Crédito da imagem: Image via source article. Used under fair use for news commentary. · source
You know that feeling when you order something online and the delivery estimate keeps slipping? First it's two days, then a week, then "we'll let you know." That's basically what's happening with AI infrastructure right now, except instead of a package, it's the entire foundation the industry needs to function.
I've been thinking about this a lot lately, honestly. We spend so much time covering the flashy stuff (new models, humanoid demos, funding rounds) that we sometimes miss the deeply unsexy problems that could actually slow everything down. And right now, the unsexy problem is power.
CyrusOne CEO Eric Schwartz laid it out pretty clearly in a recent Bloomberg interview: the AI data center expansion has become what he called an "industrial arms race." And like most arms races, it's running into supply constraints nobody planned for.
The bottleneck isn't chips. It's not even money, tbh. It's the physical infrastructure: power grids, substations, and the skilled workers needed to build them. We're talking about trillion-dollar infrastructure bets that depend on... having enough electricians and lineworkers. Which we don't.
I initially thought this was a temporary hiccup, something that would sort itself out as wages rose and workers moved into the sector. But after reading more about the scale of what's being planned, I'm less sure. The gap between what AI companies want to build and what the grid can actually support is enormous. And closing that gap takes years, not months.
Cobertura relacionada
More in Industrial
The acquisition signals Autodesk's push beyond CAD software into the messy reality of keeping physical assets running, though whether this creates genuine synergies or just a larger software bundle remains to be seen.
Aisha Patel · 6 hours ago · 8 min
More than you'd think, actually. Musk's IPO filing has some interesting implications for industrial automation.
Robert "Bob" Macintosh · 8 hours ago · 3 min
The global rush toward generative AI is pulling venture dollars away from emerging markets, and African robotics companies are feeling the pinch.
Aisha Patel · 14 hours ago · 6 min
Two days of demos, talks, and networking won't answer the hard questions about where this industry is actually headed.
Here's where it gets interesting. You might be wondering: okay, infrastructure is slow, but surely the software companies are ready to capitalize once it's built?
Not exactly. Gil Luria, an analyst at D.A. Davidson Technology Research, made a point on Bloomberg that stuck with me. He was reacting to Salesforce's latest earnings, and his take was basically: the shift to AI is taking longer than expected. Not failing, just... slower.
This is a company with massive resources, a clear AI strategy, and customers who theoretically want this stuff. And even they're finding the transition harder than anticipated. The reasons are probably mundane (integration challenges, customer readiness, the usual enterprise software friction) but the pattern matters.
The supply side can't keep up. The demand side isn't ready. That's a weird place for an industry that's supposedly moving at unprecedented speed.
Okay, so why am I, a humanoids reporter, writing about data centers and enterprise software? Because the implications ripple outward.
Every advanced humanoid, every embodied AI system, every robot that needs to process complex sensory data in real-time depends on this infrastructure. The compute has to come from somewhere. And if the power isn't there, or the data centers aren't built, or the models aren't deployed at scale, then all those impressive demos we keep seeing stay demos longer than anyone planned.
I should know this better, but I don't have great data on how much compute current humanoid systems actually require for training versus inference. What I do know is that the companies building these systems are making massive assumptions about cloud infrastructure being available when they need it. Those assumptions might be wrong.
There's something almost funny about this situation, in a dark way. The AI industry keeps talking about automation, about robots doing human work, about the future of labor. And its biggest constraint right now is that there aren't enough humans with the right skills to build the physical infrastructure it needs.
You can't automate your way out of needing someone to install a transformer. Not yet, anyway. The robots that could theoretically help with this kind of work are years away from being reliable enough. So we're stuck in this weird limbo where the future of automation depends on very traditional, very human labor markets.
I don't know what the solution is. Training programs take time. Immigration policy is a mess. Wages are rising in these trades, which helps, but you can't conjure skilled workers out of thin air.
First, whether any of the major AI companies start investing directly in workforce development. Not just writing checks to training programs, but actually building pipelines for the workers they need. It would be a long-term bet, and tech companies aren't known for patience, but it might be necessary.
Second, the policy side. There's been some movement on permitting reform for energy infrastructure, but it's slow. If data center construction keeps getting delayed by regulatory bottlenecks, the pressure for faster approvals will intensify. Whether that's good or bad depends on your perspective.
Third, I'm curious whether this creates openings for companies building smaller, more efficient systems. If compute is constrained, maybe the advantage shifts to whoever can do more with less. That could matter for robotics, where edge computing and on-device processing are already important.
Look, I'm not saying AI is going to collapse or that the hype was all wrong. The technology is real and it's improving. But there's a version of the next few years where everything takes longer than the most optimistic projections, where the infrastructure buildout hits snag after snag, where enterprise adoption moves at enterprise speed (which is to say, slowly).
That's not a disaster. It's just reality being more complicated than the pitch decks.
For those of us covering robotics and embodied AI specifically, I think it means being more skeptical of timelines. When a company says they'll have 10,000 humanoids deployed by 2028, ask them where the compute is coming from. Ask about power. Ask about the supply chain for all the boring stuff that makes the exciting stuff possible.
Because right now, the limiting factor on the AI future isn't intelligence. It's infrastructure. And infrastructure moves at its own pace, no matter how much money you throw at it.
