Open-source · IEEE JSAC 2026

The energy cost of building intelligence.

Training and inference are only part of the bill. eCAL is the first metric that adds up the energy of data collection, preprocessing, training, evaluation, and inference — end to end, in Joules per bit — so you can see the true cost of adding AI to a communication system.

1 Set up the lifecycle 2 Plot the case study 3 Draw the real data
Every AI model pays a one-time development cost before its first inference.
Computed directly with the open-source eCAL calculator (ecal-energy) — the same case-study configuration and the same sampled inference counts (γ = 10²…10⁸) as the paper's Fig. 13.
2.73×
Lower eCAL going from 100 to 10,000 inferences on the case-study MLP
141×
Higher energy per bit for a Transformer vs. an MLP at 100 inferences
7
OSI layers modeled end-to-end, from application down to physical
4
Model architectures supported out of the box — MLP, CNN, KAN, Transformer
The problem

Existing energy metrics only see part of the picture.

Telecom metrics stop at the network. Deep-learning metrics stop at the model. No one connects the two — so nobody can say what it actually costs, in energy, to make a communication system intelligent.

Blind to computation

Energy-per-Bit, PUE, CUE, and WUE measure the network and data-center layer with precision — but they have no notion of what the AI model on top is actually computing.

Blind to the network

APC, APEC, TTCAPC, and TTCAPEC weigh model accuracy against training or inference cost — but most don't even cover inference, and none account for the energy spent collecting the data in the first place.

No end-to-end view

Without a metric spanning the full OSI-to-MLOps pipeline, nobody can compare the true energy cost of "adding intelligence" to a network, or predict how that cost falls as a model is used more.

What we build

A single metric, grounded in both networking and machine learning.

eCAL formalizes every stage of the AI lifecycle as a standard OSI/MLOps data-manipulation component, then reduces the whole pipeline to one closed-form number in Joules per bit.

01

Lifecycle-wide methodology

A formal breakdown of data collection, preprocessing, training, evaluation, and inference as standardized OSI and MLOps "data manipulation components" — the shared vocabulary the rest of eCAL builds on.

02

OSI-grounded data collection

Closed-form energy formulas across all 7 OSI layers — retransmission rates, protocol overheads, wired and wireless links — validated against real Wi-Fi 6 and 5G measurement studies.

03

Closed-form FLOPs, any architecture

Exact FLOP formulas for MLP, CNN, KAN, and Transformer models, mapped to real hardware power and FLOPS profiles — from an Apple M2 to an NVIDIA H100 — with pluggable support for ResNet, VGG, and Baichuan-style LLMs.

04

One open-source calculator

A single closed-form J/bit expression spanning the whole lifecycle, shipped as a modular, extensible Python package — pip install ecal-energy — with both a CLI and an API.

The metric

One equation for the whole AI lifecycle.

eCAL reduces data collection, preprocessing, training, evaluation, and inference to a single Joules-per-bit number — so the cost of developing a model and the cost of running it can be compared on the same scale.

eCAL= eCALabs f · IS · (NS + γ·NI,P)
eCALabs=(1 + γᵥ)·(ED+γ·Einf,p)

ED is the one-time development energy — data collection, preprocessing, training, and evaluation. Einf,p is the energy of a single inference pass. γ is how many times the deployed model is actually run; γᵥ captures virtualization overhead. f, IS, NS, and NI,P normalize everything to Joules per bit of data processed.

γ smallDevelopment-dominateda single inference (6.57×10⁻⁷ J/b) is far cheaper than development (2.41×10⁻⁶ J/b) alone.
γ largeInference-dominateddevelopment cost is amortized — 2.73× lower eCAL at γ=1000 than at γ=100 for the case-study MLP.
γᵥ > 0Virtualization compounds botha flat (1+γᵥ) overhead multiplier, regardless of regime.

Validated against real Wi-Fi 6 and 5G measurement datasets, where eCAL acts as a tight upper bound on observed energy cost.

Peer-reviewed research

The work behind the metric.

Grounded in a formal lifecycle methodology, validated against real network measurement studies, and open-sourced as a working tool — not just a paper.

IEEE JSAC · 2026★ Flagship paper

The Energy Cost of Artificial Intelligence Lifecycle in Communication Networks

Introduces eCAL, the first metric to span the full AI lifecycle — data collection, preprocessing, training, evaluation, and inference — in a single Joules-per-bit number, validated on a real case-study pipeline and against Wi-Fi 6 / 5G measurement data.

  • Proposes eCAL, the first metric spanning the full AI lifecycle in J/bit
  • Shows a 2.73× energy-efficiency improvement going from 100 to 1000 inferences on a case-study MLP
  • Validated against real Wi-Fi 6 and 5G measurement datasets, acting as a tight upper bound
  • Ships as an open-source, modular, extensible calculator (PyPI ecal-energy)
Vol. 44, pp. 2427–2443 · Published 11 Dec 2025 · Chou, Hribar, Hanžel, Mohorčič, Fortuna — SensorLab, Jožef Stefan Institute
2025 IEEE NFV-SDN

Energy Cost of the AI/ML Workflow in O-RAN

Applies the eCAL lifecycle methodology to O-RAN AI/ML workflows, quantifying the energy footprint of adding machine learning to next-generation radio access networks.

  • Extends the eCAL methodology to O-RAN AI/ML pipelines
  • Analyzes energy cost across the workflow, not just training
  • Contributes toward sustainable 6G network design
pp. 1–6 · Chou, Hribar, Bertalanič, Mohorčič, Lagkas, Sarigiannidis, Fortuna
2024 IEEE CSCN

Towards the Standardization of Energy Efficiency Metrics of the AI Lifecycle in 6G and Beyond

The earliest work in this research line, arguing for standardized, lifecycle-aware energy-efficiency metrics as AI becomes native to 6G and future communication networks.

  • Motivates the need for standardized AI-lifecycle energy metrics
  • Frames energy efficiency as a first-class 6G design requirement
  • Lays the conceptual groundwork later formalized as eCAL
pp. 187–190 · Chou, Hribar, Mohorčič, Fortuna
Try it

Estimate the energy cost of your own model.

eCAL ships as an open-source Python package with a CLI and an API — plug in your architecture, hardware, and dataset size, and get a J/bit estimate in seconds.

Install & CLI
$ pip install ecal-energy
$ ecal estimate --model MLP --layers 3 --epochs 50 --hardware apple_m2
Python API
import ecal

result = ecal.estimate(
    model_type="MLP",
    model_params={"num_layers": 3, "din": 10, "dout": 2},
    num_samples=1000, num_epochs=50,
    hardware="apple_m2",
)
print(result["ecal_j_per_bit"])
Models — MLP · CNN · KAN · Transformer Hardware — Apple M2 (3.6 TFLOPS/22W) · A100 80GB (19.5 TFLOPS/400W) · H100 SXM (67 TFLOPS/700W) · generic CPU (1 TFLOPS/100W) · generic edge (0.01 TFLOPS/15W)
BSD-3-Clause licensed · Available on PyPI as ecal-energy
Why eCAL

Scaling laws gave training a shared unit of cost. eCAL gives it to the whole lifecycle.

See the true cost before you deploy

Estimate the full lifecycle cost of a model — not just training, not just inference — before committing hardware or bandwidth to it.

Compare architectures fairly

MLP, CNN, KAN, or Transformer — eCAL puts every architecture on the same J/bit scale, so efficiency claims can be compared directly instead of by proxy.

Extend it to your own stack

Open-source and modular: plug in custom protocols, architectures, or hardware profiles instead of being limited to what ships out of the box.

Get in touch

Let's make AI's energy cost measurable.

Whether you're building energy-aware AI systems, researching sustainable networking, or want to extend eCAL to a new architecture or protocol — we'd like to hear from you.