The infrastructure required to run AI is becoming the real business, and it's fracturing along three distinct fault lines: the physical constraints that no amount of software can overcome, the consolidation of economic power through technical lock-in, and the messy reality of deploying these systems into organizations that were never designed for them.
Start with the material problem. A data center consumed 30 million gallons of water without anyone noticing for months. Cowboy Space raised $275 million to build data centers in orbit because there aren't enough rockets to launch them and they're too expensive anyway. These aren't metaphorical constraints. They're hard limits on how fast the industry can scale. Meanwhile, Nvidia's CUDA ecosystem has become a software moat so deep that switching costs dwarf the value of hardware innovation itself. The company that controls the programming layer controls which customers can run what workloads, and that control compounds with every model trained on CUDA infrastructure. This is how software becomes infrastructure becomes monopoly.
The second fracture is organizational. GM laid off hundreds of IT workers to hire people with AI skills, but those new hires are being asked to do something the old IT organization was never structured to support. Finance departments are experiencing what the reports call a "quiet insurgency" where employees deploy AI while leadership scrambles to impose governance afterward. Robinhood is raising a second venture fund specifically to catch the growth-stage winners emerging from this chaos. The companies that will dominate the next five years aren't the ones with the best models. They're the ones whose existing business model and cost structure can absorb the disruption without breaking. That favors incumbents with capital and regulatory moats. The startups training AI on Hugging Face are getting hit with malware delivered through repositories impersonating OpenAI. The infrastructure is still too young to have basic hygiene.
The third fracture is political and geographic. The European Commission is drafting rules to restrict US cloud services for sensitive data. Colorado is rewriting its AI regulations after two years of collapsed deals. The IMF is warning about AI-accelerated cyberattacks on financial systems. None of these are abstract policy concerns. They're direct responses to the fact that compute infrastructure is now a national security asset and the US controls most of it. Daron Acemoglu published a paper months before winning the Nobel Prize in economics that earned him few fans in Silicon Valley. The tension isn't between regulation and innovation. It's between who gets to extract value from the infrastructure layer and who has to pay to run on top of it.
Sloane Duvall