Goldman Sachs put a number on it. By 2027, U.S. data centers will need 66 gigawatts of power. That is more than double the 31 gigawatts they consumed in 2025. The driver is AI. The math is simple: more compute, more electricity. The question is whether the industry can square that curve with its own sustainability pledges.
Right now, many of those pledges look like paper promises. Organizations sign commitments to cut emissions. Then they deploy large language models and image generators that burn through kilowatts. The report notes that these sustainability goals often do not align with actual AI usage. That is a polite way of saying the two things are in conflict.
The scale of the conflict is growing. AI workloads already account for 14 percent of global data center demand. By the time the infrastructure buildout hits its next phase, that share is projected to reach 27 percent. Nearly a third of all data center energy will be going to AI. And this is before the next generation of models arrives. The trend line points up.
Edge computing is being discussed as a possible answer. The idea is to process data closer to where it is generated — on a device, in a local server, at the network edge — rather than shipping everything to a giant data center. That could cut the energy cost of transmission. It could also reduce the need for centralized cooling and massive power feeds. But edge computing has limits. Not every AI task can run on a small chip in a factory or a phone. Training still demands clusters of GPUs in a warehouse. The split between training and inference matters.
Inference — the part where a model actually answers a query or makes a prediction — is where edge could make a real dent. If a self-driving car runs its object detection on an onboard computer instead of calling a cloud server, that saves round-trip latency and grid power. If a factory runs quality checks on a local box, the data center load drops. The report does not say how much of the projected 66 gigawatts could be shaved off this way. But the logic is straightforward.
The counterforce is the sheer momentum of the buildout. Goldman Sachs projects the doubling of power demand in just two years. That is not a gentle ramp. That is a construction boom. Companies are not waiting for edge solutions to mature. They are buying land, ordering transformers, and locking in power purchase agreements. The infrastructure is being built for centralized scale. Retrofitting that for edge distribution later would be expensive and slow.
There is also the question of what happens when sustainability commitments collide with quarterly earnings. AI is the revenue story right now. Cloud providers are selling access to GPUs at high margins. The companies using those GPUs are racing to get products out. Cutting energy use is a cost-saving measure in theory. But in practice, it often takes a back seat to speed and capability. A more efficient model that takes longer to train loses the race.
The report frames this as a need for sustainable solutions. That is accurate. But the need and the incentive are not always aligned. The data center power demand is projected to more than double. AI workloads will drive that growth. Unless the industry finds a way to decouple AI performance from energy consumption, the two curves will keep rising together.
Edge computing is one lever. Better chip design is another. So is renewable energy procurement for data centers. None of them is a silver bullet. The report does not predict which approach will dominate. It simply states the numbers and the tension. 31 gigawatts to 66 gigawatts. 14 percent to 27 percent. The gap between what companies promise and what they power.
