Revmatics: Green AI Doesn’t Need Bigger Models — It Needs Smarter Systems

One of the most commonly cited statistics surrounding AI and data centers is that data centers consume 17 billion gallons of water per year. However, scientists at Lawrence Berkeley National Laboratory puts the real bill on the total amount of water used by data centers closer to 228 billion when also accounting for electricity generation costs.

Not surprisingly, this has led to increased demand for AI that makes environmental responsibility a priority. According to Justin Wilde, co-founder, COO and CPO of AI-powered marketing platform Revmatics, James Myrick, co-founder and CCO of Revmatics and Travis Simon, CMO of Revmatics, achieving truly “green” AI requires a focus on smarter (rather than bigger) models. By building energy optimization into AI models from the start, energy use can change for the better.

A big part of the current AI energy challenges stems from the fact that builds have typically made efficiency an afterthought. As Revmatics CEO Ricky Ray Butler explains, “The problem with the industry’s obsession with brute-force AI is that it has allowed for lazy architectures and incredibly inefficient builds. When you can just throw endless computing power at a problem, you don’t actually have to build elegant software. True engineering means optimizing under constraints, not just throwing more coal on the fire to get a result.”

It’s not just the way models are designed. As Wilde explains, challenges also stem from the way users tend to use AI: “When humans manually pick AI models, they tend to default to what seems like the biggest and safest choice. They end up burning computing power ‘just in case,’ which results in a lot of unnecessary waste. There are usually more efficient models and ways of doing things, but people often don’t have the time or expertise to figure that out for themselves.”

To that end, Revmatics has gone the extra mile in taking steps to build marketing AI modeling systems that are smarter, not just bigger. In fact, the company has gone as far as patenting its own “lean pathway” that is geared toward improving efficiency. And when it comes to energy use, this focus on efficiency has proven incredibly impactful.

According to Wilde, Revmatics’ Lumara AI tool was designed from the ground up to manage its own environmental footprint. “Lumara was built so it would treat every token as a cost that needs to be justified. As a Level IV AI system, it’s designed to be self-improving, and this applies to environmental use as well. The system maximizes the value it produces in line with the energy it uses. For example, it routes simple tasks to lightweight models and only uses deep reasoning for complex tasks, so energy use is always mapped to actual computing needs.”

As Simon notes, “In production, Lumara’s per-step model selection reduces token consumption by 20 to 50 percent per workflow. We lead with tokens deliberately. Tokens are the unit a customer can actually audit on an invoice. Energy claims are nearly impossible to verify from the outside, so we’re careful about where the savings land. When Lumara routes a task to a smaller model, the reduction happens at the model provider's infrastructure, not at a specific data center we can point to on a map. A lot of vendors blur that line because the blurred version sounds better. We think the precise version ages better.”

This provides a contrarian approach to the standard practice of simply building bigger and bigger AI models that require ever-more computing power, and helps reduce the AI energy costs associated with fracking water and electricity. What Revmatics has put in practice has been backed by others. For example, an analysis from Mahmut Kandemir at Penn State University noted that developing domain-specific AI models can vastly reduce computational requirements in comparison to using large general-purpose models.

Myrick adds, “Revmatics was built to create automation with intelligence. What we’ve found is that this naturally overlaps with efficiency. You do not have one without the other. Well-engineered models and systems outcompete and outperform just throwing expensive models at a problem. So while gains are made quickly with brute force approaches, the long-term winners in both results and efficiency are the ones who build on infrastructure that has been purpose-built for the tasks at hand.”

The company’s own case studies with Lumara’s more energy-efficient approach reveal that this leaner approach to AI doesn’t compromise performance. “We’ve had clients improve their marketing KPIs by 30% to 50%, so we can see that sustainability doesn’t mean giving up performance,” Wilde says.

“Long-term, this is the right approach. Especially as Lumara reaches thousands of brands and millions of products, we’re able to shrink the environmental impact of AI in marketing at scale. It’s not about just one campaign. It’s about seriously moving the baseline for our industry’s AI environmental impact.”

For Simon, traceability and accountability are especially important in understanding whether an AI solution is truly delivering on its “green” promises. “The fastest way to lose trust in this category is publishing a gallons-saved number nobody can trace. Our rule is simple: if we publish a figure, we publish the methodology and assumptions with it. If we can’t, the figure doesn’t ship.”

The Revmatics team recognized the unsustainability of brute-force AI from the company’s founding. By building a self-improving AI focused on resource management from the start, the company has set an example of how green AI can become a true possibility.

“AI, when done right, is fundamentally an efficiency science,” Butler concludes.

“In marketing, you want to reduce how much waste in media production and spending to reach your customers. With AI, the same approach holds true. The smarter and more efficient the AI, the less energy and computing power we waste to get the answers we need. By building smarter architectures that balance model usage, we don’t just get better results for the problems we’re trying to solve with AI. We can also make sure our AI systems are working in the greenest way possible.”

As the world becomes ever more dependent on AI in practically all facets of life, this type of approach will prove ever more important.