Salesforce on Reducing AI’s Energy Use Through Efficient Models and Cleaner Power

AI’s environmental cost depends on more than the size of a model. The electricity powering its training and use, the efficiency of its hardware, and the location of its data centers all affect its emissions. Salesforce climate and emissions leaders Megan Lorenzen and Boris Gamazaychikov argue that companies can address these factors while also improving costs and response times.

The scale of the challenge is growing with AI adoption. Research from the Center for the Governance of AI estimates that the energy needed to train and run leading AI systems increased by a factor of 350 million over 13 years. That computing capacity has helped advance large language models, protein folding and autonomous vehicles, but it also relies on data centers that consume substantial electricity.

Gamazaychikov says data centers already account for a few percent of global electricity consumption and emissions. Research suggests AI could potentially triple or quadruple data centers’ energy needs. The pressure may be particularly sharp in regions where new data center construction is concentrated, since local power supply and emissions intensity shape the impact of added demand.

Efficiency starts with the model choice

Salesforce’s first recommendation is to match a model’s size to its intended task. Gamazaychikov says the company has developed domain-specific models for customer relationship management work, which he describes as more efficient than using very large, general-purpose models for those tasks. The practical point is straightforward: when a smaller model can meet a product’s needs, deploying a larger one may add computing demand without a corresponding benefit.

Model selection is only part of the calculation. Hardware efficiency and the data center’s electricity source also matter. Gamazaychikov reports that GPU efficiency has improved 36-fold in seven years, making the choice of current, efficient hardware one lever for reducing the energy needed to train and serve models. The figure speaks to hardware progress; it does not mean that total AI energy use has fallen, since demand for computing can grow as systems become more capable and widely used.

Where a model runs affects its emissions too. Salesforce maps the carbon intensity of data center electricity and favors lower-carbon sources where possible. Two deployments with similar computing requirements may therefore have different emissions if the electricity supply differs. For companies that buy AI services rather than train models themselves, Gamazaychikov says providers should disclose energy efficiency and carbon footprints. He compares this information to fuel efficiency when choosing a car: it gives buyers a basis for comparing operating impacts.

Common metrics are still missing

That comparison is difficult today because there is no uniform standard for measuring the sustainability of AI systems. Without consistent methods, a reported energy or emissions figure may not be directly comparable with another provider’s figure. Salesforce says it supports standardized measurement and disclosure so organizations can judge systems on a clearer, shared basis.

One proposal comes from Hugging Face, which Salesforce has invested in. The AI firm has suggested an “Energy Star for AI models” rating system. Gamazaychikov says Salesforce is exploring a pilot with its own models, with the aim of contributing to a simple, widely usable efficiency metric. A rating could help organizations weigh energy use alongside model quality, price and latency, though a useful system would depend on clear measurement boundaries and consistent reporting.

Lorenzen points to the absence of common standards as a reason Salesforce has advocated for the U.S. AI Environmental Impacts Act. As described in the interview, the proposal would establish a consortium to develop guidelines for measuring and disclosing AI’s environmental impact. The core issue for companies is practical: comparable information can help buyers identify the energy implications of a service and give providers an incentive to report them clearly.

Efficiency can help the budget, too

Environmental performance and operating economics often point in the same direction, according to Gamazaychikov. Smaller, more efficient models use less energy and can reduce computational costs. They may also respond faster, which can improve the experience of using an AI feature. These benefits make model sizing and efficient infrastructure relevant even for organizations whose main concern is product performance or spending.

There can still be trade-offs. The carbon intensity of electricity varies by region, so a data center with lower-cost power may not have the lowest-emissions supply. Organizations choosing where and how to run AI need to consider the full set of operating factors rather than treating energy consumption as a standalone figure.

Lorenzen also emphasizes the wider energy system. Expanding transmission capacity and easing grid congestion can help bring more low-cost renewable electricity to customers, while supporting lower emissions. That work extends beyond decisions made by a single AI team, but it shapes the power available to data centers and the cost of supplying it.

Climate uses extend beyond technology companies

Lorenzen sees the response as a shared effort across industries. Businesses can ask AI providers for transparent environmental information and develop applications that support sustainability goals. Salesforce recently completed its first AI Accelerator, which supported nonprofit organizations using AI to improve efficiency and climate action.

Salesforce’s own climate work includes maintaining a global commitment to 100% renewable energy and pursuing broader power sector decarbonization. Lorenzen describes the work as extending beyond purchasing renewable electricity to considering what is needed to decarbonize the grid. Her role includes coordinating the company’s energy transition efforts and supporting teams working on measures that can enable further progress.

What organizations can track

For teams adopting AI, the interview points to several concrete questions: Is the model sized for the task? What hardware and data center will run it? How energy-efficient is the service, and what emissions are associated with its electricity? Can the provider report those figures using a method that allows comparison?

Those questions matter because AI’s environmental footprint is shaped by deployment choices as well as by technical progress. More efficient GPUs can reduce energy per unit of computing, but rising use can increase total demand. Smaller models may lower cost and latency, while the location and electricity mix determine how much carbon their energy produces.

Salesforce’s case is that sustainability and performance can often improve together, especially when organizations avoid oversized models and choose efficient infrastructure. Wider adoption of consistent reporting would make those decisions easier to evaluate. As AI use expands, the useful measure will be whether efficiency gains and cleaner electricity keep pace with the computing demand they help make possible.