MIT unveils EnergAIzer, a fast way to estimate AI energy use in data centers

AI has become so power-hungry that estimating its energy costs is now an engineering problem in its own right. Researchers at MIT and the MIT-IBM Watson AI Lab have proposed EnergAIzer, a method that predicts within seconds how much energy a specific AI workload will consume on a chosen processor or accelerator. For data centers, that means making decisions before a model runs, rather than after wasting electricity.

The problem is no longer theoretical. According to Lawrence Berkeley National Laboratory, by 2028 data centers could consume up to 12% of all electricity in the US, and a noticeable share of that growth is tied to the development of AI. Operators have to split limited GPU resources among model training, inference and data preprocessing, but it is hard to know the energy price of each scenario in advance.

Classical methods usually break a workload into many separate operations and emulate the GPU's internal modules step by step. That can take hours or even days, which is a poor fit for fast engineering decisions.

EnergAIzer is built on a lighter model. The MIT team noticed that AI workloads often contain repeating computational patterns, especially when developers have already optimized code for the GPU: parallelized the computation, distributed data sensibly and tuned memory block transfers. Instead of a full, detailed simulation, the system uses these regular structures to quickly estimate the energy profile. The idea is to take less low-level information but extract enough from it for a reliable prediction.

The researchers did not stop at a rough estimate. They accounted for the fact that every run has a fixed energy cost for preparing and configuring the program, followed by additional costs for processing each block of data. There are also hardware factors: fluctuations in how the hardware behaves, memory access conflicts and incomplete use of bandwidth. To compensate for these effects, the team collected real measurements from GPUs and added correction coefficients. As a result, the method stayed fast but became noticeably more accurate than simple approximate models.

In practice, a user can give EnergAIzer the parameters of their workload: which model they want to run, how many input requests need processing and how long those requests are. The system returns an energy estimate within seconds. Users can also change the GPU configuration or operating frequency and see how that affects total energy use.

In tests on real AI workloads and real GPUs, the average error was about 8%, comparable to traditional approaches that take far more time. The authors also note that the method can be applied to future accelerator configurations, provided the hardware architecture does not change radically.

The work was presented on April 27, 2026, and the results are also being presented at the IEEE International Symposium on Performance Analysis of Systems and Software. The authors include MIT postdoc Kyungmi Lee, graduate student Zhiye Song, IBM Research researchers and senior author Anantha Chandrakasan, MIT's provost.

The next step is to test EnergAIzer on the newest GPU configurations and scale it to scenarios where several accelerators jointly process a single workload. That matters because large AI workloads are increasingly distributed across multiple GPUs.

The main takeaway is simple: AI energy efficiency depends not only on new chips but also on the ability to quickly measure the cost of computation before launch. If an energy estimate takes seconds rather than days, it can be built into the everyday work of data centers, model-training teams and algorithm developers. Energy consumption then stops being an after-the-fact metric in a report and becomes a real parameter for choosing architectures, tuning accelerators and planning compute.