Makora receives $1.75 million to bring AI-efficiency software to data centers

Makora has received a $1.75 million award to put its AI-efficiency software to work with data center providers. The funding, reported by Cornell Chronicle on October 16, supports a practical test: whether automated software improvements can help providers get more work from the computing resources they already have, rather than relying on additional hardware.

The award comes from the STRIDE Ventures AI Efficiency Challenge, which is supported by the National Science Foundation. It funds commercial deployment of Makora’s technology. The startup was co-founded by Mohamed Abdelfattah, an assistant professor at Cornell Tech, and is based at The Bridge, a workspace for companies on the university’s New York City campus.

Optimizing the response, not training the model

Makora focuses on inference: running a trained AI model to answer a request. A question sent to an AI assistant or a request to generate code consumes computing resources as the model produces its response. Improving that process could let organizations serve more users with the same hardware, according to Cornell Tech’s account of the company’s work.

The company uses large language models to generate GPU kernels. These are specialized pieces of code that tell graphics processing units, or GPUs, how to perform computations. That puts AI to work on the software used to run AI: the generated code handles computing operations, rather than producing the answer that an end user sees.

Makora’s tools address different levels of that work. Its MakoraSwarm platform extends automation beyond individual kernels to performance optimization across the software system. MakoraInference, meanwhile, delivers optimized AI models to customers. Those are existing parts of the company’s approach; the newly reported award supports deploying the technology with data center providers.

My work at Cornell is to research the next frontier of efficient AI technologies, while my startup is about deploying efficient AI to customers in the real world.

Mohamed Abdelfattah, Makora co-founder and Cornell Tech assistant professor

A moving target for engineers

The engineering challenge is that an efficient configuration does not stay current indefinitely. New models, computing chips and algorithms change how software needs to run. Code optimized for one pairing of model and chip may need further adjustments for another. Each new combination adds work for the engineers responsible for deploying and maintaining those systems.

Abdelfattah describes efficiency optimization as a constant game, requiring models to be redeployed as those ingredients change. His argument for automation is about keeping up with the growing number of combinations, not just improving one model once. Makora aims to automate that recurring optimization work and extract higher performance across different AI deployments.

Deployment funding, not a measured result

Abdelfattah says greater efficiency could mean faster responses, lower costs and more hardware options. He links those potential benefits to making AI more widely available as demand grows and computing and energy resources remain limited. Cornell also identifies a possible energy benefit: getting more from existing resources could help limit the energy needed to meet that demand.

Those benefits remain goals for the funded work, not quantified results from it. Cornell’s announcement provides no measured speedup, cost reduction or energy saving from the deployments. Makora will use the projects to demonstrate how its automated approach improves AI-system performance in practice—the commercial counterpart to Abdelfattah’s research on future efficient AI systems.