Every few months, a new headline declares the "real" constraint on AI infrastructure growth – ranging from power, land, or community opposition. Each framing is true, and each is incomplete.

Having recently conducted new research on global data center demand through 2030, I've come to a different conclusion: the industry isn't facing a single bottleneck at all. It's facing a convergence of them, arriving at the same time, in the same places.

That distinction matters, because it changes what "solving" the problem actually requires.

Our modeling looked at three growth paths for data center capacity through 2030 – steady, accelerated, and slower – and stress-tested each against the resources that make AI infrastructure real: electricity, land, water, capital, processors, critical minerals, and skilled labor.

We believe the industry's most likely trajectory sits between the steady and accelerated cases, meaning global installed capacity could grow somewhere between 23 and 30 percent a year for the rest of the decade. That's an extraordinary rate of expansion for any physical infrastructure sector, let alone one this capital- and resource-intensive.

Under that trajectory, the numbers get large fast. In the US alone, incremental data center electricity demand could account for roughly half of all planned generation capacity additions between now and 2030.

Transformer lead times already stretch past 100 weeks. We estimate global GPU demand could outstrip production by tens of millions of units under accelerated growth. And land – despite America's apparent abundance of it – is becoming scarce in exactly the metro corridors where power and fiber already exist, a scarcity now compounded by a sharp rise in community opposition: our sentiment analysis across seven countries found public attitudes toward data centers, positive as recently as 2024, turned negative in the US, the Netherlands, and Canada by 2026.

None of these pressures, on their own, would stop the industry. Chip shortages alone are manageable. Transformer lead times alone are manageable. Even community pushback, project by project, is manageable. What our stress test shows is that these constraints are tightening simultaneously, across the same regions, on the same timeline.

Operators will keep winning sites and securing megawatts faster than the industry can line up the transformers, chips, water rights, and workforce needed to turn those sites into usable capacity. That's not a supply problem in any one category. It's a coordination problem across all of them.

I'd also push back on one narrative that's crept into the conversation: that this is fundamentally a water story, or fundamentally a "data centers versus residential ratepayers" story. It isn't. Water usage is actually one of the stronger findings in our research – newer facilities produce zero water waste, and continued cooling innovations are helping reduce water impact even further.

And on power, the more sophisticated operators are increasingly co-investing in grid infrastructure and absorbing their own load growth through demand-response commitments, rather than quietly passing costs to other customers. The industry has real, defensible answers on both fronts. It has done a poor job telling that story, which is part of why sentiment has soured even as the underlying practices have improved.

Where I think the industry is genuinely exposed is in treating growth as a series of individual land grabs rather than a system to be engineered. Securing a site and a power interconnection are separate races right now, often run by different teams, sometimes at different speeds. That must change.

Operators need to co-plan power infrastructure with utilities years in advance rather than showing up with a load request. They need supply-chain resilience – diversified vendors, standardized designs, long-term offtake agreements for transformers and generators – built into project approval from day one, not bolted on after the first shortage hits.

They also need to treat community engagement as core infrastructure planning, not a communications afterthought, because the fastest-growing constraint on this list right now isn't a physical resource at all. It's public patience.

Policymakers have a role here too, and a narrower window than they may realize. Interconnection queues, permitting timelines, and zoning frameworks were built for a world of gradual, predictable load growth. AI has ended that world.

Governments that treat power-system modernization as strategic industrial policy – and that give operators bankable, long-term frameworks for clean power procurement – will capture a disproportionate share of the next wave of investment. Those that don't will watch capital flow to jurisdictions that solved this first.

The AI infrastructure buildout is not going to be undone by any single shortage. But it will be shaped, region by region, by how well operators, utilities, and governments coordinate across power, land, equipment, and community trust at the same time. Treat that as an engineering challenge to be orchestrated, and the industry can sustain this growth. Treat it as a race to be won site by site, and the convergence will win instead.