Cheaper inference means more of it, and the cloud rents the capacity

Every drop in the cost of running a model makes new uses economic. That tends to increase total spending rather than reduce it, and the cloud that rents the capacity gains from the volume, whoever built the model.

Price falls, volume rises

When a resource gets cheaper, people find more uses for it. Inference that was too expensive for a feature last year becomes a default this year. Each individual query costs less, and there are far more of them.

For a cloud provider, that shifts the business from selling scarce, expensive capacity to selling abundant, cheaper capacity at enormous scale. That is the business it already runs for storage and compute.

What I'm watching

• Cloud revenue growth alongside falling inference prices
• Utilisation of new capacity
• The share of workloads running on the provider's own chips, which protects its margin

Where I stand

Long, on volume outrunning price declines. The cloud gains either way the model race turns out.

What would change my mind

Inference moving substantially onto devices rather than data centres. Cheaper inference that runs locally would take volume away from the cloud rather than toward it.