There’s a bill that doesn’t show up in your AI budget spreadsheet. It doesn’t appear in the vendor proposal or the POC sign-off. But it’s real, it’s growing, and increasingly, your enterprise customers are starting to ask about it.
The bill is energy. And the conversation around it has quietly moved from sustainability reports to strategy rooms.
When Scaling Stops Being Free
There’s a useful thought experiment here. Imagine every Google search suddenly ran on a large language model. The energy required would dwarf what the current internet infrastructure consumes — by some estimates, a single AI query already uses roughly ten times more electricity than a traditional search.
Scale that by a billion queries a day, and the arithmetic gets uncomfortable fast.
Training large models is only part of the story. GPT-3’s training run consumed over 1,000 megawatt-hours — enough to power hundreds of homes for a year. Models since then have grown substantially larger. But training is a one-time event. Inference — the act of running the model at scale, day after day — is where the cumulative footprint really compounds. Research now suggests inference accounts for more than half of an AI system’s total lifecycle carbon emissions.
This isn’t a distant problem. It’s already on the agenda.
The Green AI Shift — From Footnote to Framework
For a while, sustainability in AI sat comfortably in the corporate responsibility section of annual reports. A few paragraphs. Some renewable energy credits. A net-zero pledge with a reassuringly distant deadline.
That’s changed.
What’s driving it is less idealism and more arithmetic. Enterprise customers are carrying their own ESG commitments — and they’re realising that every AI workload they run through a vendor’s data centre lands in their Scope 3 emissions. The environmental cost of AI is no longer the vendor’s problem. It’s shared.
The response from the largest players has been notable. Microsoft committed to sourcing over 10 gigawatts of renewable energy through partnerships by 2030 — a move that, for context, is roughly the output of ten nuclear power plants. Google struck partnerships to co-locate data centre capacity directly alongside new clean energy generation. Meta’s renewable energy projects added more than 15 gigawatts of clean capacity to global grids as of last year.
Nuclear, once the industry’s awkward uncle, is back at the table. Microsoft signed a power agreement tied to the restarted Three Mile Island plant. Amazon acquired a nuclear-powered data centre campus in Pennsylvania. The pattern worth noting: energy is being treated as infrastructure, not overhead.
What Efficiency Actually Looks Like in Practice
Beyond energy sourcing, the more immediately actionable shift is happening at the model level.
Model compression techniques — distillation, quantisation, sparse architectures — are delivering real gains. Research suggests quantisation alone can reduce energy per inference by up to 50%. Knowledge distillation approaches like DistilBERT achieve roughly 40% fewer parameters while retaining almost all of the performance. This is the lens worth applying: efficiency isn’t about capability tradeoffs anymore. It’s increasingly a design principle.
Hardware selection matters more than it once did. The gap between a commodity GPU and a purpose-built AI accelerator isn’t just about speed — it’s about energy per useful computation. Google’s TPUs, for instance, operate at significantly higher energy efficiency than equivalent CPU workloads for the same tasks.
There’s also the question of where and when you run workloads — a consideration that barely existed five years ago. Some enterprises are beginning to time workloads around renewable energy availability on the grid. Others are choosing data centre regions based partly on local energy mix. These aren’t radical green gestures; they’re practical cost and risk management.
The Competitive Signal Most Founders Are Missing
Here’s the thing that’s easy to miss if you’re building AI infrastructure right now: energy efficiency is quietly becoming a procurement criterion.
Enterprise buyers with serious sustainability commitments are starting to ask about the carbon cost of the AI services they’re adopting. The question won’t always appear on the initial RFP. But it’s showing up in the conversations that happen before and after the formal process.
The pattern emerging from early movers suggests that founders who can speak to environmental efficiency — not just model performance and latency — are landing in a different part of the conversation. It’s the same dynamic that played out with data security a decade ago. For a while, it was a nice-to-have. Then it became the first question asked.
Green AI isn’t about slowing down AI adoption. It’s the realisation that scaling without accounting for energy is building a fragility into the strategy — one that shows up later, in ways that are harder to fix.
The question isn’t whether to use AI. The interesting question is whether the way you’re using it is something you’d want to put in a customer briefing.
As agentic AI moves deeper into enterprise operations — a thread explored in the recent post on Agentic AI in Enterprise: The First Real Tests Begin — the energy question only gets more relevant. Autonomous systems running continuously consume differently than tools invoked on demand.
What’s your read — is energy efficiency already a real differentiator in the AI conversations you’re having, or does it still feel like a future problem?
Let’s keep learning — together.
harish.balasubramanian 
Share your thoughts