[DigitalToday reporter Choi Jae-won] A view has emerged that companies should evaluate the cost of adopting AI by examining overall economics, including infrastructure and long-term operating expenses, rather than model performance or the price per token.

On Aug. 18, Peter Griffiths (피터 그리피스), founder and chairman of Argyll Data Development, argued in a TechRadar guest column that as enterprise AI moves beyond experimentation into core operations, the standard for assessing cost must also change. What matters is not the price of a single token, but how much it costs to provide AI capabilities that are affordable, sustainable and predictable at enterprise scale.

AI services appear to charge costs by tokens, but in reality they depend on physical infrastructure such as compute devices, memory, networks, power and cooling. When AI moves beyond small experiments and begins handling millions of inference requests in production, even small differences in infrastructure efficiency can accumulate over time and lead to significant cost gaps, the column said.

Griffiths said dedicated infrastructure optimized for inference can improve energy efficiency compared with training-focused architectures. Lower power consumption can reduce cooling needs and the complexity of facility design, which in turn can cut operating costs over the full life of the infrastructure.

Usage-based token billing is not going away. It remains a flexible approach when usage fluctuates widely or when companies are still testing AI use. But as AI becomes deeply embedded in day-to-day work and token consumption keeps rising, it can become difficult to predict and control costs. Token usage also shows how much AI was used, but it cannot explain how much that use actually increased revenue or improved productivity.

As a result, building dedicated inference infrastructure in-house was presented as an alternative. Securing AI capabilities on the basis of fixed operating costs, rather than paying each time based on usage, can increase control over performance, data location and system resilience. Combining that with self-generated renewable energy and long-duration energy storage systems could also be considered as a way to reduce volatility in power costs.

Corporate AI performance metrics should also shift from 'how many tokens were used' to 'what business results were delivered,' it said. As foundation models are continuously updated, companies must manage not only costs but also the impact of model changes on performance, compliance and risk. Ultimately, the core of competition in enterprise AI is not buying the cheapest tokens, but building sustainable AI capabilities at predictable costs and creating real business value, it said.