Just about every company I speak with is trying to move faster on AI. At the same time, most are working to reduce emissions, manage rising energy demands, and make credible progress against their climate goals.
Too often, these priorities are treated as competing forces: scale AI now, come back to sustainability later. But that framing is too limited—and too costly. With AI’s massive energy demands, there’s no way to have a credible AI strategy without a sustainability strategy.
Thankfully, we’re already starting to see a strong focus on curbing AI’s carbon impact through hardware efficiency improvements. Better chips, smarter cooling, lower power draw per unit of compute; these gains are real and necessary.
But we also need to ask a more basic infrastructure question: Are we running AI workloads in the right places to begin with?
New Evidence, New Options
The carbon impact of AI’s energy demands can vary significantly depending on where the work happens. Running the same workload in one region versus another can result in very different emissions based on the local energy mix and grid efficiency.
Until recently, companies trying to reduce AI’s carbon footprint had limited options at the infrastructure level. They could choose vendors with stronger climate records, improve efficiency within existing sites, or buy offsets. Routing workloads to where the cleanest energy already exists, without a real performance trade-off, was not viable.
Thankfully, that constraint is starting to ease.
In early 2026, NTT completed a proof of concept recognized by the IOWN Global Forum. The team set out to test whether AI training could be moved from local infrastructure to remote sites powered by renewable energy sources without sacrificing performance. The answer was a resounding yes.
The project paired low-latency photonic networking with storage tools designed for fast data movement at long range. In some cases, it cut energy use by up to 30 percent by routing compute to cleaner power regions. Importantly, training speed dropped by less than one percent versus a local setup.
This proof of concept also shows how much the data layer shapes energy outcomes. Sustainability teams often focus on network or hardware choices. But maintaining training performance spanning 3,000 kilometers required more than a fast network. It required storage tools built to sustain data throughput at that range. The systems managing data are a critical part of the energy equation.
Lower-Carbon AI Doesn’t Have to Mean Slower AI
The practical value here is simple: Companies do not have to slow AI programs to make progress on their climate goals. If GPU resources are available in a region with clean power and an efficient grid, they can lower emissions by shifting where workloads run—not by asking AI teams to compromise on performance.
This is work that IT and sustainability teams can own together. And it’s becoming increasingly urgent.
Policymakers are watching data center energy demand closely, especially where it strains national grids. Investors press companies to explain how they plan to scale AI while meeting climate targets. Companies treating AI’s carbon impact as a compliance issue, rather than a design question, are finding that gap increasingly visible and costly.
Closing that gap requires real infrastructure decisions, not just better reporting. Routing AI workloads toward cleaner energy through remote, high-performance systems is one such decision that’s now backed by evidence, rather than purely theoretical projection.
Performance and Responsibility as Reinforcing Goals
For years, the gap between where data lives and where compute runs has been treated as a fixed cost. The NTT proof of concept suggests it can be a design choice instead.
Sustainability and IT leaders can use that choice together to build AI programs that scale without widening the energy gap they will need to close.
The goal must be to design systems where performance and responsibility reinforce each other, and where the data platforms enabling AI also support a more efficient and resilient operation.
