Goldman Sachs raised its year-end 2026 US data center capacity forecast by 5 gigawatts to 64 GW on Sunday, while trimming its 2027 forecast by the same amount to 90 GW. The adjustment came in a note that argued the rapid expansion of AI computing infrastructure is holding up through 2027 despite mounting political and community opposition.
The move is small in absolute terms but it signals something larger. Goldman said project data and updated capacity estimates show the US data center growth outlook through 2027 remains largely unchanged. The firm expects US data center power demand to rise 38 percent, or 12 GW, in 2026 and another 38 percent, or 17 GW, in 2027.
The timing matters because the buildout is colliding with local politics. Data centers that train and run AI models have drawn opposition in communities nationwide over electricity demand, utility costs and other local impacts. Only a third of Americans approve of the pace of data-center construction, according to a June Reuters/Ipsos poll, and just 14 percent would support a data center being built in their own community to support AI projects for firms such as Meta, Alphabet, Amazon, Microsoft and Elon Musk's xAI.
Goldman's read is that the backlash has limited near-term impact on expansion. "While US data center development faces increasing political and community opposition, project data and our updated capacity estimates suggest that the US data center growth outlook through 2027 remains largely unchanged," the note said. The firm added that public sentiment has turned more negative.
The political stakes are immediate. The backlash poses a challenge for Republicans heading into the US midterm elections on November 3, and the construction boom has become a ballot-box issue in several states. Goldman's forecast implies that even where projects face delays, the aggregate pipeline is large enough to absorb slippage without changing the national trajectory through 2027.
Why the edge is where the friction lands
The capacity story is not only about hyperscale campuses in rural counties. As AI moves from training to inference, the workloads that serve users in real time need to sit closer to the people and devices generating requests. That is the edge, and it is where the physical and political constraints tighten fastest.
Edge data centers are smaller and more distributed than the flagship campuses, which makes them easier to site in some cases and harder in others. They draw less total power but more of it per square foot, and they often land inside city limits where grid connections and permits are scarcer. The Accenture Soben Data Centre Trends Report 2026 projects the edge computing market will grow 32.2 percent annually from 2025 to reach $2.2 billion by 2030, driven by the spread of 5G and the need to run AI models locally.
For operators, the choice is becoming a tradeoff between latency and land. A model served from a hyperscale region adds round-trip delay that can break voice assistants, autonomous systems and industrial controls. Moving inference to an edge node cuts that delay but multiplies the number of sites that need power, cooling and physical security. Each new site is another potential planning dispute.
Power is the binding constraint
The 12 GW and 17 GW power demand figures are the real story behind the capacity numbers. A gigawatt of data center capacity is not just servers; it is grid interconnection, substations and, increasingly, on-site generation. Goldman's forecast assumes that power can be secured and delivered on schedule, which is the assumption most at risk from the community pushback.
The Europe picture shows the same constraint from a different angle. JLL estimates the world's four largest hyperscale cloud providers will spend $725 billion in 2026, up 77 percent from $410 billion in 2025, mostly on data center infrastructure. European AI data centers are already seeking cheaper and quicker access to energy and land, according to Reuters reporting from August. The pattern is global even if the politics are local.
What changes if the backlash hardens
Goldman's 2027 trim of 5 GW is a hedge, not a reversal. It says the firm expects some projects to slip but not to disappear. The more interesting question is whether the slippage concentrates at the edge, where sites are smaller and more numerous, or at the hyperscale tier, where a single delayed campus can erase a gigawatt of planned capacity.
It is not clear this will hold. If community opposition translates into moratoriums or zoning changes rather than just slower permitting, the 2027 number could move again. The June poll suggests the public has not been persuaded that the economic benefits of AI infrastructure outweigh the local costs, and that gap is a real risk to the schedule Goldman is forecasting.
For cloud and edge operators, the practical implication is to plan for longer timelines and to treat power as a first-order design input rather than a utility hookup. The capacity will probably arrive, but the location and the cost may not match the original plan.
The next data point to watch is whether the November 3 midterms produce state or local measures that restrict data center siting. Goldman's forecast assumes they do not. If they do, the 64 GW figure for 2026 is already locked in, but the 90 GW target for 2027 will need another revision.
There is also a quieter shift inside the industry. Operators are starting to treat cooling and water use as design constraints rather than afterthoughts, and liquid cooling is moving from niche to standard in new builds. These changes do not show up in capacity forecasts, but they affect whether a project can actually get permitted and built on the timeline Goldman assumes.
The community opposition is not uniform either. Some regions have welcomed data centers for the tax base and construction jobs, while others have organized against them over noise, traffic and strain on local water supplies. That variation means a national forecast can look stable even as individual projects face very different odds. Goldman's aggregate numbers hide that dispersion, which is worth keeping in mind when reading the headline.
For readers tracking this sector, the practical takeaway is simpler than the forecast. AI demand is large enough to keep the buildout going through 2027, but the physical and political limits are tightening at the same time. The next two years will show whether the industry can deliver capacity where it is needed, at a cost that customers will pay, and without losing the communities that host it.
