Donald Trump gave the clearest single line of the year on AI policy this weekend, and it wasn’t really about policy at all.
Asked about a rare joint call from Dario Amodei, Sam Altman and Elon Musk to slow the pace of AI development, he dismissed it outright. “Whoever wins AI wins,” Trump told reporters, brushing off warnings he characterised as coming from “negative forces” hyping up “things that won’t happen.”
Amodei’s original essay had argued plainly that “we must slow the pace at which we improve the capabilities of AI models.”
Within a day, the administration’s AI czar David Sacks had gone further, accusing the labs of trying to protect a “duopoly on frontier intelligence” and telling them to “stop pretending you need anyone else’s permission.”
It was a tidy ending to a year that never really had a federal referee to begin with.
Looking back at everything this sector has been through in 2026, from moratoriums to circular financing structures nobody had a name for 18 months ago to a genuinely new geopolitical fault line over sovereign compute, the real story isn’t that Washington declined to slow AI down.
It’s that three other forces quietly stepped into the space federal policy left empty, and between them did the job instead: local communities, capital markets, and increasingly, national governments acting in their own self-interest rather than in concert.
Local boards did what Washington wouldn’t
The clearest version of this played out at the local level. New York’s Responsible Data Center Development Act became the first statewide moratorium of its kind this year, tying permitting directly to community benefit agreements rather than leaving developers to negotiate goodwill informally, as Capacity reported when it passed.
It wasn’t an isolated case. Grassroots opposition groups have more than doubled in number since the end of last year, and Capacity’s coverage of the backlash tracked how little effect dismissive rhetoric from the administration actually had on the pace of local resistance. Community benefit agreements, once a courtesy gesture from developers keen to smooth planning approval, hardened into something closer to a functioning asset class, with lenders reportedly pricing local political sentiment directly into financing terms rather than treating it as background noise.
None of that was ever going to show up in a federal press briefing. But it did more to actually slow individual projects down, forcing site re-selection, adding months to permitting timelines, killing some proposals outright, than anything resembling a national safety framework managed to.
Trump’s weekend comments confirmed what had been obvious for months: if there was ever going to be a brake on this build-out, it was not going to come from the top. It was going to come from a planning committee in a county most of the industry had never heard of 18 months earlier, arguing about water tables and substation capacity while Washington argued about who wins the AI race.
Capital markets asked the questions nobody in Washington did
The second force did its work more quietly, in bond prospectuses and term sheets rather than planning hearings.
Meta’s attempt to raise more than $12 billion through BlackRock for its El Paso campus priced at yields north of 7%, a noticeably harder ask than a structurally similar deal it had closed less than a year earlier.
UBS strategist Matthew Mish, quoted in Capacity’s reporting on the repricing, described the pace of AI-related debt accumulation as a trend that “would make anyone familiar with credit cycles raise an eyebrow.” Investors weren’t rejecting AI infrastructure outright.
They were starting to ask, project by project, whether the underlying economics actually justified the leverage being stacked on top of them, a scrutiny that had been almost entirely absent 12 months earlier.
That same scrutiny reshaped how deals got structured rather than just how they got priced. The $35 billion Apollo and Blackstone facility behind Anthropic’s compute build-out, covered by Capacity when it closed, showed private credit stepping in where a lab’s own balance sheet couldn’t carry the weight alone.
Nvidia went further still. Its arrangement with Lambda and Hut 8 in Texas, where Nvidia is reported to hold the actual lease on the data centre while Lambda resells compute to Anthropic, turned the chipmaker into something closer to a landlord than a supplier. Lambda’s chief executive Michel Combes summed up the logic behind this wave of financing bluntly when a separate facility closed at the end of August: “Closing this Facility puts capital straight to work.”
Capacity’s explainer on the pattern traced how circular financing, in which a chipmaker’s investment and its customer’s purchase commitments feed each other in a loop, went from a niche criticism to a mainstream feature of how this sector now gets built.
What both of these forces share is that neither one waited for permission from a regulator. Local boards didn’t need federal guidance to say no to a project that strained the grid or the water table.
Bond investors didn’t need a policy framework to decide a yield was too thin for the risk being asked of them.
Between the two of them, 2026 produced something that looked, from a distance, almost like governance: pace controlled by cost of capital on one side, and by community consent on the other.
Geopolitics added a third: sovereignty as regulation
The year’s least tidy addition to this picture came from national governments, and it arrived with far less consensus than either of the other two.
At the G7 summit in Évian in June, Sam Altman, Dario Amodei and Google DeepMind’s Demis Hassabis joined heads of state for a working session on AI, and the message from the people building the technology was unusually blunt. “The technology’s future must be shaped by people, democratic institutions,” Altman told the gathering, as Capacity reported from the summit.
The timing mattered: the appeal came five days after the Trump administration had forced Anthropic’s latest models offline to comply with export controls, a move Macron publicly welcomed in principle while criticising its execution as “strictly nationalist.”
That tension, between wanting frontier AI governed collectively and every major bloc quietly building its own version of control, defined the rest of the year. The European Commission’s Tech Sovereignty Package, unveiled days after the G7 summit, set a target to triple the EU’s data centre capacity within five to seven years, backed by an estimated €200 billion in largely private investment, and paired that ambition with a Cloud and AI Development Act designed to reduce reliance on a small number of non-EU providers. Canada and Germany formed their own Sovereign Tech Alliance around the same time.
None of this was co-ordinated with Washington, and much of it was explicitly a response to Washington’s own posture, most obviously the executive order promoting a global “American AI technology stack” while declining to impose mandatory rules on the US companies building it.
That fragmentation is arguably the most consequential of the year’s three regulatory substitutes, because it doesn’t just add friction, it adds contradiction.
A procurement decision that satisfies EU sovereignty requirements may sit awkwardly against a US export control regime built around the opposite assumption, that American infrastructure should be the default choice globally. Sovereign compute became one of 2026’s defining phrases precisely because governments found it easier to build their own capacity than to agree on shared rules for anyone else’s.
What that leaves for 2027
The obvious risk in all three of these arrangements is that none of them is actually accountable to anything beyond its own interests.
A local moratorium can be exactly right for the community it protects and entirely wrong for the country’s broader AI competitiveness, and there’s no mechanism reconciling the two. A bond market repricing risk protects investors, not the communities living next to the campus the debt is funding. And a sovereignty package protects a bloc’s own strategic position, without much regard for whether the resulting patchwork of infrastructure rules makes the technology safer or simply harder to build coherently anywhere.
Trump’s comments this weekend all but guarantee that federal policy stays out of that gap through the year ahead, which means the tension between local resistance, capital appetite and sovereign ambition is going to keep resolving itself deal by deal, county by county, and jurisdiction by jurisdiction, without anyone designing the outcome on purpose.
That’s not a comfortable note to end a year of coverage on, but it’s an honest one. The infrastructure underneath the AI industry didn’t get regulated in 2026. It got priced, it got zoned, and it got quietly nationalised in three different directions at once, and those forces ended up doing more of the actual governing than anyone in Washington did.
RELATED STORIES
- Nvidia now agrees to rent back unused GPU capacity from neocloud operators if customer demand falls short
- CoreWeave’s debt hits $35bn: What it means for the neocloud refinancing wall
- What is circular financing in AI infrastructure, and should telecoms and data centre operators be worried?
- What is a neocloud, and why does it matter for digital infrastructure?
Metro Connect USA 2027
Metro Connect USA is the largest executive-level digital infrastructure event in the U.S. The only one of its kind, this 25-year-strong gathering is where decision makers come together to make deals happen.
