Two Very Different Bets
"AI" spans everything from foundation model training infrastructure to narrow, vertical applications built on top of existing models. Our research team evaluates companies across both layers, which gives us a grounded view of where the durable value is likely to sit.
How We Evaluate Both Layers
- Benchmarking harnesses and evaluation suites we run against infrastructure and application-layer claims
- Internal test infrastructure built specifically to validate the performance and reliability claims of the companies we diligence
Infrastructure: Where Defensibility Comes From Depth
Infrastructure and tooling companies tend to win on hard technical problems — efficiency, reliability, developer experience — that are difficult to replicate without genuine expertise. We look for teams solving problems we recognize as hard from our own evaluation work.
Applications: Where Defensibility Comes From Distribution and Data
Application-layer companies rarely win on model access alone. We look for proprietary data loops, workflow integration, and switching costs that persist as the underlying models commoditize.
Our Take
We evaluate both layers, but we underwrite them differently — infrastructure on technical depth, applications on distribution and data advantage.
Sources
- Bornstein, Casado, and Appenzeller, "Who Owns the Generative AI Platform?" (Andreessen Horowitz, 2023) — the widely-cited analysis of value capture across the infrastructure, model, and application layers of the generative AI stack. Our "depth vs. distribution" framing above draws directly on it.
- Kwon et al., "Efficient Memory Management for Large Language Model Serving with PagedAttention" (SOSP 2023) — see our own hands-on look at this exact infrastructure-layer dynamic in Technical Evaluation #001.
Where We Take This Next
When our evaluation supports conviction, the opportunity goes to our network of outside investors, who can invest directly — whether or not Quant Labs also commits its own capital.
Quant Labs View
What matters technically: whether a claimed advantage — infrastructure efficiency or application-layer stickiness — actually holds up once the underlying models commoditize further, since that's the pressure test both layers face.
What we would test: for infrastructure, real throughput/cost benchmarks against the current best alternative, not the alternative at launch time. For applications, whether the product's data loop and workflow integration would survive a customer swapping the underlying model provider.
What could invalidate the thesis: a foundation model provider absorbing the infrastructure layer's advantage directly (making the tooling redundant), or an application's "proprietary data loop" turning out to be replicable by a well-resourced competitor in months rather than years.
Commercial implication: we underwrite infrastructure bets on technical depth and applications on distribution and data — the diligence question is different by layer, and we don't apply one evaluation template to both.