What an Enterprise Buyer Thinks About Nebius I recently had a really interesting conversation with an enterprise buyer and client of $NBIS, who leads data and AI platform architecture at a Fortune 500 medtech company and previously worked in a similar role at a Fortune 50 pharma. I found his perspective valuable because it offers a useful way to differentiate between the so-called “neoclouds.” Most of the market tends to focus on GPU supply, CapEx, power availability and how quickly companies like Nebius can bring new capacity online. From the enterprise-buyer side, though, he argued that one of the bigger constraints on AI adoption is often data gravity, not simply compute availability. Large regulated enterprises can’t just move AI workloads wherever GPUs happen to be cheapest. Their data is already deeply embedded within governed environments involving lineage, security, compliance and existing infrastructure. Moving that data, or changing where workloads run, can therefore be much more complicated than it may appear from the outside. I think this point is particularly relevant when comparing Nebius with other AI infrastructure providers. Over time, the competitive advantage may not simply come down to who can offer the cheapest GPUs, but who can bring scalable AI infrastructure closer to enterprise data while also meeting the security, governance and reliability requirements these customers need. He also raised another point that I think could become increasingly important: AI inference spending inside enterprise platforms can grow much faster than many teams initially expect. Tools like Snowflake Cortex and Databricks Mosaic make it extremely easy for companies to integrate AI into pipelines and applications. That simplicity is obviously part of their appeal, but it can also hide the underlying compute cost. Teams can scale usage quickly without necessarily seeing the full financial impact until much later. As inference becomes embedded across more workflows, dashboards and internal applications, those costs can compound quietly and eventually create what he described as a potential “cost wall” for large enterprises. That’s also where Token Factory becomes interesting. There are already several platforms competing to make inference easier and cheaper, but Nebius controls much more of the underlying infrastructure stack. If Token Factory can leverage that integration to provide lower and more predictable inference costs at scale, it could become one of the more differentiated parts of the ecosystem. In other words, Nebius’ long-term advantage extends well beyond simply owning GPUs. Its full-stack approach allows it to structurally lower inference costs while still meeting the reliability, security and enterprise requirements large customers need. That value proposition is increasingly valuable as AI workloads move from experimentation into large-scale production. It’s also worth putting this in the context of where AI infrastructure demand is coming from today. So far, much of the attention has been on frontier AI labs, hyperscalers and AI-native startups, where access to compute, speed of deployment and raw performance often matter more than some of the enterprise considerations discussed above. But over time, I think a much broader wave of demand will come from large enterprises moving AI workloads into production at scale. That’s when issues like the ones mentioned above should become increasingly important. In that sense, some of the factors that may not matter as much to the biggest buyers today could become much more relevant to the next phase of AI infrastructure demand. From the enterprise side, the questions are increasingly becoming: Where does the data live? How difficult is it to move? How well does the infrastructure integrate with the existing enterprise stack? And what happens to inference costs once usage scales materially? Those are very different competitive dynamics from the ones investors usually focus on when comparing neoclouds, and I think they’re worth paying attention to. Credit to @AI_Investor_42, who kindly gave me permission to share some of his thoughts. For transparency, he is also long $NBIS.