Why can an organization invest in expensive AI infrastructure while researchers and employees still struggle to get the computing time they need?
In this episode of IT Infrastructure as a Conversation, I speak with Stephan Fabel, CEO of ThisWay Global, about the decisions between buying hardware and producing useful results. Our conversation connects GPU utilization, workload placement, sovereign AI, university research, and the responsibilities that come with building a data center in a local community.
Stephan describes two different problems that can disappear inside the same conversation about scale. One team may need thousands of GPUs to complete a demanding training job. Another institution may have many teams competing for a shared pool of computing resources. Software designed around a single large job does not automatically resolve the second problem. Research teams also build toward larger runs gradually, which can leave reserved capacity unused during development.
For organizations managing mixed hardware across several locations, the question becomes where each job should run and why. A deadline, budget, data location requirement, or need for an auditable record can change the answer. Stephan argues that workload placement should make those tradeoffs understandable. Starting a job immediately may produce some progress without improving its completion time, while preventing other teams from using the same resources.
We discuss how ThisWay Global approaches that problem through its Amalgamy software. The proposed benefit is matching jobs with suitable computing resources while respecting operational constraints. Stephan's account also raises a useful measurement question. Peak activity on a GPU tells us little about whether an institution is getting useful work from the capacity it pays for across the whole project.
The conversation then turns to the planned research and AI compute center at Texas A&M's RELLIS campus. Stephan describes an effort to bring industry partners, academia, and research projects together around access to accelerated computing. He points to possible applications in crop yields, energy distribution, grid management, and weather prediction. These are examples of the research he hopes access can support, rather than results established by the planned facility.
Sovereign AI introduces another set of responsibilities. For Stephan, control includes knowing where processing happens, retaining inspection rights, understanding who has access, and keeping a record of how data is handled. His description of being able to visit the infrastructure gives a physical meaning to a term often discussed through abstract policies. We also consider the operational relationship behind that control and the people needed to maintain it.
Community participation runs throughout the discussion. A data center can create demand for roads, services, and qualified workers. Stephan argues that operators should plan training and routes into technical work alongside the facility itself. His examples include vocational education, equipment operation, and opportunities for students to work with research infrastructure. These are proposals and priorities he describes, rather than guarantees of local jobs.
One story brings the access problem into focus. During an Uber journey in College Station, Stephan says a student began describing difficulties getting GPU time. The story illustrates his concern that access to computing resources can shape both research opportunities and preparation for employment. Hardware investment becomes a question about who can use the infrastructure, under what conditions, and to what end.
I'd welcome your thoughts. How should your organization balance utilization, operational control, and access when deciding where its next AI infrastructure investment belongs?