When AI Models Proliferate, the Models Themselves Aren't the Most Valuable Asset
Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle
A few years ago, the core question in AI was “whose model is strongest?” Now that question is quietly being replaced by another: when models multiply and their capabilities converge, how should users choose? Which one? How?
A telling move: a major payments company spent a valuation in the tens of billions to acquire a “model routing” company. This company doesn’t train models or sell models. What it does is more straightforward—it helps users decide “which model should handle this task” among dozens of options. A company that doesn’t produce models is worth more than many that do.
This isn’t an anomaly. It’s a structural inflection point in the AI industry: once models converge, value begins migrating to the layer “between models.”
The More Similar the Models, the More Valuable the Choice
When every model can chat, write code, and summarize documents, “using a good model” stops being an advantage and becomes infrastructure. At that point, the real cost isn’t the model itself—it’s the selection and switching.
Users face a swarm of models: some cheap, some fast, some good at long context, some good at code. Choose wrong and you waste money. Choose wrong and you waste quality. Choose wrong and you have to redo the work. These “costs of choice” used to be no one’s problem. Then a new role emerged—the layer that makes decisions between models for you.
The value of this layer isn’t in how strong it is, but in what it possesses that others don’t: who used which model for which task, what the results were, and what it cost. This cross-model data isn’t available to individual model providers, but the routing layer has all of it. With data in hand, it can make more accurate recommendations. That’s the moat.
The Real Moat of the Routing Layer: Not Technology, but Network
The value of a routing tool isn’t how sophisticated its algorithm is, but that as it connects more models and users, its data grows richer, its recommendations sharper, attracting even more users—a virtuous cycle. Classic network effects.
This also means the routing layer is a “winner-take-most” business: once one player accumulates sufficient data depth, its recommendations stay consistently more accurate, and latecomers struggle to catch up. That’s why model providers, cloud giants, and independent routing companies are all crowding into this lane.
But This Layer Has an Underestimated Crack
The routing layer looks attractive, but it has a fatal weakness: it depends on model providers.
If model providers one day build their own routing (they have every incentive and ability to do so), or if users start bypassing routing to call models directly, this layer’s value evaporates. Its position is a bit like a “toll booth”—it makes money as long as cars take the highway, but if someone builds a new road, it becomes obsolete.
So the real test for the routing layer isn’t “can it make good choices,” but “can it make both model providers and users unable to do without it?” The former relies on data, the latter on whether it’s embedded in users’ workflows.
Implications for Entrepreneurs
If you’re evaluating this space, my take is: the routing layer is a real opportunity, but not an easy business.
Its value lies in data, so you need to connect enough models and users early to capture data accumulation; its risk lies in dependence, so you need to figure out how to avoid being bypassed—embed in workflows, accumulate proprietary data, or become the irreplaceable link in a specific vertical scenario.
Simply building a “model price comparison tool” won’t hold—price comparison has no barrier. The real opportunity lies in “understanding better than model providers what users need in which scenarios,” and turning that understanding into something others can’t take away.
Key points: Value shifts to the routing layer as models converge; the routing layer’s moat is cross-model data + network effects; but dependence on model providers is a crack; opportunity lies in capturing data and embedding in workflows, not in price comparison.