The voice AI stack is splitting into four layers. Only one compounds.
Every maturing technology stack eventually unbundles into layers, and each layer captures a different kind of value. Voice AI is unbundling right now. If you're deciding where to build, it's worth being precise about which layer you're standing on — because they are not created equal.
The four layers
- The pipeline layer — speech-to-text, turn-taking, text-to-speech, orchestration. The plumbing that moves a conversation.
- The model layer — the LLM that decides what to say.
- The memory layer — identity, persistent memory, and relationship state. Who is this, what's happened between us, where do we stand.
- The application layer — the front desk, the toy, the robot the end user actually touches.
Three of them commoditize
Pipelines are already in a per-minute price war; latency and cost are converging toward zero and the vendors are interchangeable by design. The model layer is fiercely competitive and, for most builders, swappable — you can change your LLM in an afternoon. The application layer is where differentiation lives, but every application has to solve identity and memory from scratch, and most solve it badly.
One of them compounds
Memory is different in kind. Every session a product runs deepens its memory graph, and that graph is not swappable — it's the accumulated history of a relationship. A pipeline you replace loses you nothing. A memory layer you replace loses you everything the product knew. That asymmetry is what makes memory a data asset rather than a commodity input: it gets more valuable with use, and it raises switching costs for the product built on top of it.
This is also why memory shouldn't be re-implemented inside every application. Built once, as a neutral layer that sits below the apps and above the models, it can be model-agnostic — you keep your own LLM and pipeline — while the relationship asset accrues independently of both.
Where to stand
If you're building a product, own your application layer and rent everything you can below it — except memory, which you want to compound rather than commoditize. If you're building infrastructure, the pipeline war is already crowded and racing down. The layer that's still open, still compounding, and still un-owned is memory. That's the one worth building.