As enterprises deploy AI at scale, choosing the right model size and the right infrastructure can significantly reduce the energy footprint of AI workloads.

Every Earth Day brings renewed focus on how technology can reduce its environmental footprint. For the AI industry, that conversation is becoming increasingly important. Artificial intelligence is now running at scale across enterprise applications, and the infrastructure required to support those workloads continues to grow.

But scaling AI does not have to mean scaling energy consumption at the same rate. A shift is already underway across the industry toward right-sizing AI models for the task at hand. In many cases, that means deploying small language models (SLMs) that can deliver strong results for specific workloads while requiring far fewer computational resources.

When these models are paired with efficient compute platforms like Ampere CPUs, organizations can scale AI capabilities while keeping infrastructure growth and energy usage under control.

The Shift Toward Right-Sized AI
The early wave of generative AI innovation focused on the capabilities of very large models. These systems demonstrated impressive performance across a wide range of tasks and pushed the boundaries of what AI could do.

But as enterprises begin deploying AI into real production environments, a different reality is emerging. Many AI workloads today do not require the full breadth of a massive general-purpose model. They require focused intelligence within a defined domain.

Internal knowledge assistants, enterprise search tools, document analysis systems, and workflow automation platforms often operate on structured or domain-specific data. For these types of applications, smaller models that are trained or fine-tuned for a particular task can perform extremely well while requiring dramatically fewer resources to run.

This is leading to a more deliberate approach to AI architecture. Instead of defaulting to the largest available model, organizations are increasingly choosing the right model size for each workload.

Right-sizing models allows enterprises to deploy AI more broadly without dramatically expanding infrastructure requirements.

Smaller Models, Lower Energy Demand
This shift has meaningful implications for sustainability.

Every AI inference consumes compute cycles, and those compute cycles require energy. When AI services scale across an organization to serve employees, customers, and automated systems, the cumulative energy footprint can grow quickly.

Small language models help reduce this impact by lowering the amount of compute required for each request. Because they are designed for narrower tasks, they often achieve strong performance with significantly fewer parameters and lower computational overhead.

For enterprises deploying AI across thousands or millions of interactions, these efficiency gains compound. The difference between running a massive model for every request versus selecting a smaller, task-appropriate model can translate into substantial reductions in both infrastructure demand and energy consumption.

In practice, sustainability in AI increasingly comes down to architectural decisions. Choosing the right model for the job can be just as important as the hardware used to run it.

Right-Sizing Compute Matters Too
Model selection is only part of the equation. The infrastructure running these models plays an equally important role.

Much of the early generative AI conversation centered around large models running on specialized accelerators. That approach makes sense for training frontier models and for certain high-throughput inference workloads, but it is not always necessary for many of the AI applications enterprises are deploying today.

Small language models change the equation. Because they are smaller and more targeted, many SLM workloads can run efficiently on general-purpose compute platforms. Tasks like enterprise search assistants, document summarization, workflow automation, and internal copilots do not often require the scale or parallelism that accelerators are designed to deliver.

Instead, these models can run efficiently on CPUs, simplifying infrastructure and reducing the overall power required to support AI services. Ampere CPUs, for example, are designed with efficiency at their core, delivering strong performance while prioritizing power efficiency and predictable scaling for cloud environments. For organizations deploying SLMs, running inference on these CPUs allows AI services to scale without dramatically increasing power consumption.

A More Thoughtful Approach to Scaling AI
As AI adoption expands, efficiency is becoming a core part of AI system design. Enterprises are learning that deploying AI responsibly is not only about model capability, but also about selecting the right model for the task and running it on infrastructure optimized for efficiency.

Small language models represent an important part of that strategy. By focusing on task-specific intelligence rather than maximum scale, they allow organizations to bring AI into everyday workflows with far less computational overhead.

Combined with efficient platforms like Ampere CPUs, this approach enables enterprises to deploy AI widely while keeping infrastructure growth manageable and energy usage lower than it would otherwise be.

On Earth Day, the conversation about sustainable technology often focuses on long-term goals. In AI infrastructure, however, sustainability increasingly comes down to practical choices made every day: selecting the right models, running them on efficient platforms, and designing systems that scale responsibly.