Can tech companies learn to love cheaper AI models?

The AI industry is on the cusp of a major transformation, as the long-held belief that bigger AI models are inherently more powerful and successful is being challenged by mounting costs.

Early predictions indicate that the vast majority of AI workloads could shift to cheaper models within the next 12-18 months. This represents a significant departure from the current norm, where companies prioritize the most advanced models. The implications for the economics of AI are substantial, potentially affecting major AI labs that have invested heavily in cutting-edge technology.

Initial tests demonstrate that smaller, more cost-effective models can maintain performance without sacrificing quality. For example, legal AI tool Harvey successfully reduced inference costs by threefold while preserving quality through a strategic combination of models.

This evolving trend highlights a new definition of quality in AI: one that emphasizes efficiency and cost-effectiveness alongside capability. The core debate is not between proprietary and open-source models, but rather between large and small models, as companies seek to optimize their operations and reduce expenses. This shift poses a challenge to the "scaling-first" approach that has previously dominated the industry, encouraging a more measured and efficient use of computational resources.

However, it remains uncertain whether this cost pressure will genuinely drive enterprises away from larger models. Companies might instead find alternative ways to economize, such as reducing the number of calls to existing models or optimizing current deployments. If smaller models prove equally effective for most applications, it could significantly impact the demand for high-end inference and raise questions about the justification for the substantial costs of developing frontier models.