Claim
AI foundation models are in a temporary supply crunch that resembles pricing power but is not. Every observable structural dynamic points toward commoditization: competing frontier models on similar training regimes, trillion-dollar capacity additions in progress, per-token efficiency improvements each generation, and no clear winner-takes-all dynamic at the model layer. Value will concentrate further up the stack, as it did in mobile and semiconductors.
Mechanism
Token pricing sits between sellers' marginal cost and buyers' ROI. The supply crunch is unstable: multiple labs are building frontier models from shared scientific foundations and similar data sources; data center capex commitments are measured in trillions; and inference efficiency improves with each model generation. Inference carries 40-50% gross margins today, but those figures exclude training costs that currently exceed revenue.
Historical analog: TSMC generates $53 billion in net income, less than Apple alone, despite being the indispensable manufacturing layer beneath everything Apple makes. Cellular carriers built a trillion dollars of infrastructure over two decades while application-layer companies captured the majority of consumer and enterprise value. In both cases, the necessary infrastructure layer remained compressed while value accumulated in what ran on top of it.
Conditions
Holds when: multiple frontier models exist with equivalent capability and a shared training regimen; supply can scale with capital investment; and application-layer products generate higher gross margins than the model API providers beneath them.
Fails when: one lab achieves a durable capability gap that capital alone cannot replicate; or when regulatory barriers limit competition among frontier model providers.
Evidence
"We're in a supply crunch, and this is unstable. All of the variables are in play."
"Every dynamic we can see points to the latter" [commoditization].
"Token price is a function of supply and demand at a level between the sellers' marginal cost and the buyers' ROI."
"Every path to foundation models having market dominance...requires something to change."
"All the value was captured by other people further up the stack."
Comparators cited: cellular industry at approximately $1 trillion annual revenue with $200 billion capex annually; TSMC net income at $53 billion, less than Apple alone; mobile data traffic grew several orders of magnitude over 20 years while carrier margins did not.
Signals
- Multiple frontier models show indistinguishable benchmark performance on core tasks.
- Per-token prices fall quarter-over-quarter even before new supply capacity is online.
- Enterprise buyers route workloads across multiple model providers without friction.
- Application and tooling companies report higher gross margins than model API providers.
Counter-evidence
Evans acknowledges genuine uncertainty: the supply crunch creates conditions where current pricing could persist if demand grows faster than supply. A lab with a durable capability gap could sustain premium pricing if no competitor closes it. Demand-side dynamics remain unclear; only software development currently shows strong product-market fit at scale, and ROI across most enterprise use cases is unproven. If demand grows faster than supply, pricing power could persist longer than the structural signals suggest.
Cross-references
- Reselling inference at cost-plus yields zero margin as model prices compress: Tunguz argued for value-based pricing on the sell side; Evans maps the structural conditions that determine whether value-based pricing is sustainable for model providers.