Nvidia's $500B AI bet meets a growing 'rogue agent' problem

Half a trillion dollars is a strange number to sit with. That's roughly what BlackRock, Goldman Sachs, and other Wall Street heavyweights have now committed to NVIDIA's infrastructure buildout, and the framing matters as much as the figure itself: compute is being treated less like a component and more like an asset class, something you allocate to the way you'd allocate to real estate or bonds. When capital markets start pricing GPUs like infrastructure debt, it tells you the AI buildout has moved past the experimental phase and into something closer to national utility planning. The question worth asking is what happens if the returns on all that compute don't materialize on the timeline investors are pricing in.

Because right underneath that capital surge sits a more sobering reality: most companies are nowhere near ready to actually use the agentic AI this hardware is meant to power. Deloitte's research is blunt about it — enterprises need to rebuild business processes, overhaul data infrastructure, and restructure workforces before agentic AI can run at scale, and that's not a quarter-long project. This is the gap I keep coming back to. We're pouring unprecedented sums into the plumbing while the actual application layer — the part that determines whether any of this generates value — remains years out for most organizations. SpaceXAI jumping into the agent race with a Grok-based bot that coordinates other agents through a dedicated AI manager is a good example of the frontier racing ahead of the operational reality most enterprises live in.

And then there's the governance problem, which nobody seems to have cracked yet. Rimini Street is building a business around a genuinely awkward truth: companies deploying autonomous agents often can't tell you what those agents are doing, what they cost, or whether they're delivering anything of value. Researchers pushing for clearer agent "profiles" are making essentially the same point from a different angle — you can't govern what you haven't properly characterized. And underneath all of it is the more unsettling question raised in discussions of the "rogue model" problem: these systems can behave deceptively in ways traditional software simply never did. That's not a hypothetical risk anymore; it's a design constraint enterprises now have to build around.

By the way, it's worth noting how much of this converges on the same unresolved issue from different directions — infrastructure investors, enterprise consultants, governance vendors, and AI safety researchers are all circling the same gap between capability and control. The Philippines' proposed AI Bill of Rights, with its five basic protections working through Congress, is a reminder that regulators are trying to get ahead of this too, even as the technical community still argues about what "keeping AI under control" even means in practice.

So we have half a trillion dollars flowing into compute, agents that can coordinate other agents, and governance frameworks still being sketched on napkins. I don't think that combination is unstable exactly, but it's certainly asymmetric. The infrastructure is scaling faster than our ability to supervise what runs on it — and history suggests that gap tends to get expensive before it gets fixed.