The discussion about the limits of AI is almost always about models: how smart is it, what can't it do yet, when does the next version arrive.
By Anthony Raaijmakers
The discussion about the limits of AI is almost always about models: how smart is it, what can't it do yet, when does the next version arrive. But there's a limit that rarely comes up in that discussion, and one that in practice may hit harder: the physical infrastructure that keeps AI running. That gets stuck on copper, steel, and a power grid that has to handle it.
Sightline Climate, a research organization tracking the US data center market, published a report this spring on the gap between planned and actual data capacity. According to that research, of the 16 gigawatts of capacity meant to go live this year, only about 5 gigawatts was actually under construction, and 2027 doesn't look better. The explanation for that gap isn't a lack of money or chips, but one layer deeper: high-voltage transformers have lead times approaching four years, as pv-magazine reported earlier this year based on US transformer market data. Switchgear installations that go with large data centers can't be reordered quickly either.
"AI scales infinitely" is a claim you hear a lot, and technically it's true — given enough compute. But that compute hangs off a physical chain that can't be sped up with a software update. Grid operators are dealing with infrastructure that's already heavily loaded. Transformer manufacturers have backlogs stretching years out. Major AI labs and hyperscalers are already feeling this: new data centers get delayed or built in locations where power is actually available, regardless of whether that's ideal for latency or regulation.
For an SMB, this changes the relevant question. Not "how do I get more compute," but "what am I using the compute I have for?" If capacity is scarce and costs stay high, it matters more whether the application you're building actually delivers something: a process that runs faster, an error that no longer slips through, a decision that's better supported. The next win isn't in more compute — it's in better choices about what you use it for.
This applies doubly to companies that don't build data centers themselves. You don't set the pace of grid expansion. But you do decide which problem you tackle first, and whether investing in AI fits what your organization actually needs right now.
We start every AI project with the question that rarely gets asked in the capacity debate: what does this need to concretely deliver for your organization? Not as a brake on ambition, but because most of the gain sits in targeted applications that solve a recognizable problem, not in broad experimentation with expensive compute time.
The physical limits of AI are real and will become tangible over the coming years. The choice of how you respond — which problem you tackle first and what you build with it — is yours.
Want to make that call before you invest, rather than after? You can give us a call.
Source: Sightline Climate — Data Center Outlook, published May 2026 | pv-magazine USA — transformer lead times, published 2026-05-11