Universities have long been pivotal in fostering innovation, shaping careers and making invaluable contributions to global knowledge. The University of Edinburgh, with its world-renowned alumni, are joining other leading universities in supporting companies to develop the technology underpinning modern society. Net AI, a spin-out company based at the Bayes Centre, the University's AI and data science innovation hub, is one such case: a company born out of research at the intersection of 5G networks and artificial intelligence, now making technology that could help critical infrastructure systems function more cost-efficiently and reliably at scale. 

The company's origin stems from the research of one of the founders, Paul Patras, and a team he built with his research collaborators and former PhD students — a working relationship that gave Net AI its technical backbone, and part of why the company prioritises desk space at the Bayes Centre rather than setting up elsewhere. Paul still holds a position in the School of Informatics, and the proximity, he says, brings "a shared entrepreneurial culture that feeds motivation," even among founders working on entirely different problems. The research itself began when 5G was still in its infancy: the industry was eager about network slicing – the idea that operators could carve one physical network into several virtual ones, each tailored to a different application, rather than building separate infrastructure for every use case.

The difficulty was in effectively managing these slices –Paul realised, "you risk either over-provisioning – running more infrastructure than you need – or under-provisioning, which risks breaking service level agreements."

The fix meant giving operators real-time insights into traffic demand, service by service, so capacity could be configured dynamically. A patent followed, and Net AI was spun out — with backing, from its earliest round, from Old College Capital, the university's affiliated investment fund. 

The COVID-19 pandemic, oddly, helped the company formation: with investors sitting on time and money and unable to meet the founders in person, Net AI took dozens of meetings fast. Paul recalls people telling him "this is a genuinely good technology that will make a lot of sense" – but the market proved not to be ready. Most 5G rollouts still ran 4G cores, and few handsets could request a specific slice.

As Paul puts it, "we had a very powerful tool, but we were selling performance-enhancement drugs to people who actually needed painkillers."

The technical staff now sit as far afield as Spain, France, and the US — but the company has kept a single physical anchor throughout: desk space at the Bayes Centre, where the whole team still gathers twice a year for in-person all-hands meetings. 

While the war in Ukraine sent energy prices soaring, operators were already under pressure from growing demand for high-speed connectivity, customer churn, and shrinking profit margins. Energy costs are now outpacing sales growth by 50%, while customers increasingly expect reliable service across multiple connected devices. Mobile operators are trying to meet that demand efficiently enough to protect margins, without giving customers a reason to switch providers. That is where Net AI’s focus began to shift: from early questions around 5G network slicing to the more immediate problem of keeping infrastructure running cost-efficiently at scale. For Paul, the opportunity is both practical and commercial. Mobile networks account for two to three percent of global emissions, and the same efficiency gains that reduce that footprint can also help improve the economics of connectivity in harder-to-serve areas. “If you can play a part in that and make money doing it, why not?” he says. At the same time, he is careful not to overstate Net AI’s role: “We can’t move the needle on our own, we’re just trying to play our part.”

 That shift changed the shape of Net AI’s product: instead of building tools for a future 5G use-case, the company focused on helping operators run the infrastructure they already have more efficiently. A forecasting engine predicts traffic across thousands of antennas from a single model, deliberately biased against underestimating demand, since that "hurts service quality, and as a result customer retention suffers"; an anomaly-detection engine catches problems before they escalate; and through its ARIA-backed work, the team built synthetic network datasets and realistic test environments that allow AI-based energy-saving decisions to be evaluated safely before they are applied to live networks. Radio access networks account for 70 to 80% of a deployment's energy use, and Net AI's tools typically cut that by 10 to 35%, rising to 60% at individual antennas in lower-traffic areas. 

Paul credits his time in the 2021 cohort of the Bayes Centre's AI Accelerator, a programme that gives early-stage founders mentorship, access to office space, and investor connections, with teaching him fundamentals he hadn't previously needed —"what a cap table is, what cash flow is, what a term sheet is."

Even the hires beyond the technical team have come through university channels — the marketing lead was found through the student job board — proof, Paul says, that the connectivity between very different backgrounds at the Bayes Centre "works well," and that the institution's contribution to a spin-out isn't limited to engineers. 

Looking ahead, Net AI's near-term plan is straightforward — closing a funding new round, building a commercial function, and eventually, in Paul's words, scaling toward an acquisition by "a large vendor or a large system integrator. "Growth, for him, doesn't mean leaving Edinburgh behind: "scaling doesn't have to mean moving; we can find people outside the UK while still keeping our base here." The technology is reaching beyond telecoms, into satellite communication, data centres, financial backbone and energy distribution networks, wherever there is room to make critical infrastructure more efficient. Research that started in one narrow, technical corner has become something with wider consequences, still tethered to the Bayes Centre, to the place it began.