Training AI on Sunshine: Carbon-Aware Scheduling in Practice
TL;DR
Here is a clear, practical guide to training AI: the fundamentals, the best practices that actually move the needle, common mistakes to avoid, concrete data points, and a short FAQ. Everything is structured so you can apply it to real projects today.
Key takeaways
- For AI, inference at scale usually dominates lifetime energy — invest in quantization, distillation, batching, and caching, not just efficient training.
- Bake carbon into your CI/CD and observability: what you can measure on every deploy is what teams will actually optimize.
- PUE only measures facility overhead; a low PUE running on a dirty grid can still be high-carbon, so pair it with grid carbon intensity and CFE metrics.
- Measure carbon as intensity, not just totals: adopt the SCI rate (grams CO2e per functional unit) so efficiency wins are visible even as usage grows.
- Prefer managed, high-utilization cloud and modern efficient hardware over always-on self-managed servers running at low utilization.
This is a practical, up-to-date guide to Training AI — what it is, why it matters in 2026, and how to apply it in real projects. It is written for developers and founders who want clear answers and proven best practices, not filler.
Whether you're just starting out or leveling up, treat this as a working reference you can return to. Every section is built to be skimmed, applied, and shared.
What is green software engineering?
Green software engineering is the discipline of building and operating applications that emit less carbon, defined by the Green Software Foundation around three activities: energy efficiency, hardware efficiency, and carbon awareness. Energy efficiency means doing the same work with fewer joules; hardware efficiency means using existing devices longer and at higher utilization to amortize their embodied carbon; and carbon awareness means doing more work when and where electricity is cleaner. The guiding insight is that everything a program does ultimately draws electricity from a grid whose carbon intensity varies by hour and region, so software choices have physical emissions consequences. Crucially, the field frames carbon as a first-class engineering property, on par with latency, cost, and reliability, rather than an offsetting or procurement problem handled only by facilities teams.
Sustainable data center design
Sustainable data centers attack energy waste across cooling, power delivery, and hardware. Modern facilities raise cold-aisle temperatures, use free-air and evaporative or warm-water liquid cooling, and contain hot and cold aisles to shrink the fraction of energy spent moving heat rather than computing. Power delivery is tightened with high-efficiency UPS systems and higher-voltage distribution to cut conversion losses, and Google famously used DeepMind reinforcement learning to tune cooling setpoints in real time. Beyond efficiency, operators pursue heat reuse (piping waste heat to district heating networks, as several Nordic sites do), water-stewardship metrics like WUE, and increasingly the reporting of embodied carbon from construction and servers, which is a growing share of total lifecycle emissions as operational efficiency improves.
How Software Carbon Intensity (SCI) works
Software Carbon Intensity, now standardized as ISO/IEC 21031:2024, expresses a software system's carbon as a rate rather than a total: SCI = ((E times I) + M) per functional unit R. Here E is the energy the software consumes, I is the location-based marginal carbon intensity of the electricity powering it, and M is the embodied carbon of the hardware amortized over its useful life. R is a unit of work chosen by the team, such as per API request, per user, or per benchmark run, which makes the score comparable across releases. Because it is a rate, SCI keeps teams honest as they scale: total emissions may rise with growth, but a falling SCI proves the software itself is getting more carbon-efficient. Unlike annual corporate greenhouse-gas inventories, SCI is designed to be recalculated frequently and owned directly by engineering teams.
Getting started with green software
A pragmatic starting path is measure, then reduce, then shift. Begin by turning on your cloud provider's carbon dashboard and picking one meaningful functional unit to compute an SCI-style rate you can track release over release. Next, harvest the large, low-risk efficiency wins: right-size instances, enable autoscaling and scale-to-zero, delete idle resources and stale data, choose efficient instance families and regions, and cache aggressively to avoid repeated compute. Then introduce carbon awareness for genuinely flexible workloads by wiring a grid-signal API or the Carbon Aware SDK into batch schedulers so deferrable jobs prefer low-carbon windows and regions. Finally, make it durable by adding a carbon or energy metric to dashboards and CI so regressions are visible, and by upskilling the team through resources like the GSF Green Software Practitioner course.
Business and regulatory drivers
Green tech is increasingly compelled by regulation and cost, not just goodwill. The EU's Corporate Sustainability Reporting Directive (CSRD) and the associated European Sustainability Reporting Standards are pulling many large companies and their suppliers into mandatory, audited disclosure that includes Scope 3 emissions, which is where most software and cloud carbon lands. The EU Energy Efficiency Directive adds specific reporting obligations for larger data centers, and jurisdictions elsewhere are tightening efficiency rules amid surging AI-driven demand. Commercially, energy efficiency correlates tightly with cloud cost — a right-sized, high-utilization system is usually both cheaper and greener — so FinOps and sustainability programs increasingly reinforce each other rather than compete, giving engineers a rare metric that finance, operations, and ESG teams all want to see fall.
Measuring your cloud carbon footprint
Cloud carbon footprinting attributes the emissions of your rented compute, storage, and network back to your accounts, services, and teams. Each major provider now ships a native tool — AWS Customer Carbon Footprint Tool, Microsoft's Emissions Impact Dashboard, and Google Cloud Carbon Footprint — while the open-source Cloud Carbon Footprint project estimates emissions across AWS, Azure, and GCP from billing and usage data using published coefficients. These figures are typically Scope 3 for the customer (upstream emissions embedded in a purchased service) and Scope 1 and 2 for the provider, which is a frequent source of accounting confusion. A recurring challenge is transparency: providers differ in whether they report location-based versus market-based emissions, how they handle renewable-energy purchases, and how much embodied hardware carbon they include, so cross-vendor comparisons require care.
Training AI: Key Facts and Data
According to recent industry research and the official documentation linked below:
- Training a single large frontier AI model can consume energy on the order of several gigawatt-hours, and inference at scale is now widely regarded as the dominant lifetime energy cost for popular deployed models rather than the one-time training run.
- Google reports a fleet-wide trailing-twelve-month PUE of approximately 1.09 as of recent years, among the lowest at scale, achieved through custom cooling, ML-driven optimization, and warm-water and free-air cooling designs.
- The Green Software Foundation's Software Carbon Intensity (SCI) specification was published as ISO/IEC 21031:2024, giving green software its first formal international standard and moving carbon measurement from aggregate reporting toward a per-unit rate engineers can act on.
Quick-Reference Summary
A map of what this guide covers:
| Topic | What you'll learn |
|---|---|
| What is green software engineering? | Green software engineering is the discipline of building and operating applications that emit less carbon |
| Sustainable data center design | Sustainable data centers attack energy waste across cooling, power delivery, and hardware. |
| How Software Carbon Intensity (SCI) works | Software Carbon Intensity, now standardized as ISO/IEC 21031:2024, expresses a software system's carbon as a rate |
| Getting started with green software | A pragmatic starting path is measure, then reduce, then shift. |
| Business and regulatory drivers | Green tech is increasingly compelled by regulation and cost, not just goodwill. |
| Measuring your cloud carbon footprint | Cloud carbon footprinting attributes the emissions of your rented compute |
How to Get Started with Training AI
A simple path that works:
- Learn the fundamentals of Training AI from primary sources, not just tutorials.
- Build one small, real project end to end.
- Get feedback, refactor, and add tests.
- Ship it publicly and document what you learned.
- Repeat with a slightly harder project each time.
Build It with a World-Class Full Stack Developer
Sandeep Kumar Chaudhary is a full stack world-class developer. If you want to turn this into a real, production-ready product, get in touch — message directly on WhatsApp at +9779802348957 for a fast, no-pressure consult.
You can also explore the projects already shipped to thousands of users, or start a conversation here.
Final Thoughts
For AI, inference at scale usually dominates lifetime energy — invest in quantization, distillation, batching, and caching, not just efficient training. The developers and teams who win in 2026 pair strong fundamentals with consistent shipping. Start small, stay curious, build in public, and revisit this guide as your skills grow.
Sources and Further Reading
Frequently Asked Questions
What is training ai?
Sustainable data centers attack energy waste across cooling, power delivery, and hardware. Modern facilities raise cold-aisle temperatures, use free-air and evaporative or warm-water liquid cooling, and contain hot and cold aisles to shrink the fraction of energy spent moving heat rather than computing. This guide covers training AI end to end — core concepts, best practices, concrete data, and a step-by-step approach you can apply right away.
Which cloud region should I choose to lower emissions?
Prefer regions whose grids are dominated by clean generation, such as hydro-heavy or nuclear-heavy grids, and consult provider sustainability pages and services like Electricity Maps for regional carbon intensity. Cloud providers also publish which regions are lower-carbon, and some offer the grid's carbon intensity directly. Balance carbon against latency, data residency, and cost, since the cleanest region is not always viable for a given application.
Does moving to the cloud automatically reduce my carbon footprint?
Often but not automatically. Hyperscale clouds typically run at higher utilization, use more efficient hardware, and have lower PUE than a self-managed server room, so migrating a lightly used on-premises workload usually cuts energy. However, the cloud's elasticity can also encourage over-provisioning and always-on resources, and the actual carbon depends on the region's grid and the provider's energy sourcing. You still need to right-size, choose clean regions, and shut down idle capacity to realize the benefit.
What is the difference between green software and sustainable software?
The terms are often used interchangeably, but "green software" narrowly means software designed to emit less carbon through energy efficiency, hardware efficiency, and carbon awareness, as framed by the Green Software Foundation. "Sustainable software" is sometimes used more broadly to also include social and economic sustainability, such as maintainability, longevity, and equitable access. In everyday engineering practice most people mean carbon and energy reduction regardless of which term they use.
Is training AI models the main source of AI's carbon emissions?
Training a large model has a big one-time energy cost, but for widely used models the cumulative energy of serving inference to millions of users typically dominates over the model's lifetime. That is why efficiency work increasingly targets deployment through quantization, distillation, batching, and caching, not just the training run. The right emphasis depends on usage: a rarely queried research model may be training-dominated, while a popular production model is almost always inference-dominated.
Sandeep Kumar Chaudhary
Full Stack Software Developer· Nepal's SEO, AEO, GEO & AIO expert and share-market educator. More about me