In the past two years, large frontier AI models, boasting billions to trillions of parameters, have gained significant traction, integrating into various applications and business tools almost daily. However, the rapid expansion of these resource-intensive models raises concerns about their sustainability. This has prompted major tech companies, including Microsoft, Google, and AWS, to explore nuclear power to support the massive data center infrastructure needed for future AI growth. While cutting-edge models from OpenAI, NVIDIA, Google, and Anthropic demonstrate unprecedented capabilities, they often exceed requirements for many applications, likened to using a Formula 1 car for daily commuting. Thus, there's a growing interest in smaller models, typically containing hundreds of millions to less than 10 billion parameters, that are efficient and cost-effective. NVIDIA recently introduced its NIM (NVIDIA inference Microservice) technology, which streamlines AI deployment and allows for industry-specific applications, such as drug discovery and code generation, while minimizing computational demands.

The collaboration between NVIDIA and Accenture exemplifies how specialized microservices and expertise can facilitate faster AI adoption in enterprises. Similarly, IBM unveiled its Granite 3. 0 models, competitive with Llama and Mistral, highlighting the effectiveness of small language models in various applications. IBM’s open-source solutions can be implemented on multiple cloud platforms, including its own watsonx, demonstrating a strong focus on enterprise needs. The future of enterprise AI looks toward a blend of models and adaptable infrastructure that prioritizes outcome-based projects, driving advancements in areas like automation and digital labor. The notion that a single, large-scale model can address all enterprise needs is misguided; many defined use cases would benefit more from smaller, specialized models that offer better energy efficiency and cost management, while simplifying data governance. Research into sophisticated AI will continue, but for many business applications, smaller foundation models represent a more sustainable approach. This not only allows for broader and more efficient AI deployment but also significantly reduces operational costs, making it an attractive choice for enterprises seeking to harness the potential of generative and agentic AI solutions.