The standard commoditization story says the foundation model moat is eroding from below: open-weight models keep getting cheaper and better, so proprietary APIs lose pricing power. That story is correct. It's also incomplete.
There are two more pressures running at the same time. Together, they're doing something the commoditization thesis alone doesn't capture: they're hollowing out the model layer from three sides at once.
Pressure one: open weights
DeepSeek and Qwen have crossed the threshold that matters for most production workloads. Not AGI-level reasoning. Not the frontier. The threshold where the model is good enough to deploy, cheap enough to run locally, and malleable enough to fine-tune for a specific domain — and where paying for a proprietary API is a hard question to justify.
Startups shipping on Chinese open-weight models aren't making an ideological statement. They're making an economic one. When the quality gap closes to "acceptable" and the cost gap is ten to one, the decision isn't hard.
The question that follows is what the proprietary providers are actually selling. For genuinely hard reasoning tasks, there's still a clear answer. For everything else — summarization, extraction, classification, routine generation — that answer is getting shakier.
Pressure two: architectural alternatives
AMI Labs raised over a billion dollars on the thesis that next-token prediction is architecturally wrong for physical-world reasoning. I wrote about this in April, but the implication for the model layer specifically is worth pulling out separately.
If world models are the right architecture for embodied AI, then the current scaling path — bigger transformers, better RLHF, more compute — optimizes for the wrong thing. That's not an immediate threat to most language use cases. But it means the next major category of AI application may not be won by the incumbents, regardless of how good their current models are.
The foundation model moat assumes a single architectural trajectory. AMI Labs' billion-dollar bet says that assumption is wrong, at least at the frontier.
Pressure three: regulatory friction
State-level AI regulation in the US is becoming operationally real, without any corresponding federal framework to organize against. Hiring algorithms, automated evaluation, transparency disclosures, and algorithmic accountability requirements are accumulating at the state level — California, Colorado, Illinois, now more — and they vary enough to turn compliance into a product design problem.
Every foundation model provider is now carrying compliance uncertainty as a cost. Every application layer built on top of them is too. This doesn't kill the model layer. It taxes it — and the tax rate is rising.
Where the value went
The model layer is being squeezed from below by open weights, forked architecturally by world models, and taxed from above by regulation. The layer that benefits from all three pressures is orchestration.
Not because orchestration is glamorous. Because it's the only layer that gets more valuable as model selection becomes more complex. When you're routing between DeepSeek and Claude depending on task type, handling compliance gating by jurisdiction, and building in fallback paths for the day a vendor changes pricing or an architecture shift makes a different model category relevant — the routing layer is load-bearing in a way the model layer isn't. It's indifferent to which model wins. It abstracts the choice — and captures the margin that used to live in the model itself.
What's different from the standard commoditization argument is the simultaneity. The pressures aren't sequential. Open weights, world models, and regulatory friction are all compressing the model moat at the same time, from different vectors.
The practical question
If you're building on a single foundation model provider without a routing or abstraction layer underneath, you've bet on one horse surviving all three pressures intact. That's a specific bet. It might pay off. But you should know you're making it.
The more defensible position: treat model access as an infrastructure commodity (because it increasingly is), build the orchestration layer that handles selection and compliance, and stay architectural about which tasks actually require frontier reasoning versus which ones just got routed there by default.
The durable bet isn't on which model wins. It's on the infrastructure that makes it not matter.