Goldman Sachs raised its 2026 year-end US data center capacity forecast by 5 gigawatts to 64 GW on Sunday, then reduced its 2027 year-end forecast by 5 GW to 90 GW in the same research note.
Two edits to the same spreadsheet, pointing in opposite directions. The 2026 number went up because projects that were already financed, permitted and under construction are landing faster than the bank expected. The 2027 number went down because projects still waiting on power and local approvals slipped out of the window.
The gap between those two moves is the real story of the US buildout in late 2026. Compute hardware stopped being the hard part. Power, interconnection queues and county zoning hearings now decide when a gigawatt-class campus comes online, and the schedule slips get pushed into whatever year still has room.
The Forecast Moved in Two Directions at Once
The bank's note, reported by Reuters, holds the growth rate steady while reshaping when the growth arrives. Goldman forecasts US data center power demand to grow 38%, or 12 GW, in 2026, and another 38%, or 17 GW, in 2027.
"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, according to Reuters.
"Largely unchanged" is doing a lot of work in that sentence. Total capacity across the two years is close to where it was, but the composition changed. Near-term megawatts got pulled forward. Megawatts that depend on a substation being built got pushed back.
This is what a constrained buildout looks like on paper. When land and servers are easy to line up but grid connections take 24 to 36 months, the projects with permits in hand keep their dates and everything else drifts. Analysts at the bank described the 2027 reduction as timing rather than a verdict on demand, which is a defensible read given that hyperscalers have kept capital expenditure commitments intact.
Still, it's worth being blunt about the numbers. Cutting 5 GW from a single forecast year is not a rounding error. A gigawatt is roughly the output of a large nuclear unit, and the US added about 12 GW of data center demand growth in 2026 under the bank's own estimate. Moving 5 GW means about a third of a year's worth of growth shifted out of 2027.
A Third of Americans Approve. Fourteen Percent Want One Nearby.
Goldman did not flag political risk in the abstract. It cited a June Reuters/Ipsos poll in which just a third of Americans approved of the pace of data center construction in the US, and only 14% of respondents said they would support a data center being built in their community to support AI projects for companies including Meta, Alphabet, Amazon, Microsoft and xAI.
"Public sentiment has also turned more negative," the bank added.
That gap between 33% approving nationally and 14% welcoming a site locally is the operative number for anyone underwriting a 2027 in-service date. National approval is a headline. Local opposition is a permit.
Local opposition is exactly what has been generating headlines all year. Reuters reported in September that Democrats in the rural Midwest are campaigning on data center backlash, framing electricity demand and utility bills as a pocketbook issue. US midterm elections land on November 3, and the bank's note said the backlash poses a political problem for Republicans running in districts where new campuses are visible from the highway.
One honest reading of the forecast revision: Goldman may have moved some 2027 capacity into 2026 precisely because 2026 approvals are locked and 2027 approvals are not. The utility bills and the polling numbers get worse before they get better, and utility cost allocation is heading into a legislative fight rather than a private negotiation. If state regulators and municipal bodies push back harder next year, the 2027 shortfall could widen rather than close.
Finland Is the Other Side of the Same Coin
Applied Digital made a different bet on October 6. The Dallas-based developer said it had signed an agreement for access to up to 1 GW of potential power capacity in Finland, its first development outside the United States. Initial power is expected to be available starting in 2028.
"Finland stood out because many of the characteristics that have supported our success in North Dakota are also present in the Nordics, a favorable climate, an attractive energy ecosystem, strong connectivity and the ability to develop at meaningful scale," Wes Cummins, the company's chairman and CEO, said in the announcement.
Cummins was careful to frame it as a hedge rather than a pivot. "This is a deliberate first step, not a change in our strategy. Our focus remains firmly on executing the projects in front of us in the United States," he said. The company is marketing the site and talking to hyperscale customers now.
Pair the two announcements and the picture is coherent. Capacity is not disappearing. It is moving to wherever power is available on a known date. When Goldman trims a US year, operators do not stop building. They shop for a substation in a jurisdiction that returns the call.
That is the part the market has been slow to price. Applied Digital's Finnish campus will not produce a watt until 2028 at the earliest, and the company structured the deal to limit its initial exposure. It's a real signal that US power scarcity is now severe enough to justify a multinational search, but it's not a 2027 fix for anyone.
Transformers Are the Quiet Bottleneck
Power availability is not one constraint, it's a stack of them. A site can have a signed interconnection agreement and still stall on large power transformers, switchgear, and the turbine and gas plant capacity that utilities use to serve a new 200 MW load. Lead times on some grid equipment have stretched to multiple years.
This is why the split in Goldman's forecast is the most informative datapoint of the week. A forecast that shifts capacity between years, while keeping the two-year total roughly flat, is a forecast describing a logistics problem, not a demand problem. Demand for training and inference capacity did not change between Sunday and the Sunday before. The interconnect and permitting calendar did.
It also explains why liquid cooling, closed-loop designs and siting in cool climates have stopped being differentiators and become table stakes. Cummins pointed to climate and an energy ecosystem, and he is not being sentimental. Every degree of ambient temperature is a cheaper cooling bill, and cooling is where a meaningful share of a facility's power draw goes after the chips themselves.
What to Watch
The next real test of Goldman's read arrives in Q4 earnings from the hyperscalers, when Meta, Alphabet, Amazon and Microsoft update capital spending guidance for 2027. If that guidance holds near current levels, the bank's claim that the buildout is a scheduling problem will hold with it. If guidance gets trimmed, the 5 GW cut looks less like a slip and more like the start of a correction.
Watch the November 3 midterms too. Local data center fights produced campaign issue out of a utility bill, and the outcomes in the Midwest and mid-Atlantic will tell you how much siting risk just got added to 2027 and 2028 timelines.
And watch where the next Finland deal lands. If two or three more developers follow Applied Digital into the Nordics, Canada or Iberia before the end of the year, the constraint has stopped being a US political story and become a global power procurement story. That is the outcome worth planning around, and it's not clear anyone is ready for it.
More on capacity, siting and pricing in our cloud and edge computing coverage, with earlier reports in the archive.