The fastest-growing cost in artificial intelligence may not be chips — it may be the electricity that feeds them. On October 12, the University of Michigan announced that NVIDIA, the world's dominant AI computing company, is evaluating how to add two U-M energy-saving tools to Megatron-LM, NVIDIA's widely used open-source platform for training large language models. According to the university, the software can cut the energy required to train AI by up to 50% — without impacting training speed or model performance.
The news matters well beyond one lab's publication record. Megatron-LM is not a niche research codebase: companies can use it directly or as the foundation for their own private, customized training platforms, and it is the platform used to train NVIDIA's flagship Nemotron models, which users can download onto their own hardware. If the U-M tools make it into Megatron-LM, efficiency improvements developed in an academic lab would propagate into a large share of the world's AI training runs. For ongoing coverage of the infrastructure race behind AI, follow the latest AI developments on our homepage.
What Kareus and Perseus Actually Do
The two tools attack different sources of waste in distributed training, and their savings stack.
Perseus: Managing Processor Speed
Perseus addresses over-provisioned compute. In a large training run, not every processor is doing useful work at every moment, yet all of them draw power. By managing processor speed to match the actual work available at each point in training, Perseus cuts the energy needed to train AI by up to 30%, according to U-M.
Kareus: Smarter Timing for Compute and Communication
Kareus targets a subtler problem. Training modern models requires coordinating hundreds to thousands of processors, and the work splits between computation and chip-to-chip communication. Kareus reduces training energy by an additional 15% to 30% by finding the most efficient timing and resource sharing between those training and communication tasks — scheduling each so that neither sits idle waiting for the other.
Together, the university says, the tools can save around half of the energy used to train AI while delivering the same training speed and final model performance. Kareus is also a peer-reviewed result: the tool was presented at the USENIX Symposium on Operating Systems Design and Implementation (OSDI) in July, one of the top venues for systems research.
Why NVIDIA's Interest Is the Real Story
Academic efficiency research usually faces a long road to production. This one may be different, because of where it landed. A doctoral student, Ruofan Wu, spent this past summer at NVIDIA integrating the U-M software into Megatron-LM as part of an internship — meaning the tools are not a proposal but working code inside NVIDIA's training stack, now under formal evaluation.
"AI is becoming power-constrained as models become larger and more people want to train models, but we can achieve the same training speed and performance using less energy," said Wu, a doctoral student in computer science and engineering at U-M, in the university's announcement. "There are a lot of opportunities to optimize AI energy use at the software and hardware level, and one of the most rewarding parts of the internship was the opportunity to bring ideas from my research into production systems," she added.
That word — power-constrained — is the industry's new watchword. The U-M research team, which also includes doctoral students Jeff Ma and Jae-Won Chung along with associate professor Mosharaf Chowdhury, is working with NVIDIA on the integration. The team's framing is straightforward: the goal is to let AI do more useful work for less energy and cost by finding the most efficient way to coordinate training across hundreds to thousands of processors.
Energy Is Now a First-Class Constraint
The scale of frontier training is what turns percentage savings into strategic numbers. Meta said this week that its new Muse Wonder models were trained at Prometheus, which it describes as a gigawatt-scale data center. When a single training campaign runs at that scale, a percentage-level energy reduction is not a rounding error; it is the difference between fitting a training run inside a power budget or not.
The economics follow the physics. Energy has become one of the gating resources for AI expansion. Software that halves the electricity bill of a training run — using existing hardware — changes the calculus for every company training its own models, and particularly for smaller players for whom energy costs are the binding constraint.
It also reframes the environmental debate around AI. Much of the public discussion has focused on inference energy — the power consumed by every chatbot query. But training is where the largest single blocks of energy are concentrated, and a peer-reviewed, vendor-agnostic software approach that cuts it by up to half offers something rarer than a hardware roadmap: efficiency gains that can arrive with the next open-source release rather than the next data center build.
What Happens Next
The university says NVIDIA is evaluating how to add the software to its open-source training platform. If it ships in Megatron-LM, adoption could spread quickly given how widely the framework is used across the industry and in academia. The U-M team's broader research program continues to look for energy savings across the software and hardware stack — an agenda that seemed niche a few years ago and now looks like the center of the game.
