Posts

Encouraging to see SCI for AI recognized in the G7 French Presidency’s work on measuring and monitoring the environmental impact of AI.

What makes this recognition meaningful is that SCI for AI was shaped through deep collaboration across nearly 20 organizations spanning industry, academia, and sustainability practitioners. The specification was developed as a community effort to extend the ISO/IEC 21031 SCI standard into the AI domain, helping establish a practical and standardized way to measure AI emissions across different AI paradigms and lifecycle stages.

Recognition like this reflects the growing importance of moving from broad discussions on AI sustainability to measurable and actionable approaches that can support better infrastructure, investment, and policy decisions.

A big thank you to everyone in the Green Software Foundation community who contributed to shaping the specification.

Day 7 of Green, Efficient AI is live — Shape the Answer Before You Ask.

Most teams optimise what goes into an AI call. Few optimise what comes back.

Output tokens cost roughly three to five times what input tokens cost, across every major provider. The return trip of the call is where the meter runs fastest — and where the least design effort has been spent.

You decide what the model returns before you call it, or you pay to fix what it returns after.

How to stop paying twice for the same answer, in today’s issue.

Green Software Foundation #greenai #efficentai

Day 8 of Green, Efficient AI is live — Right-Size the Agentic Loop.

A team built an agent to triage bug reports. They expected it to cost about fourteen times a single call.

It cost several times more than that.

Nothing was broken. By turn twelve, the agent was re-sending every search, every file, and every line of its own reasoning — at full price, on every turn. Most of its tokens weren’t spent solving the bug. They were spent re-reading the loop.

A call pays once. A loop pays for everything it remembers, on every turn it takes.

How to right-size it, in today’s issue.

Green Software Foundation #greenai #efficentai

Most people know AI consumes electricity.

But very few realize that AI may also have a water story.

As Agentic AI evolves from answering questions to reasoning, planning, coordinating tools, and acting autonomously, an important question quietly emerges:

What physical resources support intelligent systems?
When AI systems run, infrastructure runs.
Servers generate heat.
Data centres cool infrastructure.
Electricity powers computation.

And in many cases, water becomes part of the story — directly through cooling systems or indirectly through power generation.

What makes this conversation even more interesting is scale.
One interaction feels small.
One extra reasoning step feels insignificant.
One retry seems harmless.

But what happens when millions of people rely on always-on intelligent systems working quietly in the background?

And perhaps the bigger question:

Does location matter?

A litre of water associated with AI infrastructure in a water-rich region is not always the same conversation as infrastructure running in areas already facing drought, groundwater depletion, or increasing scarcity.

Because the future of AI may not simply be about building smarter systems.

It may also be about building systems that are efficient, responsible, and thoughtful about the invisible resources quietly supporting intelligence.

🎙️ Episode 90 of Agentic AI — The Future of Intelligent Systems is now live: Understanding the Water Footprint of Agentic AI

Listen here:
https://lnkd.in/dvx8jnj2

Day 6 of Green, Efficient AI is live — The Cheapest Tokens Are the Ones You Never Retrieve.

A team was sending their model an 8,600-token prompt to answer a question
whose actual answer was one paragraph.

Twenty passages retrieved. One useful. Nineteen paid for in full.

The model was never the problem. The retrieval layer was — and most teams have never looked at it.

Most prompts in production are not written. They are assembled. And the cheapest tokens in any prompt are the ones you never retrieve.

How to find and fix the waste, in today’s issue.

Green Software Foundation #greenai #efficentai

Day 5 of Green, Efficient AI is live — Spend Tokens Like They Cost Money.

What if the largest pool of waste in your AI system is not the model you chose, or the calls you made, but the words you packed into every prompt?

Most production prompts are two to ten times longer than the task requires. The next principle is what to do about it.

Green Software Foundation #greenai #efficentai

Day 4 of Green, Efficient AI is live - Make the Call Count Smaller.

The cheapest, fastest, lowest-carbon inference is the one the system never made.

Most teams focus on model cost. Bigger model. Smaller model. Better prompt.

But one of the biggest opportunities in AI efficiency lives somewhere else:

Calls your system never needed to make in the first place.

The same question answered forty thousand times.
The same document fetched again inside an agent loop.
The same retry sent back to the model for a problem the network should have handled.

Major providers already reduce some waste below the API line through prompt caching.

The bigger opportunity lives above it.

Dedupe → Reuse → Carry → Catch

One ladder for reducing unnecessary inference.

The cheapest call is the one you never make.
The second cheapest is the one you make once and remember.

Day 4 of Green, Efficient AI ↓

Green Software Foundation #greenai #efficentai

What if the future of AI becomes too expensive to sustain?
Not because models are costly.

But because intelligence itself becomes something we consume continuously.

We are moving toward a world where AI systems won’t just answer questions.
They will quietly prepare for meetings, coordinate schedules, negotiate prices, summarize research, monitor health patterns, optimize spending, and make decisions in the background.

And many of these systems won’t work alone.
They will reason.
Retrieve memory.
Call tools.
Collaborate with other agents.
Quietly operating throughout the day.

Now multiply this by millions of people.
And perhaps hundreds of agents per person.

This is where a hidden economy begins to emerge.
Tokens.
Often treated as a technical metric.

But perhaps better understood as the currency of intelligence.
Because every token represents computation.

And computation carries consequences:
💸 Rising operational costs
Which increases…
⚡ Energy consumption
Which requires more…
💧 Water for cooling infrastructure
Leading to growing…
🌍 Carbon emissions
And when systems reason unnecessarily, carry oversized context, or overuse tools…
♻️ Waste multiplies.
As Agentic AI scales, the question may no longer be:

How intelligent can AI become?
But:
How intelligently should intelligence operate?

The latest episode of Agentic AI — The Future of Intelligent Systems explores the hidden economy powering Agentic AI — from personal agents and token demand to cost, energy, water, carbon, and the future of efficient intelligence.

🎧 Listen here:
https://lnkd.in/dHJTjYeG

Day 3 of Green, Efficient AI is live — Right-Size the Model.

One principle, written for everyone who builds, funds, designs, operates, or uses AI.

A frontier model summarising a meeting transcript is a Ferrari delivering a pizza.
It works. It also consumes an order of magnitude more energy, water, and cost than a small model that would have done the same job indistinguishably to the end user.

And sometimes the honest answer is not a smaller model. It is no model.

The smallest technology that solves the task is the right one. Not the most powerful. Not the newest. Not the one the demo used.

Green Software Foundation

Day 2 of Green, Efficient AI is live: The Invisible Bill.

Every AI response carries five costs at once — energy, water, carbon, hardware, money. One sits on a dashboard the team already has open. The other four are simply waiting to be brought into view.

What it means: efficient AI begins the moment all five become visible at the point of decision — every query, every retry, every agent loop, every GPU hour.

Who it’s for: every seat at the table — engineering, platform, architecture, sustainability, business. No single role solves efficient AI alone, and every role has something to bring.

Green Software Foundation

Every AI feature ships an invisible bill.
Cost. Energy. Water. Carbon. Hardware.
Most teams see one.

Almost none see the full picture — and that gap is about to define the next era of how we build software.

Today, I am launching a new LinkedIn newsletter:

Green, Efficient AI

Building AI that is intelligent in what it delivers — and efficient in how it runs.
The newsletter is an almost-daily read built around one small, actionable principle at a time.

No long essays. No abstract frameworks.
Just one idea per issue — useful enough to bring into your work tomorrow.

Waste in AI doesn’t live in one place.
It runs through every layer — from the model to the prompt to the hardware itself.

Which means every role has something to contribute:
Engineers across the lifecycle. Architects. Designers. Platform teams. Sustainability leads. Finance partners. End users.
And increasingly, the agents consuming AI on our behalf.

No single role solves efficient AI alone.
That is the conversation this newsletter is for.

Day 1 is live.

It explains why this newsletter exists and how it works.

If you build AI, fund it, design it, operate it — or simply use it every day — I’d love for you to subscribe and read along.

Green Software Foundation

AI can answer faster.
But can it ask better questions?

That may be one of the biggest shifts in the age of AI.

For years, expertise was measured by how quickly answers could be produced. Today, answers are abundant. AI can generate them in seconds.

The differentiator is changing.

The future may belong to people who can:

• Ask better questions
• Connect ideas across domains
• Think critically beyond the obvious
• Apply judgment when certainty is low
• Bring empathy, context, and responsibility to decisions

Perhaps the real challenge is not learning how to compete with AI.
It is learning how to become more than AI.

A recent reflection from my latest book, Be More Than AI: The Reason AI Isn’t Enough:
AI can amplify intelligence.
But meaning, judgment, empathy, and purpose remain deeply human.

In a world where AI becomes increasingly capable, the question shifts from “What can AI do?” to:
What makes humans indispensable?

More thoughts, ideas, and episodes from the book:
📘 Buy the book - https://lnkd.in/dKuU_2Sr
🌐 https://bemorethan.ai/

What is one human capability you believe becomes more valuable as AI advances?

Every token your AI generates has a price tag. Three, actually.

There is the bill — from your model provider, your hyperscaler, or your own inference infrastructure. There is the energy drawn from the grid to produce it. And there is the carbon released into the atmosphere because of it.

For a long time, the assumption was that engineering teams had at least the first one under control. The age of agents has dismantled that assumption.

In the latest edition of Technology Bytes, I unpack why token economics is breaking — and why the answer is not a smaller AI bill, but a different way of thinking about AI altogether.

The next decade of AI will not be won by whoever spends the most on compute. It will be won by whoever wastes the least.

Read the full edition 👇
#GreenSoftware #SustainableAI #TokenEconomics #TechnologyLeadership #AI Green Software Foundation

Last week was my Day 2 at the The Linux Foundation — and I was already at the All Hands meeting.

Our CEO, Jim Zemlin opened the event by reflecting on the culture and values that continue to shape the Linux Foundation and its communities, and one slide in particular stayed with me:

Helpful. Hopeful. Humble.

Those words felt real because they were reflected in the people around me and in the spirit of the open communities we build. It’s about helping others move forward, staying hopeful about our collective future, and remaining humble enough to know that real impact comes through collective effort and shared purpose.

One of those people is Sean Mcilroy. He has been the backbone of the Green Software Foundation projects and operations since the early days. He didn’t wait for me to “onboard” — he immediately started connecting dots, making introductions, and sharing context. That is what community actually looks like.

Swapping Mumbai’s humidity for snow while running on jet lag made the week memorable, but the shared spirit of the community is what grounded me. Having been part of the GSF journey since its inception, seeing the broader Linux Foundation ecosystem up close reinforced why this work matters. Sharing a picture with Sean from the trip.

At the Green Software Foundation, we’re tackling a challenge no single organization can solve alone: building sustainable software and AI systems. As AI adoption accelerates and the token economy scales, the need for standards, measurement, tooling, reduction strategies, and collective responsibility becomes even more important.

But weeks like this remind me that the people here aren’t waiting for permission or applause. They’re just building.

Grateful to be here.

After five impactful years, closing a defining chapter at Accenture.

Five years ago, tech sustainability was still an emerging space. What began as a set of ideas evolved into building awareness, creating assets, platforms, and frameworks, driving real client impact, and contributing to new ISO standards that are shaping how organizations approach sustainability in software and AI.

Over time, this work gained industry recognition, with leadership reflected in evaluations by Forrester and Everest Group.

Grateful to Accenture for the opportunities, and to the Tech Sustainability Innovation team and extended teams who helped build and scale this work.

I’m excited to begin a new chapter with the Green Software Foundation and The Linux Foundation, stepping in as Executive Director.

Having been involved with the Green Software Foundation since its inception, this transition feels both natural and meaningful.

There is still a long way to go—and much more to build together as a community.

Sustainability is a shared responsibility. It requires collaboration across organizations, industries, and communities to make it the default, not the exception, with a clear focus on balancing cost, carbon, and energy in how software and AI are built and operated.

Thankful to Asim Hussain for building a strong foundation, and looking forward to working with the ecosystem to take this forward at scale.

Onward! 🚀

Gadhu Sundaram Jonathan Turnbull Sanjay Podder Carolin Rubner Vinjosh Varghese Yusuke Kobayashi Todd Moore Sean Mcilroy Jamie A Cowan Russell Trow Pindy Bhullar Gosia Fricze Joseph Cook Kirsty Du Plessis Jenya Saprykina Jabari George

AI can process.
AI can generate.
AI can calculate.

But can it think?
Not even close.

Because thinking isn’t one thing. It has layers.
Ask — before you solve, question whether it’s the right problem.
See — connect patterns from your experience, not just data.
Feel — sense what the situation actually needs.
Decide — make the call when there’s no clear answer.
Own — stand behind your thinking.

I call it The Human Stack.

AI operates on the surface.
The depth? That’s entirely yours.

Episode#3 of Be More Than AI is live.

📖 And if you want the full deep dive — I’ve written a book on this. Link in comments.

Children are already growing up with AI.
The real question is — are we teaching them how to use it responsibly?

Little AI Explorer is now live on the iOS App Store:
https://lnkd.in/d24DexEx

This has been a journey of building something simple, yet meaningful — a space where curiosity meets AI in a way that feels natural, safe, and engaging.

Why Little AI Explorer?
AI is becoming part of everyday life.
But one critical piece is often missing — ethics and responsibility.

If children are exposed to AI early, they shouldn’t just learn what it can do. They should understand:

• How to question what AI tells them
• Why not everything generated is correct
• What responsible usage looks like
• Where human judgment still matters

Little AI Explorer is built as an app for kids — designed to:

• Cultivate early curiosity about AI
• Introduce responsible thinking from day one
• Build awareness of ethical AI usage
• Provide a safe space to explore and learn

Because the goal isn’t just early exposure.

It’s building a generation that grows up with AI — thoughtfully, ethically and responsibly.

For more details, visit - https://lnkd.in/dN5E8jta

This is just the beginning.
Would love your thoughts and feedback 🙌

What does Google’s Gemma 4 release actually change for Sustainable AI?

Most conversations will focus on benchmarks.
That’s not where the impact is.
The shift is architectural.

Google DeepMind released Gemma 4 with four open models:
→ E2B and E4B (effective models)
→ 26B MoE (4B active at inference)
→ 31B dense

The smaller models are where this becomes practical.

This is less about model size, and more about where inference happens. Gemma’s “effective parameter” design means:

• Only part of the model is active per request
• Lower memory footprint
• Lower compute per inference

This makes on-device inference viable for a broader set of use cases.

On-device inference improves sustainability under specific conditions:

• High-frequency tasks (e.g., typing assist, voice input)
• Repeated interactions where network calls are avoided
• Models that fit efficiently within device constraints

There are also trade-offs:

• Edge devices can be less efficient per unit of compute than optimized data centres
• Battery consumption shifts energy usage to the device
• Hardware lifecycle impact remains part of the equation

The practical opportunity is in hybrid AI architectures:

1. Edge-first inference (E2B / E4B)
   Handle frequent, low-complexity tasks locally
2. Cloud escalation (larger models like 31B)
   Route only complex queries to higher-capacity models
3. Selective compute (MoE / effective models)
   Activate only the required subset of the model during inference
4. Context-aware routing
   Decide dynamically between edge and cloud based on latency, cost, and energy

Sustainability outcomes are driven by inference strategy, not just model efficiency. The impact comes from:

• Reducing unnecessary large-model calls
• Keeping repetitive workloads closer to the user
• Designing systems that avoid excess compute

Gemma 4 expands the design space.
The sustainability outcome depends on how it is used.

For Gemma 4 details , visit https://lnkd.in/dp7HFbpB

#Gemma4 #SustainableAI #EdgeAI #GreenSoftware #OnDeviceAI #AIEfficiency Green Software Foundation #google

Telling people AI will take their jobs is easy.

It gets attention.
It creates urgency.
It spreads fast.

Because fear sells.
But it also reduces a complex shift into a one-sided story—driven by fear and shallow measures of productivity.

AI is changing work. That part is real.
But it’s not just removing work.
It’s changing what the work actually is.

The part where you write the report? AI can do that.
The part where you analyze the data? AI can do that.
The part where you build the presentation? AI can do that.
The part where you write the code? AI can do that.

But deciding what the report should say,
questioning what the data is missing,
choosing what actually matters,
designing the system, the responsible architecture, the trade-offs—
that part is still yours.

The autopilot work is going away.

What’s left isn’t just “skills.
Thinking.
Seeing.
Feeling.
Deciding.
Owning.
Not tasks. Not outputs.

The things that were always yours.

So the real question isn’t:
Will AI take my job?
It’s:
What part of my work actually requires me?

I wrote this book because something didn’t sit right.
We’ve been measuring human value through output and execution.
AI didn’t remove that value.
It exposed how little we understood it.

This book maps what remains—what humans uniquely bring, why it’s irreplaceable, and how to use AI to your advantage.

I call it The Human Stack:

• Ask More — question what’s worth solving
• See More — expand beyond what’s visible
• Feel More — understand what’s not said
• Decide More — act under uncertainty
• Own More — take responsibility

📘 Be More Than AI
Be the Reason AI Isn’t Enough
Now available on Amazon - https://amzn.to/4v9iiBU

AI will keep evolving.
The Human Stack is not what it replaces. It’s what it reveals.
Own it.

AI doesn’t remove work. It reveals what the real work was all along.

Take something simple — email.
In many roles, a good part of the day goes into:
Reading threads.
Understanding context.
Figuring out what’s needed.
Drafting responses.

Now AI can do most of that in seconds.
And when it does… something interesting happens.
👉 The effort disappears.
👉 But the outcome doesn’t.

Which raises a deeper question:
If AI handles more of the doing…
what does “work” actually become?

🎥 Episode 2 of “Be More Than AI” explores this shift.
Because as tools remove friction,
what remains becomes much clearer.