Model Routing Is an L2d Story — and a Law I Receipt.

A line I keep hearing in operator and investor conversations right now:

The framing is right. The layer attribution is usually wrong. Most people are calling routing a new category. Structurally, it isn't. It's a single sublayer of L2 — L2d Model Routing & Composition — getting pulled into the spotlight because the layer above it (L5 Execution) finally has a reason to use it at scale.

Worth decoding properly, because the difference between “routing is a category” and “routing is a sublayer that benefits L5” changes who you'd back, who you'd build, and which roadmap items you'd cut.

The three reasons, mapped to the framework

The argument for routing usually shows up as three bullets. Each one lands cleanly on the stack.

1. Cost optimization. Frontier intelligence for planning and review, cheaper or open-source models for the bulk of the workload. This is pure L2d Model Routing & Composition — the sublayer was literally defined for this. It's also the most visible signal of Law I (Intelligence Commoditizes Downward) in action: once you can swap a frontier call for an OSS call mid-workflow without quality loss, the frontier model has been commoditized for that step. Not someday. That step, that workload, today.

2. Capability maximization. Different models are better at tool use, coding, or domain knowledge. Route accordingly. Still L2d as the capability — but the decision of which model to call for which step is made at runtime by L6a Agent Loops and L6c Role Routing & Task Decomposition. This is the L2↔L6 seam, and it's why standalone “router” startups keep getting absorbed into orchestration frameworks. The capability lives at L2d; the caller lives at L6.

3. Risk mitigation. If a government restricts a model or a provider has a Fable-style incident, you want to swap providers without rebuilding the workflow. That's L3a Compliance & Export Controls treating routing as a compliance primitive, not just a cost lever. Once routing is load-bearing for compliance, it stops being optional infrastructure — and L3's grip on the stack tightens by one more notch (consistent with Law IV — Generation and Verification Must Be Separate).

Three reasons, three layers touched: L2d (the capability), L6a/L6c (the runtime that uses it), L3a (the compliance reason it becomes mandatory). One framework, no new vocabulary required.

Who actually captures the value

Here's where the popular framing gets the conclusion right for the wrong reason.

The claim is that routing shifts value to “the applied AI layer.” Correct. In framework language: L5 Execution captures the surplus. But not because L5 owns routing. Because L5 finally gets to treat L2 as a swappable commodity — which is the textbook definition of Law I.

This is the same pattern that played out with databases under SaaS, with compute under SaaS, and with CDNs under media companies. The layer that commoditizes doesn't capture the value it releases — the layer above does. Routing is L2 telling on itself.

Which means the interesting question isn't “who builds the best router.” It's “which L5 companies are positioned to harvest the L2 surplus when routing is free.” Two answers stand out:

L5 companies with L1b Proprietary Data. If your moat is data the model can't get elsewhere, the choice of model becomes irrelevant — you win on inputs, not on inference. Harvey, Bloomberg-style verticals, anything with regulated corpus access.

L5 companies with L8 Memory that compounds. If the system remembers the user, the workflow, and the institution across sessions, swapping the underlying L2 is a config change. The moat is in L8, not in which model answered the last call. This is Law III (Surface Captures Attention; Chain Captures Power) read forward.

Where routing as a product gets compressed

Standalone routing products are in a familiar spot. The capability is real. The defensibility isn't.

L2 providers will ship routing themselves (OpenAI's model picker, Anthropic's tier selection, Bedrock-style multi-model endpoints) — that's L2d absorbing the function back into L2. Orchestration frameworks (LangChain, CrewAI, the agent stacks) will ship routing as a built-in primitive — that's L6 absorbing it upward. Both directions of compression. Same logic as every other middleware story in the last two cycles.

The companies that survive at L2d will be the ones that turn routing into something neither L2 nor L6 can easily replicate: cross-provider evals tied to L3b Quality Gates, enterprise audit trails tied to L3a, or workload-specific routing policies tied to a vertical's compliance regime. In other words — L2d only survives by anchoring into L3. Pure routing is a feature.

What to do with this

If you're a product leader at an L5 company: routing is good news for you. It means the cost and capability ceiling above you is about to drop. Spend the surplus on L1b (proprietary data) and L8 (memory). Do not spend it on building your own router — that's the layer commoditizing.

If you're an investor: when you see a pitch deck for a routing company, ask one question — what L3 surface are they anchored to? No answer means they're a feature waiting to be absorbed. A real answer (audit, compliance, eval governance) means they may have a wedge.

If you're a founder building at L2d: the capability is necessary. Necessary capabilities without anchors get absorbed. Pick your anchor early.

The bigger pattern

Model routing isn't the story. It's a symptom. The story is that L2 is becoming the layer everyone routes through rather than the layer everyone builds on top of. That's what commoditization looks like from the inside — and it's exactly what the framework predicts.

Watch for the same pattern at L7 over the next twelve months. Surface routing — picking between ChatGPT, Gemini, Copilot, and Claude for the same task — will follow the same arc. Same law. Different layer.

The full framework — 10 layers, 50 sublayers, 4 structural laws — is free at supplychainofai.com. No signup, no paywall.