new framework paper to review and ready to be published by marisolpalmero · Pull Request #2041 · cncf/toc

marisolpalmero · 2026-02-12T18:17:24Z

it could be great if the formatting of the tables could be reviewed by a technical writer, making it more attractive for the reader. thanks

nate-double-u

Comment removed. (Sorry for the noise)

nate-double-u · 2026-02-17T18:56:43Z

+
+[^22]:  [European Union, AI Act: risk, transparency, obligations](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)
+
+[image1]: <data:image/png;base64,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>


Is there a better way to provide these images? I don't think they are showing up correctly in the rendered page

which format will you required?

I don't think we have a requirement. PNG is good. As it is, these images aren't showing up.

@nate-double-u could you review this again?
Images are now getting rendered correctly.

Images appear now. I'll take myself off the reviewers list as this isn't my area of expertise.

Hi @nate-double-u, I don't think that removing yourself from the reviewers list worked, could you check again please? The requested changes review from you is still there, which will block the PR from getting merged. You should be able to dismiss it with a comment. Thank you for looking into this!

Signed-off-by: Payam Mohajeri <payamohajeri@users.noreply.github.com>

marisolpalmero · 2026-03-19T14:41:34Z

hi, which is the next step to get the document published?

marisolpalmero · 2026-04-17T10:57:22Z

Good to see that we are on next phase, with reviews pending. Could we have an update on next steps? many thanks!

payamohajeri · 2026-04-17T15:15:08Z

Hi Marisol, I think this topic is taken over by @GenPage and TAG Infrastructure for review. #1675 (comment)

I just joined the tag channel on slack offering my support in case something is needed from us.

xing-yang · 2026-05-24T14:53:54Z

+| Product Owner | Manage demand and feature usage patterns | Inferences per user action |
+| Governance Lead | Define incentives, controls, and compliance evidence | Sustainability OKRs met |
+
+This table can be adapted per organization, but each persona should have at least one measurable KPI that links day-to-day decisions to sustainability outcomes.


Suggest adding the following:

Cluster Operator: Add explicit cluster-level efficiency responsibilities

Security Engineer: Include sustainability in security scanning and policy enforcement

xing-yang · 2026-05-24T14:59:07Z

+| OpenTelemetry [^14] | Metrics & observability | Monitor sustainability KPIs |
+| Kueue [^15] | Job queuing & resource sharing | Efficient resource utilization for AI training |
+
+Taken together, these projects make sustainability actionable within the existing cloud native control plane: measure resource use, expose it as telemetry, and use scaling and scheduling to reduce waste. This enables teams to integrate sustainability KPIs into operational workflows (dashboards, alerts, policy gates) rather than treating them as an external reporting step.[^16]


Suggest adding the following:

Cluster API: Multi-cloud efficiency and right-sizing

Crossplane: Infrastructure provisioning with sustainability policies

Flux/ArgoCD: GitOps for sustainable deployment workflows

xing-yang · 2026-05-24T15:18:37Z

+Even as hardware becomes more efficient, overall demand can still rise due to increased model scale and usage. The AI Index 2025 reports rapid improvements in hardware energy efficiency, while also noting that the power required for training has continued to increase.5 This reinforces why the deployment environment and the control plane matter: sustainability improvements depend on measuring the right signals and then using orchestration and scheduling to reduce idle capacity and unnecessary data movement.
+
+For Kubernetes environments, Kepler-based approaches are one practical path to connect workload operations to energy-aware optimization workflows.16 At the systems level, the IETF GREEN working group provides a standards-oriented framing for energy measurement and control in ICT systems.[^19]
+


Can you add content regarding container runtime optimization, storage sustainability, and network efficiency?

angellk · 2026-06-08T20:10:53Z

Hi @marisolpalmero -- it is not clear which TAG has worked on this with you or which TAG leads have already reviewed. Could you please open this initiative under a TAG after TAG Chairs and Tech Leads review?

angellk · 2026-08-04T15:25:50Z

@marisolpalmero re-opening per @GenPage - he will have a TAG Infrastructure Tech Lead assigned and help drive this with you. 🙏

We asked Nate to remove the already addressed comment as it was blocking this PR to get merged.

didiViking · 2026-08-05T16:36:43Z

We invited TCG-environmental-sustainability and Green Software Foundation for reviews. Currently this PR looks pretty much completed.

JulianLegler

I think this white paper reflects the current state of art around these topics very well and tries to go beyond. I am more of an expert on the sustainability part and recently started diving deeper into the AI part, but it reads complete and well from my point of view.

Just one minor idea for argumentative consistency and some syntax and structure things. But I am not sure what the intended final publication format is, so it's probably more about the content.

JulianLegler · 2026-08-06T11:39:40Z

+
+</div>
+
+| TAG / Group Home: [https://tag-runtime.cncf.io/wgs/cnaiwg/](https://tag-runtime.cncf.io/wgs/cnaiwg/)  Authors (listed alphabetically): Adel Zaalouk Andrew Block, Red Hat Marisol Palmero Nimisha Mehta Payam Mohajeri Prateek Kumar Viktor Lu Vincent Caldeira, Red Hat | GitHub Issues: [CNCF TOC Issue 1675](https://github.com/cncf/toc/issues/1675) |


Author names are not clearly separated. Probably beneficial to have them separated by commas or other means to make sure no names get mixed up.
Red Hat is mentioned two times.

JulianLegler · 2026-08-06T11:47:28Z

+At the same time, the computational intensity of AI systems continues to rise. The Stanford AI Index 2025 reports that training compute for notable AI models now doubles approximately every five months, while the power required for training increases annually. Although hardware efficiency continues to improve, these gains are currently outpaced by increases in model scale, training frequency, and deployment volume.[^5]
+
+The impact is global. The International Energy Agency projects that data-center electricity demand will nearly double to around 945 TWh by 2030, approaching 3% of global electricity consumption, with significant growth across North America, Europe, and Asia-Pacific.3
+


Do we want a argumentation chain from energy consumption to carbon impact? Carbon is picked up in the next chapters but this chapter only provides sources for energy increase. One could think that maybe future energy consumption can be achieved carbon neutral. But I think there are recent papers suggesting that production of green energy plants can not keep up with the current rising energy demands. So therefore, increasing energy consumption actually means more carbon because we cannot build green energy that fast to keep up.

https://dl.acm.org/doi/10.1145/3757892.3757899

Possible text snipped: "Despite improvements in hardware efficiency and increased adoption of renewable energy sources, the total data center consumption demands in 2030 are estimated to grow 4 times as fast as the decarbonization of the energy grids. Therefore, leading to an increase in carbon emissions."

Although I think this study is very America-centric.

JulianLegler · 2026-08-06T12:10:14Z

+
+At the same time, the computational intensity of AI systems continues to rise. The Stanford AI Index 2025 reports that training compute for notable AI models now doubles approximately every five months, while the power required for training increases annually. Although hardware efficiency continues to improve, these gains are currently outpaced by increases in model scale, training frequency, and deployment volume.[^5]
+
+The impact is global. The International Energy Agency projects that data-center electricity demand will nearly double to around 945 TWh by 2030, approaching 3% of global electricity consumption, with significant growth across North America, Europe, and Asia-Pacific.3


Suggested change

The impact is global. The International Energy Agency projects that data-center electricity demand will nearly double to around 945 TWh by 2030, approaching 3% of global electricity consumption, with significant growth across North America, Europe, and Asia-Pacific.3

The impact is global. The International Energy Agency projects that data-center electricity demand will nearly double to around 945 TWh by 2030, approaching 3% of global electricity consumption, with significant growth across North America, Europe, and Asia-Pacific.[^3]

JulianLegler · 2026-08-06T12:12:24Z

+4. **Act**: Autoscale, binpack, batch, or shift workloads.  
+5. **Report**: Sustainability KPIs and governance evidence; feed back into measurement.
+
+Even as hardware becomes more efficient, overall demand can still rise due to increased model scale and usage. The AI Index 2025 reports rapid improvements in hardware energy efficiency, while also noting that the power required for training has continued to increase.5 This reinforces why the deployment environment and the control plane matter: sustainability improvements depend on measuring the right signals and then using orchestration and scheduling to reduce idle capacity and unnecessary data movement.


Suggested change

Even as hardware becomes more efficient, overall demand can still rise due to increased model scale and usage. The AI Index 2025 reports rapid improvements in hardware energy efficiency, while also noting that the power required for training has continued to increase.5 This reinforces why the deployment environment and the control plane matter: sustainability improvements depend on measuring the right signals and then using orchestration and scheduling to reduce idle capacity and unnecessary data movement.

Even as hardware becomes more efficient, overall demand can still rise due to increased model scale and usage. The AI Index 2025 reports rapid improvements in hardware energy efficiency, while also noting that the power required for training has continued to increase.[^5] This reinforces why the deployment environment and the control plane matter: sustainability improvements depend on measuring the right signals and then using orchestration and scheduling to reduce idle capacity and unnecessary data movement.

JulianLegler · 2026-08-06T12:14:48Z

+
+Even as hardware becomes more efficient, overall demand can still rise due to increased model scale and usage. The AI Index 2025 reports rapid improvements in hardware energy efficiency, while also noting that the power required for training has continued to increase.5 This reinforces why the deployment environment and the control plane matter: sustainability improvements depend on measuring the right signals and then using orchestration and scheduling to reduce idle capacity and unnecessary data movement.
+
+For Kubernetes environments, Kepler-based approaches are one practical path to connect workload operations to energy-aware optimization workflows.16 At the systems level, the IETF GREEN working group provides a standards-oriented framing for energy measurement and control in ICT systems.[^19]


Suggested change

For Kubernetes environments, Kepler-based approaches are one practical path to connect workload operations to energy-aware optimization workflows.16 At the systems level, the IETF GREEN working group provides a standards-oriented framing for energy measurement and control in ICT systems.[^19]

For Kubernetes environments, Kepler-based approaches are one practical path to connect workload operations to energy-aware optimization workflows.[^16] At the systems level, the IETF GREEN working group provides a standards-oriented framing for energy measurement and control in ICT systems.[^19]

JulianLegler · 2026-08-06T12:16:01Z

+
+# Conclusion {#conclusion}
+
+Sustainable AI in cloud native environments is primarily an operations and governance problem: AI demand is rising, and the infrastructure footprint depends on how systems are deployed, measured, and continuously optimized. Recent analyses show data-centre electricity demand growing rapidly and projected to increase substantially toward 2030, with AI as a major driver.3 At the same time, leading AI systems continue to increase compute requirements, reinforcing the need to treat sustainability as part of the platform control plane rather than an external reporting activity.5


Suggested change

Sustainable AI in cloud native environments is primarily an operations and governance problem: AI demand is rising, and the infrastructure footprint depends on how systems are deployed, measured, and continuously optimized. Recent analyses show data-centre electricity demand growing rapidly and projected to increase substantially toward 2030, with AI as a major driver.3 At the same time, leading AI systems continue to increase compute requirements, reinforcing the need to treat sustainability as part of the platform control plane rather than an external reporting activity.5

Sustainable AI in cloud native environments is primarily an operations and governance problem: AI demand is rising, and the infrastructure footprint depends on how systems are deployed, measured, and continuously optimized. Recent analyses show data-centre electricity demand growing rapidly and projected to increase substantially toward 2030, with AI as a major driver.[^3] At the same time, leading AI systems continue to increase compute requirements, reinforcing the need to treat sustainability as part of the platform control plane rather than an external reporting activity.[^5]

JulianLegler · 2026-08-06T12:16:18Z

+
+Sustainable AI in cloud native environments is primarily an operations and governance problem: AI demand is rising, and the infrastructure footprint depends on how systems are deployed, measured, and continuously optimized. Recent analyses show data-centre electricity demand growing rapidly and projected to increase substantially toward 2030, with AI as a major driver.3 At the same time, leading AI systems continue to increase compute requirements, reinforcing the need to treat sustainability as part of the platform control plane rather than an external reporting activity.5
+
+To make sustainability actionable, telemetry must be used to trigger concrete optimization decisions. The CNCF ecosystem provides the primitives to measure and act (energy telemetry, cost signals, autoscaling, scheduling, and observability), but the goal is to close the loop between measurement and change.16 Actionable steps for sustainable cloud native AI:


Suggested change

To make sustainability actionable, telemetry must be used to trigger concrete optimization decisions. The CNCF ecosystem provides the primitives to measure and act (energy telemetry, cost signals, autoscaling, scheduling, and observability), but the goal is to close the loop between measurement and change.16 Actionable steps for sustainable cloud native AI:

To make sustainability actionable, telemetry must be used to trigger concrete optimization decisions. The CNCF ecosystem provides the primitives to measure and act (energy telemetry, cost signals, autoscaling, scheduling, and observability), but the goal is to close the loop between measurement and change.[^16] Actionable steps for sustainable cloud native AI:

JulianLegler · 2026-08-06T12:16:47Z

+5. Define a small set of operational sustainability KPIs  
+6. Instrument and standardize telemetry
+
+The practical outcome of this approach is a measurable loop: telemetry reveals waste (idle accelerators, oversized services, inefficient placement), and platform controls convert those insights into action (autoscaling, scheduling, placement, and time shifting). This makes sustainability improvements repeatable across teams and workloads, while aligning operational behavior with the growing scale of AI and evolving governance expectations.3 5 23  


Suggested change

The practical outcome of this approach is a measurable loop: telemetry reveals waste (idle accelerators, oversized services, inefficient placement), and platform controls convert those insights into action (autoscaling, scheduling, placement, and time shifting). This makes sustainability improvements repeatable across teams and workloads, while aligning operational behavior with the growing scale of AI and evolving governance expectations.3 5 23

The practical outcome of this approach is a measurable loop: telemetry reveals waste (idle accelerators, oversized services, inefficient placement), and platform controls convert those insights into action (autoscaling, scheduling, placement, and time shifting). This makes sustainability improvements repeatable across teams and workloads, while aligning operational behavior with the growing scale of AI and evolving governance expectations.[^3] [^5] [^23]

JulianLegler · 2026-08-06T12:18:30Z

+| Environment | Common fit | Dominant sustainability drivers | Primary levers | Hardware access considerations |
+| :---- | :---- | :---- | :---- | :---- |
+| Public cloud | Elastic training/inference; burst capacity | Region energy mix; overprovisioning; data egress | Right-sizing; autoscaling; placement policies; energy telemetry | Access via rentals; quotas/capacity can constrain peaks |
+| Private / on-prem | Data control; predictable workloads | Utilization; cooling efficiency; upgrade cycles | Consolidation; scheduling; instrumentation; capacity management | May have older accelerators or no accelerators available 5 |


Suggested change

| Private / on-prem | Data control; predictable workloads | Utilization; cooling efficiency; upgrade cycles | Consolidation; scheduling; instrumentation; capacity management | May have older accelerators or no accelerators available 5 |

| Private / on-prem | Data control; predictable workloads | Utilization; cooling efficiency; upgrade cycles | Consolidation; scheduling; instrumentation; capacity management | May have older accelerators or no accelerators available [^5] |

JulianLegler · 2026-08-06T12:18:55Z

+
+Where an AI system runs largely determines its sustainability profile. Deployment choices influence utilization (idle vs busy resources), cooling efficiency, the electricity grid mix, and how much data must move across networks. The same model can therefore have very different operational impacts depending on whether it is served on-device, on-prem, in a public cloud region, or across a hybrid footprint.
+
+Deployment decisions are also constrained by hardware reality. Some organizations operate older accelerators, while others lack accelerator capacity entirely and cannot obtain enough GPUs to meet demand. The AI Index 2025 highlights that cutting-edge AI increasingly requires compute and financial resources that are not available to academia, with leading models predominantly produced by industry.5 This access gap matters for sustainability: many efficiency strategies depend on having the right hardware and then keeping it highly utilized. The table below summarizes common environment trade-offs and the levers typically available in each setting.


Suggested change

Deployment decisions are also constrained by hardware reality. Some organizations operate older accelerators, while others lack accelerator capacity entirely and cannot obtain enough GPUs to meet demand. The AI Index 2025 highlights that cutting-edge AI increasingly requires compute and financial resources that are not available to academia, with leading models predominantly produced by industry.5 This access gap matters for sustainability: many efficiency strategies depend on having the right hardware and then keeping it highly utilized. The table below summarizes common environment trade-offs and the levers typically available in each setting.

Deployment decisions are also constrained by hardware reality. Some organizations operate older accelerators, while others lack accelerator capacity entirely and cannot obtain enough GPUs to meet demand. The AI Index 2025 highlights that cutting-edge AI increasingly requires compute and financial resources that are not available to academia, with leading models predominantly produced by industry.[^5] This access gap matters for sustainability: many efficiency strategies depend on having the right hardware and then keeping it highly utilized. The table below summarizes common environment trade-offs and the levers typically available in each setting.

marisolpalmero requested review from a team as code owners February 12, 2026 18:17

marisolpalmero mentioned this pull request Feb 12, 2026

[Initiative]: Cloud Native AI Sustainability Papers #1675

Open

nate-double-u reviewed Feb 17, 2026

View reviewed changes

nate-double-u previously requested changes Feb 17, 2026

View reviewed changes

payamohajeri requested a review from a team as a code owner March 9, 2026 17:30

payamohajeri force-pushed the 1675-ai-sustainability-papers branch from c83ff8b to acc24fb Compare March 9, 2026 17:33

new framwork paper

cf71736

Signed-off-by: Payam Mohajeri <payamohajeri@users.noreply.github.com>

payamohajeri force-pushed the 1675-ai-sustainability-papers branch 2 times, most recently from 6ff9d6f to 0863a88 Compare March 9, 2026 18:40

payamohajeri added 2 commits March 9, 2026 18:44

docs: update whitepaper and extract embedded images to files

975c18e

Signed-off-by: Payam Mohajeri <payamohajeri@users.noreply.github.com>

docs: center images in whitepaper

e6f214a

Signed-off-by: Payam Mohajeri <payamohajeri@users.noreply.github.com>

payamohajeri force-pushed the 1675-ai-sustainability-papers branch from 0863a88 to e6f214a Compare March 9, 2026 18:45

payamohajeri added 2 commits March 9, 2026 19:19

docs: replace images with higher quality versions

fdb1c81

Signed-off-by: Payam Mohajeri <payamohajeri@users.noreply.github.com>

docs: add Kueue to CNCF projects table and renumber footnotes

d77ac46

Signed-off-by: Payam Mohajeri <payamohajeri@users.noreply.github.com>

github-actions Bot added needs-triage Indicates an issue or PR that has not been triaged yet (has a 'triage/foo' label applied) needs-kind Indicates an issue or PR that is missing an issue type or kind (a kind/foo label) labels Apr 22, 2026

github-project-automation Bot added this to TAG Workloads Foundation, TAG Operational Resilience, CNCF TOC Board, TAG Infrastructure, TAG Security and Compliance and TAG Developer Experience Apr 22, 2026

github-actions Bot added the needs-group Indicates an issue or PR that has not been assigned a group (toc or tag/foo label applied) label Apr 22, 2026

github-project-automation Bot moved this to New in CNCF TOC Board Apr 22, 2026

xing-yang reviewed May 24, 2026

View reviewed changes

angellk self-requested a review June 2, 2026 15:51

angellk closed this Jun 8, 2026

github-project-automation Bot moved this from New to Done in CNCF TOC Board Jun 8, 2026

riaankleinhans mentioned this pull request Jun 16, 2026

[TOC Meeting][Public] 2026-06-16 #2069

Closed

angellk reopened this Aug 4, 2026

JulianLegler reviewed Aug 6, 2026

View reviewed changes


		[^22]: [European Union, AI Act: risk, transparency, obligations](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)

		[image1]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAO8AAAAmCAYAAAAsl8zFAAAI2UlEQVR4Xu2dzY0cRxKF0wSZQBNkQmohA2hC38UDLzvQsUyQCXPnRR6oAEG6SsBCe24T6IHEYHXORH0VP5nZzQtVH/CgyXgvoqoaE8OeIQcq/3n34e9RPf3215DY36NycnISw6XpEZczE/t7xPs8OTkBXJoecTkzsb9HvM8JZIaliJFsBmc1faNDDrWM3QevQehrLa+xu9Fze56T99LDz6W/h/N7pfE8rx7BntZXjXqmr3J5+ZCeLHoyGZzhKfrkrmXsPjib0I8U3VcGZ2WM5gX2RH3M9Urjeayzz8LLV8PL9FUt78dyfMBMJPMz2N8ji1ryjCabSb9HM3BGNmck22BP1MdcrzSzHlmLn6+Gl8lZ3h8+vP/+3Yfqicv59Pv/3/746/+qJ/Zrfffuw/Ph+uPL+1yOD9crTeRlsLdpLdtbU9ab5IsOqWXsPjiT0O/VKOzP5vTmGsxnvcz0SjPiXXfuHmY1VdV7ZS+vLNVnM4AL/P6XP6bebn263sLrTywvHyzqjzK9Mwj7ol7mrGwtsU+yeZkvMCP6dpfIYX9TVRkNcxnMj/Q2RvuyfOY3olxVddGiTRcuTe/yCo9Y4AcsL1+Unl4vMzpHkGce7WOePVXV6VlEs4TMb/xZ+rMW7M3m9GQazFLXl2QM+zKyfOYLzDBXVV20aNOFSzOyvP/97a9v713gL7C89zAziz3XnevDPk1VdXoW0Swh8zUjWcJeimS+xspatYzRnixv/ayFZH4te3/RpguXZmR5hXsX+Ctc3l6ivqrq9CyiWULma34q+2zduTG8jiVN5Gmuxc5atYzRnp58lFlV3fKFWvb+ok0XLs3o8gr3LPC5vC+qypOPR2ZyFsl8orPWD9Q8rOuw9nyrW55HlIs8iy+RZ+YaeBa1HHORNrg0M8srzC7wnctbi/dgc8zMmukR2FeVJx+PzOQskvlkNN+w+qK3lV6dRLnIs/hSeS/n1TW1HHORNrg0s8srzCzwncsr2A82x8ysmR4h6quqTs8imiVkPtHZZW+FeNdhvXlWjWSZteQZzUhW6M1bOda8dzG1HLORNrhsTS+BBxAt8KdrLbz24PXtB5tjZtZMj5D1ae8tPA2/R1137kZ2Lc1IlkS99CxZMNOjiJGs0Ju/lmOW8qhln1u06cKlmVygEJnlLfAXWN6eXi8zOkeI3hZ69PyVTOY3enI9mcZIlmS99ClCv1fen3ACsxkjeWYpj1r2uUWbLlwaivkZ2ixrgR+wvEs5vkiyHB7Ri8k5vbAv6pXXgNmqAzeY8ejJ9WQE5qKsRdb7czlmojz9EXn05hojeWZ7+2rZZxdtunBpLLFnFD2LC/x0//IKfKGa5JMly2giL2Ipx14Rv1jRj67DjGhVvvWnd+8szUXVqVF6+pnx8vToE2b52jeYy7g339NXyz67aNOFS3PTR9bYNwJn6QV+eszyCnyxesQ/oelHWraWF+j3KoLZTB7M9WiG3hnMWfnI8+jp6clo7s2LrjpgUMuxJxeXRiS/TPDpv3+yLleZIfop9NPjllc4PqCvN1vLDmYiLVvLjuxtIdUDezxFMJtplpE5zOr8m8CL6OnpyWhG89Y7ooxajj25uDQiWd5P5uefEtOTRRRvBs562r5APHJ5G8cHfdXlNXaA2UjL1mLyXI55rVEu5Tijqb6kfNjj6V5G53l5r54hn5tZb+aT0bww2lPLsScXl0bUllewFrh5M3CWJ/adnJwALo1IL6/ABdbeDLyeJfacnJwALo2IyyvoBaY3A69JMX9ycgK4NN7yCm2BWZ+F1z2X9+RkAC5NtLwC/572Xnjtc3lPTjrhLwT0/GICvwe+Z9ke8IsJJyf/Trg0PcsrPGqBz+U9OZmES9O7vMLTA/4hxwOXd/u7r70smLFytcS+9lan7vV7dUHXOYNqeOdo/oK6wPlWv4bZqI9+lFlQr+XY085vVa3BrK5RES1Tnbon5gjzIutbUa9f2DwuzcjyCvcu8AOW92M5vhhaDXmB6Fk5oao6PUF7q1OnrAzRdfZTDe8sWlRdyOqWPJizekZe81ZbUK/lmPdmCFad1436hefiZ9hPMadhltJ4dWGrc2lGl1e4Z4EfsLzeQ7KW5XS9ou7NEa2Gp+EMnjWjdYGens+eVluMmpWNsHpINJt1696EWvws69rrrbEuRPM1kU/Pm6f/VZgmuoetxqWZWV5hdoHvXF7v4ci1xLnm/XQ7V1VjH+ur4Wkq6tbMxmhdoKfne95yO8vbTmZ6yfquJc40r73mvLdGLcc57axFT2PVhFa/OHVrvmbE41ljXYfXfz54XJrZ5RVmFvhBy5thvTga+hW1Just+vq5Y8O6BmfzrBmtC/T0fF6rfbzg7M2O4HzO4pnQbx8vLXCjluMcq+/Z8BpWTeA9WDX6mhGPZ0K/navrcWlEt/8FiSyVqaftlwk87ZZXxH5olWtS6kYj+FAeLedl6VectS/S/nrLCMxpXYwMGa0L9Di/fax/22WBZ81u9Yp6Q/dS9C3ot4+XFrhRy3GOPlecmfVqAu9B13iO+i3o8Uzo6zPvYfuYS9MjLmcm9vfodpMZfCgP/Ylr0TzrbbP2rdp6O+uaJSuj0T/cIV5doMf5b1ATLTfvomqk1SvqDa+vcS1xhvei39Vo9H2zxrOVFaya0OrvcfZEvLpAj2eNdQ3v/CouTY+4nJnY3yN10xntYWRBrTrPnG3VK2p6sS63Wjuvt7OuZVg56z4aXl2gZ83RNdHieJpWq6g3rB6SzfbqVq19YdU1jZ7peVZN1zmDurwkN9ivoWddT8h+YGXVNnFpesTlzMT+HuGmI/YPdJSGnperRt07r0Ytg9f27qMx4nmzdH3ZW4d70HrzGtthXYOM/FWRQN/LWjUhy1uS5RFWVSOXYntWrWF5vDalsWr715NL0yMuZyb29wg3neF9klgwY+Vq8b1G81ej1gPvIeqLfHrRvFZfUBd4L1a/pifT4Nyojzkr69W9bz04L8rIDIuox8LzeA9WRvC81z4uTY+4nJnY3yPe8cnJyZ5/AJDg8hZwHHckAAAAAElFTkSuQmCC>

		Even as hardware becomes more efficient, overall demand can still rise due to increased model scale and usage. The AI Index 2025 reports rapid improvements in hardware energy efficiency, while also noting that the power required for training has continued to increase.5 This reinforces why the deployment environment and the control plane matter: sustainability improvements depend on measuring the right signals and then using orchestration and scheduling to reduce idle capacity and unnecessary data movement.

		For Kubernetes environments, Kepler-based approaches are one practical path to connect workload operations to energy-aware optimization workflows.16 At the systems level, the IETF GREEN working group provides a standards-oriented framing for energy measurement and control in ICT systems.[^19]


		</div>

		\| TAG / Group Home: [https://tag-runtime.cncf.io/wgs/cnaiwg/](https://tag-runtime.cncf.io/wgs/cnaiwg/) Authors (listed alphabetically): Adel Zaalouk Andrew Block, Red Hat Marisol Palmero Nimisha Mehta Payam Mohajeri Prateek Kumar Viktor Lu Vincent Caldeira, Red Hat \| GitHub Issues: [CNCF TOC Issue 1675](https://github.com/cncf/toc/issues/1675) \|

		At the same time, the computational intensity of AI systems continues to rise. The Stanford AI Index 2025 reports that training compute for notable AI models now doubles approximately every five months, while the power required for training increases annually. Although hardware efficiency continues to improve, these gains are currently outpaced by increases in model scale, training frequency, and deployment volume.[^5]

		The impact is global. The International Energy Agency projects that data-center electricity demand will nearly double to around 945 TWh by 2030, approaching 3% of global electricity consumption, with significant growth across North America, Europe, and Asia-Pacific.3


		# Conclusion {#conclusion}

		Sustainable AI in cloud native environments is primarily an operations and governance problem: AI demand is rising, and the infrastructure footprint depends on how systems are deployed, measured, and continuously optimized. Recent analyses show data-centre electricity demand growing rapidly and projected to increase substantially toward 2030, with AI as a major driver.3 At the same time, leading AI systems continue to increase compute requirements, reinforcing the need to treat sustainability as part of the platform control plane rather than an external reporting activity.5


		Sustainable AI in cloud native environments is primarily an operations and governance problem: AI demand is rising, and the infrastructure footprint depends on how systems are deployed, measured, and continuously optimized. Recent analyses show data-centre electricity demand growing rapidly and projected to increase substantially toward 2030, with AI as a major driver.3 At the same time, leading AI systems continue to increase compute requirements, reinforcing the need to treat sustainability as part of the platform control plane rather than an external reporting activity.5

		To make sustainability actionable, telemetry must be used to trigger concrete optimization decisions. The CNCF ecosystem provides the primitives to measure and act (energy telemetry, cost signals, autoscaling, scheduling, and observability), but the goal is to close the loop between measurement and change.16 Actionable steps for sustainable cloud native AI:

	\| Private / on-prem \| Data control; predictable workloads \| Utilization; cooling efficiency; upgrade cycles \| Consolidation; scheduling; instrumentation; capacity management \| May have older accelerators or no accelerators available 5 \|
	\| Private / on-prem \| Data control; predictable workloads \| Utilization; cooling efficiency; upgrade cycles \| Consolidation; scheduling; instrumentation; capacity management \| May have older accelerators or no accelerators available [^5] \|


		Where an AI system runs largely determines its sustainability profile. Deployment choices influence utilization (idle vs busy resources), cooling efficiency, the electricity grid mix, and how much data must move across networks. The same model can therefore have very different operational impacts depending on whether it is served on-device, on-prem, in a public cloud region, or across a hybrid footprint.

		Deployment decisions are also constrained by hardware reality. Some organizations operate older accelerators, while others lack accelerator capacity entirely and cannot obtain enough GPUs to meet demand. The AI Index 2025 highlights that cutting-edge AI increasingly requires compute and financial resources that are not available to academia, with leading models predominantly produced by industry.5 This access gap matters for sustainability: many efficiency strategies depend on having the right hardware and then keeping it highly utilized. The table below summarizes common environment trade-offs and the levers typically available in each setting.

Uh oh!

Conversation

marisolpalmero commented Feb 12, 2026

Uh oh!

nate-double-u left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nikimanoledaki Jun 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

marisolpalmero commented Mar 19, 2026

Uh oh!

marisolpalmero commented Apr 17, 2026

Uh oh!

payamohajeri commented Apr 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

angellk commented Jun 8, 2026

Uh oh!

angellk commented Aug 4, 2026

Uh oh!

didiViking commented Aug 5, 2026

Uh oh!

JulianLegler left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

nate-double-u left a comment •

edited

Loading

nikimanoledaki Jun 1, 2026 •

edited

Loading

payamohajeri commented Apr 17, 2026 •

edited

Loading

JulianLegler left a comment •

edited

Loading