Every gold rush has two kinds of participants. There are the people panning in the river, and there are the people selling them shovels, boots, and dinner. The panners mostly go broke. The shovel sellers buy the town.
I spend my days looking at AI agents — what they actually do, where they break, which ones survive contact with a real workflow. And the thing I keep running into is that agents are the panners. They are the visible, exciting, occasionally disappointing part of the story. The money question is different: who sells them the shovels? In 2026, when the buyer is the U.S. federal government and the shovel is a data center, that question gets a lot more interesting.
Nvidia is no longer a chip company in any useful sense
One framing I saw recently described Nvidia as the company that went from selling GPUs to gamers to becoming the AI arms dealer of the 21st century. That is a little theatrical, but it captures something real. Nvidia continues to lead AI infrastructure in 2026, and it leads not because it makes the fastest silicon but because it sits at the chokepoint where every serious AI workload has to pass through.
From where I sit, that matters for a specific reason. When I evaluate an agent platform, I am indirectly evaluating a compute supply chain. The agent that handles your document review is renting time on hardware somebody had to buy, install, cool, and power. Nvidia captures value from that chain whether the agent on top of it is brilliant or useless. That is a very different risk profile than betting on any individual AI product.
Where defense spending changes the math
Commercial AI budgets are discretionary. If the productivity story wobbles, a CFO can pause a rollout in a quarter. Federal defense spending does not behave that way. It moves on multi-year authorizations, program cycles, and procurement rules that are slow to start and equally slow to stop.
Two names come up repeatedly in this intersection. Palantir supports both AI and defense sectors, sitting at the software layer where government data becomes something an analyst or an automated system can act on. TTM Technologies operates further down the stack, in the physical components that AI and defense hardware depend on. One sells the interpretation, the other sells the substrate. Neither is a pure AI story, and that is arguably the point.
The pattern I find worth attention is that federal AI work rewards companies that already know how to sell to the government. Clearance requirements, compliance frameworks, and procurement processes function as a moat that no amount of model quality can cross quickly. A startup with a better agent architecture still needs years to get through the door. An incumbent with existing contracts does not.
The broader cast
The infrastructure story extends past these three. Marvell, Salesforce, and Microsoft each play roles in the 2026 AI buildout, at different layers — networking silicon, enterprise software, and hyperscale cloud capacity respectively. Most investors in 2026 seem to agree that putting money into established AI leaders is the clearest way to participate in the boom, which is a consensus view and should be treated as one.
I want to be direct about the limits of what I can tell you here. Current data lacks specific 2026 updates for several of these companies. I am not going to fill that gap with numbers What follows is a way of thinking about the space, not a forecast.
How I’d frame it as an agent person
A few things I keep in mind when the AI infrastructure conversation comes up:
- Agents are demand, not supply. Every new agent deployment adds inference load. The interesting question is who gets paid per unit of that load regardless of which agent wins.
- Government timelines are a feature. Slow procurement cuts both ways, but it makes revenue less sensitive to AI sentiment swings than a commercial SaaS contract is.
- Consensus is already priced. When most investors agree the established leaders are the clearest way in, that agreement is not a secret. It is an input.
- Layers behave differently. Silicon, boards, cloud capacity, and application software all respond to the same buildout on different schedules and with different margins.
None of this is investment advice — I curate agents, not portfolios. But the discipline is similar. When I assess an agent, I ask what happens to it if the hype cools and only the useful parts survive. Applied to infrastructure, the same question sorts the field fast. Defense-linked AI work is one of the few areas where the answer does not depend much on the hype holding up.
Shovels still sell after the rush ends. Someone has to dig the next hole.
🕒 Published: