When the Model Is Free: Pricing Pressure Across the Industry
Falling prices are good news for anyone buying inference and awkward for anyone selling it. The real danger to standalone vendors is not a rival’s price list.
The price of a given level of model capability has fallen dramatically and continuously. A task that cost dollars per thousand operations a few years ago now costs cents, and open-weight alternatives make some of it effectively free.
Cheaper inference, denser hardware, free rivals
- Architectural and engineering efficiency: better inference at the same quality.
- Hardware improvements: more throughput per unit of capital.
- Competition, including free alternatives that set a ceiling on what can be charged.
Software margins without software physics
Software companies are used to falling costs, because software has near-zero marginal cost. Serving a model has a real marginal cost that falls but never reaches zero. A company that charges for usage is therefore exposed on both sides: price falls, and unit cost falls more slowly.
The defences are familiar from other infrastructure businesses. You can move up the stack into applications where the model is a component and the customer pays for outcomes. You can specialise in workloads where reliability, compliance or latency justify a premium. Or you can own a distribution channel that makes your model the default regardless of price.
Why a twelve-month contract looks expensive by month twelve
For anyone buying these services, the practical implication is that multi-year commitments at fixed prices are risky and portability is valuable. The cost of switching has fallen along with the price, and the vendor that looks expensive in a proposal may be behind the market within twelve months.
The rival is not another model vendor
The most serious pricing pressure comes not from competitors’ price lists but from bundling. When an adequate model is included in a productivity suite a company already pays for, the standalone vendor must justify a separate line item, and most buyers will not.
This is the dynamic that reshaped databases, analytics and collaboration software, and there is no obvious reason AI should be different. Vendors with a distribution channel are structurally advantaged over vendors with better technology.
Narrower markets that are harder to lose
The response that has worked in other infrastructure markets is to move into workloads where general-purpose products are inadequate: regulated industries with specific audit requirements, workflows with hard latency constraints, or domains where proprietary data provides an advantage no general model can match.
Each of those markets is smaller and harder to win, and each is more defensible once won. That trade-off is the choice facing every model provider that is not also a platform.
Image credit and licence details for every photograph on this site are listed on the credits page. This article is editorial content; it carries no sponsored material.