AI is rebuilding the layer stack of every industry, including its own.
Every general-purpose technology reorganizes the economy into layers, and every re-layering moves the profit pools: some layers commoditize into utilities, new chokepoints form, and scarcity relocates. Re-layerings are when new companies get to win. Incumbents own the industry as it is, and builders own the redraw. That AI is transformative is consensus. Which layer of which stack captures the transformation is not, and that question decides which companies should be built, which founders should be backed, and where returns are made.
Being right about the technology is not enough.
Personal computing reshaped the world exactly as promised, and the companies that made the PCs were destroyed anyway. Once the design standardized, the assembly layer commoditized, and the industry's profits migrated to the two layers where scarcity persisted: the chip and the operating system. Two companies captured over 80% of the industry's profits while a dozen PC makers fought over the rest. Then mobile and cloud re-layered the stack, both winners missed the shift, and an entire generation of new companies was built in the layers that opened. The returns went to those who were right about the layer, and stayed with those who saw each shift coming.
How we read a stack.
For every industry AI touches, we break it into its layers, and for each layer we ask five questions:
- Scarcity Does value pool here because the output is hard to substitute, and is that scarcity structural or temporary?
- Standardization Is the output becoming interchangeable? Interchangeable layers become utilities, and their profits drain to their neighbors.
- Customer ownership Who owns the relationship? Whoever does keeps a claim on value even when their technology commoditizes.
- Capital intensity Does capital protect this layer or expose it? High barriers protect incumbents in shortage and destroy them in glut.
- Complements Who wants this layer to become a commodity, and how powerful are they?
The answers tell us, for each layer, whether a great business is possible there, likely there, or forbidden there. The exercise is never finished: value migrates each cycle, and the layers must be reread as the stack develops.
What this tells us to do.
Some layers decide a company's fate before it is founded. No team can build a great business on an undifferentiated product, and we say no there, even to exceptional founders, because the kindest thing this analysis can do is keep builders off dead terrain. Some layers reward whoever gets through with structural strength: the chokepoints in hard technology, where physics, process knowledge, or infrastructure gates the market. There we back founders at the moments of discontinuity when closed layers briefly open. And some layers decide nothing, because the company decides everything. In applications, the same layer holds the best and worst businesses in AI, separated only by speed, depth of integration, and proximity to the customer's real work. There, the founder is the thesis.
Where we are looking now.
The framework currently points us at the physical foundations of AI: energy and grid infrastructure, photonics and interconnect, advanced silicon and manufacturing, critical materials. At defense and dual-use systems, where the customer and the moat are unlike anywhere else in technology. And at vertical AI applications, where agents take on real work in law, healthcare, finance, and the industrial economy. These are where the layers look most alive to us today. The list changes as the stack does.
The work.
The framework is constant. How we act on it changes with the stage. In the Lab, we start companies where the ceiling justifies building from zero. In early-stage, we back founders where the contest is live and the layer is worth fighting for. In our access vehicles, we structure positions in the winners once they are visible. One framework, entered at three moments, from the first line of code to the last private round.