For infrastructure, ML and sustainability practitioners who need to measure the energy and carbon of AI training and inference, report it credibly, and cut it without hurting results.
The curriculum is being prepared and reviewed. Enrollment is not open yet.
About this course
AI's energy demand has become a planning constraint for utilities, a reporting obligation for operators, and a question every customer and regulator now asks. Answers are often wrong in both directions: headline numbers that ignore idle power, cooling and embodied emissions, or offsets that hide real consumption. This course teaches practitioners to produce numbers they can defend and reductions that are real.
You start by measuring: GPU power through NVIDIA DCGM and NVML, node and rack power, estimation tools such as CodeCarbon and Zeus, and Kubernetes-level attribution with Kepler. You then convert energy into carbon and water using grid intensity data, PUE and WUE, and the GHG Protocol's location-based and market-based Scope 2 methods, and add the embodied carbon of servers and GPUs. A reporting module covers the Software Carbon Intensity standard (ISO/IEC 21031:2024) and the Green Software Foundation's SCI for AI, where US climate disclosure stands (California's SB 253 requires large companies doing business in the state to report Scope 1 and Scope 2 emissions), and, for US companies that run data centers or sell AI models in Europe, the EU Energy Efficiency Directive's data-center reporting database and the A-to-G rating scheme the Commission adopted in September 2026, and the energy documentation expected of general-purpose AI model providers under the EU AI Act. The second half is about reduction: choosing smaller and distilled models, quantization and efficient serving, GPU power capping and frequency tuning, and carbon-aware and time-shifted scheduling. Facility power and cooling design is covered in depth in inf-340. Regulatory and tool facts are current as of October 2026; this course is educational and not legal advice.
What you will be able to do
Measure the energy of a training job and an inference service using DCGM, NVML-based tools and facility meters
Convert measured energy into location-based and market-based carbon and water estimates using PUE, WUE and grid intensity
Estimate embodied emissions of AI hardware and allocate them across workloads
Calculate an SCI score for an AI service using the SCI and SCI for AI methodology
Identify which reporting obligations, such as California SB 253 emissions disclosure, EU data-center reporting and AI Act energy documentation, apply to a US organization
Reduce energy per useful output through model choice, quantization, batching and GPU power capping
Design carbon-aware scheduling for flexible training and batch inference workloads
Spot misleading AI sustainability claims and write a defensible footprint statement
Syllabus
6 modules · 22 lessons · about 12 hours
1
Module 1: Where AI energy goes
The physical footprint of AI and the vocabulary used to describe it.3 lessons
Training versus inference: why inference now dominates lifetime energy for popular models
From chip to grid: GPU, node, cooling and facility overhead
Energy, carbon and water: the three footprints and how they diverge
2
Module 2: Measuring energy
Getting trustworthy energy numbers from hardware and software.4 lessons
GPU power telemetry with DCGM and nvidia-smi: what the counters measure and miss
Software estimators: CodeCarbon, Zeus and ML.ENERGY-style measurement
Attributing power in shared clusters: Slurm energy accounting and Kepler on Kubernetes
Designing an energy benchmark: energy per token, per image and per training step
3
Module 3: Carbon and water accounting
Turning kilowatt-hours into emissions and water figures that hold up.4 lessons
Grid carbon intensity: average versus marginal, hourly data from Electricity Maps and WattTime
GHG Protocol Scope 2: location-based, market-based, PPAs and 24/7 carbon-free energy
PUE, WUE and the ISO/IEC 30134 metrics
Embodied carbon of GPUs and servers and how to amortize it
4
Module 4: Standards, disclosure and regulation
The frameworks and rules that shape AI footprint reporting.4 lessons
Software Carbon Intensity (ISO/IEC 21031) and the SCI for AI specification
EU data-center reporting under the Energy Efficiency Directive and the A-to-G sustainability rating scheme
AI Act energy documentation for general-purpose AI models and cloud provider carbon reports
Greenwashing risks: offsets, averages and claims that will not survive scrutiny
5
Module 5: Efficient models and serving
Reducing energy per useful output at the model and software level.3 lessons
Right-sizing the model: small, distilled and mixture-of-experts models
Precision and quantization: FP8, FP4 and INT4 and their energy effect
Serving efficiency: batching, KV-cache reuse, speculative decoding and utilization
6
Module 6: Efficient operations
Running GPU fleets and schedules for lower energy and carbon.4 lessons
GPU power capping and clock tuning: finding the energy-optimal operating point
Carbon-aware and time-shifted scheduling for training and batch jobs
Idle power, utilization and consolidation in GPU clusters
Building a Green AI program: targets, dashboards and a footprint statement
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by NVIDIA (DCGM, nvidia-smi), CodeCarbon, Zeus / ML.ENERGY, Green Software Foundation (SCI, SCI for AI), Electricity Maps, WattTime, Kepler, or AWS, Google Cloud and Microsoft Azure carbon tools. Product and company names are trademarks of their respective owners.
A plain-language tour of the stack that runs AI: what a GPU server is, why clusters need special networks and storage, how jobs are scheduled, what an AI data center needs in power and cooling, and how cloud and on-premises options differ. No prior infrastructure background required.
The operator's view of data-center GPUs: what the current NVIDIA and AMD systems are, how they are packaged and connected, how the driver, CUDA/ROCm and container stack fit together, how to monitor and triage failing GPUs, how to share and size GPU memory, and what power, cooling and sourcing decisions follow.
The Linux skills that differ on an AI cluster: choosing and imaging the OS, packaging GPU and RDMA drivers that survive kernel updates, NUMA and PCIe topology tuning, cgroup v2 isolation, shared identity and file systems, scientific software environments with Lmod, Spack and containers, and patching a fleet without breaking jobs.