You’ve felt it. That uneasy knot when you realize your AI assistant is getting smarter — but also more locked into one company’s ecosystem. Maybe you shrugged it off as competition. It’s not.
The real threat isn’t that AI will become sentient and overthrow us. It’s that three companies already control the physical stuff needed to build it: compute, energy, and data centers. And that concentration is creating a feudal AI landscape where innovation belongs to a few, and the rest of us just rent access.
The bottleneck isn’t innovation — it’s electricity. Every breakthrough in AI requires more computing power. More GPUs. More cooling. More power plants. And building a data center today takes years of permitting, grid upgrades, and billions of dollars. Only those with existing capital and political clout can even play.
You’ve probably read think pieces about AI ethics or regulation. Those matter. But they miss the ground truth. The real lever is infrastructure — the physical scarcity of compute and energy is what locks in power, not just clever algorithms or training data.
I saw this firsthand. A startup with a brilliant new architecture couldn’t train its model because every GPU cluster was already booked by the Big Three. So they pivoted to consulting. Another team tried to build a distributed training network using spare compute from gamers — but latency and energy costs crushed them.
Concentration of compute is the new concentration of capital, and it’s creating a feudal AI landscape. The people who control the chips control the future. And they don’t owe you access.
Here’s the twist: most debates about AI safety focus on alignment or rogue agents. But a brittle, centralized system is inherently unsafe. If one company’s data center goes down — or its priorities shift — entire research agendas die. Diversity of thought requires diversity of infrastructure. You can’t have decentralized governance with centralized compute.
So what do we do? Stop pretending that open-source models alone solve this. A model is useless without the compute to run it. The real campaign should be for public compute infrastructure — think national AI labs, not corporate cloud credits. We need energy policy that prioritizes small-scale, geographically distributed data centers over mega-clusters.
Because right now, your future AI experience is being shaped by a tiny group’s decisions, not by market competition or democratic oversight. And if you think that’s fine, ask yourself: would you let three companies own all the printing presses in the world?
FAQ
Q: Isn't this just a temporary supply-chain issue that the market will fix?
A: No. Compute and energy are structurally scarce due to physical limits (grid capacity, chip fab lead times) and regulatory barriers. Market forces will only deepen concentration because the incumbents have already captured the infrastructure, making it harder for newcomers to enter.
Q: What can a regular AI user do about this?
A: Demand public AI infrastructure — like national research clouds or co-op data centers. Support policies that break up vertical integration in AI (e.g., separating cloud compute from model training). And be skeptical when companies claim 'open source' while controlling the only hardware that can run the models.
Q: Aren't we better off if big companies handle safety through centralized control?
A: That's the argument they want you to believe. But centralized control means single points of failure — a bug, a bad decision, or a government takeover can wipe out an entire research paradigm. Distributed infrastructure with transparent safety audits is more resilient and less prone to catastrophic alignment failures.