The model layer gets the attention. The infrastructure layer gets the margin. Underneath every impressive demo sits a stack of compute, data and orchestration that has to exist for any of it to work, and that stack is where a lot of durable value will be captured over the next decade.
The catch is that "AI infrastructure" is not one market. It is a dozen, with very different competitive dynamics. Calling something a picks-and-shovels play is the beginning of the analysis, not the end.
Three layers, three questions
We find it useful to compress the stack into three layers and ask a different question of each.
Compute: who controls scarcity?
At the bottom is raw compute: chips, interconnect, data centres, power. This layer is real, enormous, and largely spoken for by incumbents with balance sheets a startup cannot match. The early-stage opportunity is rarely in competing head-on. It is in the inefficiencies around scarcity: utilisation, scheduling, specialised silicon for narrow workloads, and the unglamorous tooling that squeezes more useful work out of constrained hardware.
Backing a startup to out-build a hyperscaler on compute is not a thesis. It is a hope.
Data: who owns the proprietary loop?
The middle layer is data: pipelines, labelling, synthetic generation, evaluation, retrieval. This is the layer where defensibility is most often claimed and least often real. Generic tooling here commoditises quickly because the buyers are sophisticated and the switching costs are low.
The exceptions are companies that sit inside a proprietary data loop, where using the product generates data that makes the product better in a way a competitor cannot copy. That is a moat. A wrapper around a public model is not.
Orchestration: who becomes the default?
The top of the infrastructure layer is orchestration: the frameworks, agents, observability and deployment surfaces that developers actually touch. Value here is about becoming the default, the thing teams build on and do not want to rip out. Distribution, developer love and timing matter more than raw technology. This is where a small team can still win, and also where most of them quietly lose to an open-source project.
The defensibility test
For any AI infrastructure company, we run the same blunt test:
- Does it get harder to displace over time? Switching costs, data loops, standards. If not, today's growth is rented.
- Is it exposed to the model layer eating it? Many tools are one model release away from being a feature. Ask what happens when the frontier labs ship the obvious version.
- Is the buyer sophisticated and price-sensitive? Selling shovels to people who can forge their own is a margin trap.
A company that passes all three is rare. Most pitches pass one, dress it up as three, and hope you do not check.
The Banyan view
We are structurally interested in AI infrastructure because it is where real cash flows are forming while the headlines chase the models. But the picks-and-shovels framing is a trap if you stop there. The money is not in selling shovels. It is in selling the one shovel that stays scarce after the rush, owning a proprietary loop, or becoming the default that nobody wants to replace.
Everything else is a commodity with good marketing. Our job is to tell the difference early.