> Markdown version of [/videos/570-optimizing-your-ai-ml-workloads-for-sustainability?t=2439](https://www.wearedevelopers.com/videos/570-optimizing-your-ai-ml-workloads-for-sustainability?t=2439). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Optimizing your AI/ML workloads for sustainability

Deployment accounts for nearly 90% of a machine learning model's energy usage. Discover how to systematically engineer your cloud architecture to balance powerful AI capabilities with planetary limitations.

- **Speakers:** Sohan Maheshwar
- **Event:** WeAreDevelopers LIVE
- **Published:** April 21, 2023
- **Duration:** 46:31
- **URL:** https://www.wearedevelopers.com/videos/570-optimizing-your-ai-ml-workloads-for-sustainability

## Summary

Machine learning models are growing exponentially in parameter size and complexity, driving staggering computational demands and significant carbon footprints. While migrating workloads from on-premises data centers to the cloud can initially reduce energy usage by up to 80%, practitioners must actively engineer for environmental efficiency under the shared responsibility model. Balancing powerful AI capabilities with planetary limitations requires a systematic, sustainability-focused approach across the entire machine learning lifecycle, from initial problem framing to production inference.

Sustainable AI architecture begins with utilizing pre-trained models from Amazon SageMaker or Hugging Face instead of training from scratch, alongside migrating compute to purpose-built, highly efficient hardware like AWS Trainium, Inferentia, and Graviton3 processors. During data processing, employing efficient columnar formats like Parquet, upgrading log compression from standard gzip to zstandard (zstd), and automating Amazon S3 storage lifecycle classification drastically minimizes wasted compute and storage. In the model development phase, defining "acceptable performance criteria" prevents excessive resource consumption spent chasing marginal accuracy gains. Because deployment accounts for nearly 90% of an ML model's operational cost and energy usage, engineering teams must smooth traffic peaks utilizing queues and buffers to properly right-size instances. Further optimizations include hosting multiple models behind a single endpoint and negotiating impact-friendly service level agreements (SLAs) utilizing asynchronous batch transforms where real-time inference is unnecessary.

Tracking these improvements requires moving beyond aggregate carbon numbers to calculate normalized proxy metrics—such as emissions per user—using the AWS Customer Carbon Footprint Tool and the Sustainability Pillar of the AWS Well-Architected Framework. Beyond internal cloud optimizations, machine learning acts as a vital tool for solving overarching global environmental challenges. By leveraging open datasets from the Amazon Sustainability Data Initiative (ASDI) and deploying ML at the extreme edge with AWS Snowball or IoT Greengrass, organizations securely analyze climate data, track ocean health, and monitor endangered species in remote, disconnected environments.

**Keywords:** ai workload sustainability, machine learning lifecycle optimization, cloud carbon footprint reduction, AWS shared responsibility model, S3 intelligent storage tiering, log compression optimization, purpose-built AI hardware, AWS Trainium and Inferentia, Graviton3 processors, ML inference right-sizing, asynchronous impact-friendly SLAs, AWS well-architected sustainability pillar, Amazon sustainability data initiative, edge machine learning deployment, normalized proxy emissions metrics

## Chapters

1. **High carbon footprint of complex machine learning models** (00:50) — Moving toward massive neural networks exponentially increases the energy consumption and carbon emissions required for training data.
1. **Reducing energy usage by migrating workloads to cloud infrastructure** (05:24) — Cloud data centers roughly optimize structural energy consumption compared to traditional on-premises hardware systems.
1. **Adopting the shared responsibility model for cloud computing sustainability** (06:14) — While providers optimize physical infrastructure operations, organizations must strictly govern application code efficiency and data usage to minimize environmental impact.
1. **Integrating sustainability goals into machine learning problem framing** (09:23) — Defining precise optimization targets allows teams to leverage pre-trained models and managed services instead of wasting processing resources on reinvention.
1. **Applying storage lifecycle policies to machine learning data** (12:44) — Automatically transitioning infrequently accessed records to cold storage prevents the ongoing environmental cost cycles of running high-availability server clusters.
1. **Optimizing log storage and file formats for reduced footprints** (16:03) — Implementing advanced compression algorithms and columnar file formats significantly decreases total storage volume and data transfer network overhead.
1. **Selecting purpose-built hardware for model training and inference** (19:08) — Utilizing custom silicon processors designed for intensive algorithmic processing yields inherently higher inference performance per watt than general-purpose compute instances.
1. **Developing machine learning models against acceptable performance thresholds** (21:09) — Establishing realistic architectural accuracy criteria limits the massive computing waste associated with running extended training epochs for marginal statistical gains.
1. **Right-sizing environments and optimizing endpoints during model deployment** (23:51) — Implementing asynchronous buffered application requests and multiplexed model endpoints smooths network traffic peaks entirely to eliminate heavily over-provisioned infrastructure instances.
1. **Negotiating impact-friendly SLAs to minimize idle capacity** (27:18) — Accepting acceptable minor latencies during cold failovers or batch processing runs enables absolute total infrastructure utilization on actively provisioned nodes.
1. **Choosing cloud data center regions powered by renewable energy** (29:18) — Deploying cloud infrastructure deployments closer to large renewable energy initiatives capitalizes entirely on unique regional advantages regarding green power distribution.
1. **Tracking normalized carbon emission metrics in cloud environments** (33:04) — Measuring total systemic energy consumption mathematically relative to product user growth ensures an accurate analytical reflection outlining authentic architectural operational optimizations.
1. **Accessing large environmental datasets to analyze global climate challenges** (36:46) — Processing freely available public remote sensing records intelligently empowers engineering groups everywhere to build sophisticated predictive tools forecasting complex natural global phenomena.
1. **Deploying machine learning models at the edge for conservation** (40:39) — Operating localized computer vision inferencing environments immediately on rugged hardware perimeter hardware facilitates crucial endangered ecosystem analytics regardless of internet constraints.

## Related Moments

- [Balancing heavy compute demands with environmental sustainability goals](https://www.wearedevelopers.com/videos/916-beyond-the-hype-real-world-ai-strategies-panel) (from "Beyond the Hype: Real-World AI Strategies Panel")
- [Balancing human-centric AI collaboration with environmental sustainability practices](https://www.wearedevelopers.com/videos/1016-insight-into-ai-driven-design) (from "Insight into AI-Driven Design")
- [Addressing the sustainability and power consumption of AI](https://www.wearedevelopers.com/videos/2127-five-things-in-tech-that-matter-now-wearedevelopers-world-congress-2026-closing-keynote) (from "Five Things in Tech that Matter Now - WeAreDevelopers World Congress 2026 Closing Keynote")
- [Managing massive power consumption scaling in AI data centers](https://www.wearedevelopers.com/videos/1106-the-future-of-computing-ai-technologies-in-the-exascale-era) (from "The Future of Computing: AI Technologies in the Exascale Era")
- [Assessing the energy demands versus sustainability benefits of computation](https://www.wearedevelopers.com/videos/100155-the-sustainability-race-ai-s-promises-pitfalls-and-potential) (from "The Sustainability Race: AI's Promises, Pitfalls and Potential")
- [Establishing cross-chain accountability for sustainable machine learning usage](https://www.wearedevelopers.com/videos/100155-the-sustainability-race-ai-s-promises-pitfalls-and-potential) (from "The Sustainability Race: AI's Promises, Pitfalls and Potential")

## Related Articles

- [Is Software Development Making the Climate Crisis Worse?](https://www.wearedevelopers.com/magazine/551-is-software-development-making-the-climate-crisis-worse)
- [MLops – Deploying, Maintaining And Evolving Machine Learning Models in Production](https://www.wearedevelopers.com/magazine/115-mlops-deploying-maintaining-and-evolving-machine-learning-models-in-production)
- [MLOps – What’s the deal behind it?](https://www.wearedevelopers.com/magazine/125-mlops-what-s-the-deal-behind-it)
- [Stephan Gillich - Bringing AI Everywhere](https://www.wearedevelopers.com/magazine/489-stephan-gillich-bringing-ai-everywhere)

## Related Jobs

- [Data Scientist](https://www.wearedevelopers.com/jobs/ext/1351648-data-scientist) at **Almedia**
- [Staff, Machine Learning Engineer (L4)](https://www.wearedevelopers.com/jobs/ext/1202639-staff-machine-learning-engineer-l4) at **Twilio**
- [Machine Learning Engineer](https://www.wearedevelopers.com/jobs/ext/588393-machine-learning-engineer) at **Twilio**
- [AI Software Engineer (Germany)](https://www.wearedevelopers.com/jobs/48317-ai-software-engineer-germany) at **Sunhat**
- [Machine Learning Engineer](https://www.wearedevelopers.com/jobs/ext/1355348-machine-learning-engineer) at **TWILIO**
- [Staff Software Engineer, Copilot Experiences](https://www.wearedevelopers.com/jobs/ext/164361-staff-software-engineer-copilot-experiences) at **GitHub**