Many of today’s artificial intelligence, or AI, systems spend their time repeating themselves.
Every time an AI system analyzes data to generate a prediction, recommendation or decision, it must move information between databases and machine learning software. That process is often surprisingly inefficient. Across large-scale online systems, the wasted computation adds up fast. In many companies, the biggest expense is no longer training AI models but continuously running them on streams of incoming data — a process researchers say can account for up to 90% of machine learning costs.
Jia Zou is an assistant professor of computer science and engineering in the School of Computing and Augmented Intelligence, part of the Ira A. Fulton Schools of Engineering at Arizona State University. While much of the AI industry focuses on building larger and more powerful models, Zou is trying to solve a different problem. Her research focuses on the enormous amount of wasted computation hidden inside modern AI systems.
“There’s a huge opportunity right now to make AI systems more efficient,” Zou says. “By eliminating redundant computations and data transfers, while intelligently scheduling co-optimization techniques, we can reduce costs and dramatically improve response times.”
For the past several years, funded by the U.S. National Science Foundation, Zou and her collaborators have been building systems that allow databases and AI models to work together more efficiently. Instead of constantly moving information back and forth between separate systems, their projects, InferF and CactusDB, keep data processing and AI analysis in the same place, reducing wasted computation and speeding up responses.
Teaching AI systems to stop wasting work
Much of that inefficiency comes from the way modern AI systems are built. In most companies, data storage and AI processing still happen in separate systems. Information must be retrieved from databases, transferred into machine learning software, analyzed and then sent back again, a process that consumes time, energy and computing power.
“Jumping between the database system and the machine learning system creates a lot of problems,” Zou says.
Those problems include privacy risks, data transfer costs and sluggish response times. But they also create wasted computation.
That inefficiency is the focus of InferF, a new system developed by Zou and her team that was recently presented in June at the 2026 ACM SIGMOD/PODS Conference in Bengaluru, India, one of the field’s leading conferences for database research.
The project tackles a common problem in modern AI systems. Companies often combine information from multiple databases before sending it to AI models for analysis. But that process can unintentionally duplicate huge amounts of data, forcing AI systems to perform the same computations repeatedly.
In the paper submitted to the conference, the researchers cited an example involving the Internet Movie Database, or IMDb. The platform’s original structured data requires only about 1.2 gigabytes of storage. But after the data is combined through a series of operations, the resulting dataset can swell to more than a terabyte.
That explosion consumes storage space and generates computational waste. InferF attempts to solve this problem by breaking AI workloads into smaller pieces that can run closer to where the data already lives, reducing repeated computation and unnecessary data movement. According to the researchers, the system was up to 11.3 times faster than existing approaches.
From board games to big data
To solve that problem, CactusDB borrows ideas from the same family of search algorithms that helped power AlphaGo, the AI gaming system that defeated world champion Go player Lee Sedol in 2016.
Today, most companies process data and run AI models in separate systems. A database retrieves the information, then another system runs the AI model, and the results are passed back again. CactusDB combines those steps into a single system that plans them together.
When a request comes in, the software evaluates different ways of organizing the task before it runs. It determines where computations should happen, how data should move through the system and which operations can be combined or avoided entirely. The goal is to reduce unnecessary computation and minimize the amount of data being transferred between systems.
To help make those decisions efficiently, CactusDB uses a technique called the Monte Carlo tree search, which allows the system to quickly identify promising execution strategies without exhaustively testing every possibility.
“So, database systems originally just manage data,” Zou says. “But now in our system, databases can also be used to manage models.”
In benchmark tests, the researchers reported speed improvements of up to 441 times faster than alternative systems.
The approach could become increasingly important as AI systems move into environments where computing power, energy and internet connectivity are limited.
Beyond bigger models
Zou’s team is already exploring applications with medical researchers and civil engineers. One collaboration involves helping clinicians quickly find and synthesize information from large collections of medical records. Another focuses on helping engineers analyze data from smart buildings, where AI tools could help identify patterns in energy use and building performance.
“In these scenarios, there is a huge amount of data to process, but people still need answers quickly,” Zou says. “Our goal is to make these systems faster and more efficient.”
As AI use expands, the cost of wasted computation is no longer only a technical problem. These systems require enormous amounts of electricity, cooling and infrastructure. Recent ASU research found that heat from large data centers can raise temperatures in nearby neighborhoods by as much as 4 degrees Fahrenheit.
If AI is going to scale into hospitals, buildings, financial systems and everyday services, simply building bigger models may not be enough. The future of machine learning may also depend on teaching those systems how to waste less.
