with local grid factor
        return 0.0  # Implement with self._grid_co2_intensity_g_per_kwh

For GPAI Model Providers (AI Office Reporting Package)

When the AI Office requests an energy consumption report under Art.53(1)(d), the response package should include:

import json
from typing import List

class Art53dReportingPackage:
    """
    Art.53(1)(d) energy consumption report package for AI Office submission.
    Prepare this documentation proactively; the 'upon request' trigger
    means you must be able to deliver it without delay.
    """
    
    def __init__(
        self,
        model_name: str,
        model_version: str,
        provider_entity: str,
        reporting_date: datetime,
    ):
        self.model_name = model_name
        self.model_version = model_version
        self.provider_entity = provider_entity
        self.reporting_date = reporting_date
        self.training_records: List[TrainingRunEnergyRecord] = []
        self.inference_records: List[InferenceEnergyRecord] = []
    
    def add_training_record(self, record: TrainingRunEnergyRecord) -> None:
        self.training_records.append(record)
    
    def add_inference_record(self, record: InferenceEnergyRecord) -> None:
        self.inference_records.append(record)
    
    def total_training_energy_mwh(self) -> float:
        return sum(r.total_facility_energy_mwh for r in self.training_records)
    
    def weighted_inference_wh_per_1k_tokens(self) -> float:
        """Token-weighted average across all inference measurement periods"""
        total_tokens = sum(r.total_tokens_processed for r in self.inference_records)
        if total_tokens == 0:
            return 0.0
        total_energy_wh = sum(r.total_inference_energy_wh for r in self.inference_records)
        return (total_energy_wh / total_tokens) * 1000
    
    def validate_for_submission(self) -> list[str]:
        """
        Check completeness before Art.53(1)(d) submission.
        Returns list of missing items.
        """
        missing = []
        
        if not self.training_records:
            missing.append("MISSING: At least one TrainingRunEnergyRecord required")
        
        for i, tr in enumerate(self.training_records):
            if tr.measurement_method == "billing_allocation":
                missing.append(
                    f"WARNING: Training record {i} uses billing_allocation — "
                    "hardware_telemetry or pdu_metering preferred for AI Office submission"
                )
            if tr.measurement_uncertainty_pct > 15:
                missing.append(
                    f"WARNING: Training record {i} has >{tr.measurement_uncertainty_pct}% "
                    "uncertainty — document methodology justification"
                )
        
        if not self.inference_records:
            missing.append("MISSING: At least one InferenceEnergyRecord required")
        
        if self.weighted_inference_wh_per_1k_tokens() == 0:
            missing.append("MISSING: Valid inference energy measurements needed")
        
        return missing
    
    def to_json_summary(self) -> str:
        return json.dumps({
            "model": f"{self.model_name} {self.model_version}",
            "provider": self.provider_entity,
            "report_date": self.reporting_date.isoformat(),
            "training_total_mwh": round(self.total_training_energy_mwh(), 2),
            "inference_wh_per_1k_tokens": round(
                self.weighted_inference_wh_per_1k_tokens(), 4
            ),
            "training_records_count": len(self.training_records),
            "inference_periods_count": len(self.inference_records),
            "validation_issues": self.validate_for_submission(),
        }, indent=2)

Art.53(1)(d) × Art.51: The Systemic Risk Threshold

Art.53(1)(d) only applies to GPAI models that have crossed the Art.51 systemic risk threshold. Art.51 specifies two classification paths:

Art.51(1)(b)(i) — Compute threshold: Training compute ≥ 10^25 FLOPs. Current models in this tier include GPT-4 class (estimated 2.15×10^24 FLOPs — slightly below, depending on counting methodology), Gemini Ultra class, Claude Opus class, and Llama 3.1 405B.

Art.51(1)(b)(ii) — AI Office designation: The AI Office may designate a GPAI model as having systemic risk based on capabilities assessment, even below the compute threshold. This designation triggers Art.53 obligations including Art.53(1)(d).

Implication for energy measurement: If your model is below the compute threshold today but you expect to cross it in the next training run, you should establish your energy measurement infrastructure now. Retrofitting measurement after crossing the threshold will require reconstructing training energy from billing records — an approach the AI Office consultation process suggests is insufficient for compliance.

from enum import Enum

class SystemicRiskBasis(Enum):
    COMPUTE_THRESHOLD = "Art.51(1)(b)(i) — 10^25 FLOPs threshold"
    AI_OFFICE_DESIGNATION = "Art.51(1)(b)(ii) — AI Office designation"
    NOT_SYSTEMIC_RISK = "Below threshold, no Art.53 obligations"

def assess_art51_status(
    training_flops: float,
    ai_office_designated: bool = False
) -> SystemicRiskBasis:
    """
    Determine Art.53(1)(d) applicability from Art.51 status.
    """
    if ai_office_designated:
        return SystemicRiskBasis.AI_OFFICE_DESIGNATION
    if training_flops >= 1e25:
        return SystemicRiskBasis.COMPUTE_THRESHOLD
    return SystemicRiskBasis.NOT_SYSTEMIC_RISK

def art53_1d_applies(
    training_flops: float,
    ai_office_designated: bool = False
) -> bool:
    """Art.53(1)(d) energy reporting required?"""
    status = assess_art51_status(training_flops, ai_office_designated)
    return status != SystemicRiskBasis.NOT_SYSTEMIC_RISK

Downstream Implications: Art.55 and SaaS Providers

If you are a SaaS developer who integrates a GPAI API rather than training models yourself, Art.53(1)(d) affects you through Art.55 downstream provider obligations.

Art.55 requires GPAI providers to make available to downstream providers:

"… technical documentation, information and tools … sufficient to comply with the obligations applicable to providers and deployers of high-risk AI systems …"

For energy specifically, the GPAI CoP energy efficiency chapter extends this to require that systemic risk GPAI providers publish:

Per-token inference energy figure (updated at least annually)
Hardware generation and deployment region information sufficient for carbon intensity calculation
Any significant changes in energy efficiency following model updates

Practical implication for SaaS developers: When evaluating GPAI API providers, check whether they publish their inference energy per token. A provider with Art.53(1)(d) obligations but no published energy figure either: (a) has not yet established their energy measurement infrastructure, or (b) is not treating Art.53(1)(d) as applicable to their model

Both situations are relevant to your own compliance posture if you are building an Annex III high-risk AI system on top of their API.

For EU-sovereign infrastructure choices, the jurisdiction of your inference infrastructure affects the carbon intensity calculation — EU data centres subject to the EU ETS (Emissions Trading System) and renewable energy targets have generally lower grid emission factors than non-EU data centres. This is one dimension where hosting infrastructure choice has a direct compliance documentation impact under the energy efficiency chapter.

Enforcement Timeline and Penalties

Art.53(1)(d) has been applicable since 2 August 2025 — the date on which Chapter V of the EU AI Act entered into force for GPAI providers.

Enforcement is led by the AI Office (European Commission) for GPAI models with systemic risk, as established by Art.88. Market surveillance authorities in each Member State have concurrent jurisdiction for GPAI models deployed in their territory.

The penalty for non-compliance with Art.53 obligations (including Art.53(1)(d)) is established by Art.99(3):

"Non-compliance with obligations applicable to providers of GPAI models with systemic risk … shall be subject to administrative fines of up to EUR 15 000 000 or, if the offender is an undertaking, up to 3 % of its total worldwide annual turnover …"

Note the asymmetry: Art.53 penalties (3% / €15M) are lower than Art.5 prohibited practices (7% / €35M) but higher than Art.52 baseline GPAI penalties (1% / €7.5M for some infringements). The energy reporting obligation sits at the 3% / €15M tier.

Given that the AI Office can request energy consumption records covering the period since 2 August 2025, providers without measurement infrastructure for that period are exposed to the Art.99(3) penalty tier for the historical gap.

Art.53(1)(d) × Art.56 Code of Practice: Compliance Strategy

For most GPAI systemic risk providers, the GPAI Code of Practice (Art.56) is the primary compliance pathway. The relationship between Art.53(1)(d) and the CoP for energy reporting is:

Compliance layer	Function
Art.53(1)(d)	Legal obligation to report energy consumption upon request
GPAI CoP (Art.56)	Presumption of conformity with Art.53(1)(d) if CoP commitments fulfilled
Delegated act (Art.97(2))	Measurement methodology specification (finalised after 15 May 2026 consultation)

Providers who sign the GPAI CoP and fulfil their energy efficiency commitments benefit from a presumption of conformity with Art.53(1)(d) under Art.56(5). This means the AI Office would need to rebut that presumption before asserting a violation — a higher evidential bar than if no CoP commitment exists.

The practical compliance strategy before the delegated act is finalised:

Implement measurement infrastructure now (hardware telemetry, not billing estimation)
Document training energy for all systemic risk training runs since August 2025
Establish per-token inference energy baseline for current production deployment
Sign the GPAI CoP and commit to the energy efficiency chapter
Respond to the consultation to influence the measurement methodology that will govern your obligations
Prepare an Art.53(1)(d) reporting package so you can respond to an AI Office request within 30 days

25-Item Implementation Checklist

Scope and Classification (Items 1–5)

Art.51 status verified — training compute calculated (FLOPs) or AI Office designation letter held on file
Art.53(1)(d) applicability confirmed — written legal assessment of whether energy reporting obligation applies
Scope of 'model' defined — documented decision on what training runs and inference deployments are within scope
Historical gap assessed — energy records covering period since 2 August 2025 located or reconstruction plan documented
Consultation submission prepared — written response to AI Office delegated act consultation submitted before 15 May 2026

Measurement Infrastructure (Items 6–12)

Hardware telemetry enabled — GPU/TPU power draw recorded at hardware level (not estimated from billing)
PUE documented — Power Usage Effectiveness for each data centre in training cluster on file
Training run records complete — TrainingRunEnergyRecord populated for all Art.51-threshold training runs
Inference energy metered — per-token inference energy measured in production, not estimated from benchmark
Geographic disaggregation — energy records disaggregated by data centre location
Measurement uncertainty quantified — percentage uncertainty documented for each measurement method
Billing cross-validation — energy figures cross-checked against data centre billing records for consistency

Reporting Readiness (Items 13–18)

Art53dReportingPackage assembled — complete reporting package for current model version prepared
Validation issues resolved — validate_for_submission() returns no blocking issues
30-day response capability confirmed — internal process exists to deliver report to AI Office within 30 days of request
Training energy summary figure confirmed — total MWh for primary training run documented
Inference energy per-token figure confirmed — Wh/1k tokens figure documented for current production deployment
Third-party verification feasible — telemetry and billing records sufficient for independent verification held accessible

GPAI Code of Practice (Items 19–22)

GPAI CoP signed — provider registered with AI Office and signed CoP or committed to sign before deadline
CoP energy chapter commitments reviewed — specific energy commitments cross-checked against current measurement capability
Annual energy report commitment acknowledged — process for annual publication of energy consumption summary established
Efficiency improvement trajectory documented — year-over-year inference energy efficiency improvement tracked

Downstream and Systemic (Items 23–25)

Art.55 energy disclosure to downstream providers — per-token inference energy figure published in API documentation
Carbon intensity calculation documented — grid emission factor for each deployment region recorded for CSRD cross-reference
Delegated act monitoring assigned — responsible person assigned to monitor delegated act finalisation and update measurement methodology when published (estimated Q1 2027)

EU-Sovereign Infrastructure and Energy Reporting

For systemic risk GPAI providers who select EU-sovereign inference infrastructure, the energy compliance dimension has a concrete advantage: EU data centres are subject to the EU ETS, the Energy Efficiency Directive, and the Renewable Energy Directive — meaning their energy mix documentation, grid carbon intensity data, and PUE certification are readily available through EU regulatory channels.

Non-EU data centres may have equivalent or superior energy efficiency, but the provenance documentation needed for the geographic disaggregation requirement in the delegated act will require additional verification steps. When the delegated act methodology includes geographic emission factor disaggregation (as current consultation documents propose), EU-jurisdiction deployment provides a cleaner paper trail.

For downstream SaaS developers building on GPAI APIs: if your provider's inference infrastructure is hosted in an EU-sovereign cloud, you can reference their published EU ETS documentation when calculating your own scope 3 AI emissions for CSRD reporting. If their infrastructure is outside EU jurisdiction, you are dependent on their voluntary disclosure for that calculation.

Summary

Art.53(1)(d) energy consumption reporting is a live obligation for GPAI systemic risk providers. The measurement methodology is being finalised through a consultation process closing 15 May 2026 — which means the decisions made in the consultation window directly determine what you will be required to implement.

The key actions before 15 May 2026:

Verify Art.51 systemic risk status and confirm Art.53(1)(d) applies
Implement hardware-level energy telemetry for training and inference
Reconstruct training energy records since August 2025 if not already measured
Submit a response to the AI Office delegated act consultation
Ensure GPAI CoP energy efficiency commitments are consistent with consultation-phase methodology

After 15 May 2026, monitor the delegated act drafting process and update your measurement infrastructure to match the final methodology when it enters into force (estimated Q1 2027).