Enterprise AI needs workload-specific models, hardware and deployment

Businesses moving AI from pilot projects into production face costs that can rise as applications are used more often or agents call several models to complete a task, argues Emma Cloney, a Core42 executive, in an opinion article for TechRadar.

Cloney says companies should match models and accelerators to each workload instead of defaulting to the most powerful options. Smaller models can handle routine classification, extraction and routing; mid-sized models can support everyday reasoning; and larger models may suit complex reasoning, strategic planning and tasks where accuracy and nuance matter.

Hardware choices also depend on the work. Low-latency applications need responsive systems, while high-volume workloads may benefit more from throughput and cost efficiency. Cloney argues that using one type of hardware for every task can mean paying for performance that some applications do not use. She says a mix of accelerators can give companies more flexibility, naming NVIDIA, AMD, Qualcomm, Microsoft and Cerebras as possible providers.

As an example, the article says Cerebras has run OpenAI's gpt-oss-120b at around 3,000 tokens per second, which Cloney says is valuable when response time is critical. She also highlights deployment location and local control, particularly for governments and companies handling sensitive information. The article notes demand in Europe for infrastructure developed and operated according to local requirements.

Cloney's central argument is that businesses should choose AI infrastructure according to workload needs, including speed, cost, control and data residency, rather than commit every project to one model, hardware type or deployment route.