Instead of Models and GPUs, Focus on the Right Workloads for AI

HEALTHTECH: What sort of assumptions do healthcare leaders have when they discuss their AI infrastructure concerns with you? What conversations would you like to have?

FOSTER: I want to talk to healthcare leaders about workloads, not just the chatbots or infrastructure. Most of the questions have been about where various AI technologies are taking place — training, inferencing and data processing. And the challenging part about healthcare is that even with constraints such as performance, cost, security and compliance, you don't get to choose which aspects of AI will work; they all have to work. It's not optional in healthcare. So, bringing all that together is really the conversation we want to have. This is often looked at as a procurement decision, but it's really a placement decision in terms of, how do these things come together at the organizations? What's the clinical workflow? Where is the data actually created, and where does it live? What are the costs to do things such as move data back and forth between different areas? Five years from now, I think people aren't going to be challenging you about what model it is you picked for a certain job.

WATCH: How did Children’s Hospital Colorado move from data silos to data-driven decisions?

HEALTHTECH: Many healthcare organizations work with legacy systems and may not necessarily be prepared for AI workloads. What does that transformation look like in practice for providers of different sizes?

FOSTER: It doesn't mean an entire forklift-type upgrade. A refresh cycle is going to be a great catalyst to start with a lot of these AI projects. “How do I leverage what I'm doing from a refresh cycle to help prove our ROI and build an incrementally better system that can take advantage of AI as it goes?” No one in healthcare is starting from zero. Everybody has infrastructure that's already been deployed, and when they replace things that are happening out at the branch, out at the edge, for example, those are going to be opportunities for them to improve operations and clinical workflow by leveraging AI.

HEALTHTECH: How do you address the potential tension between adhering to strict compliance and security measures with deploying flexible, scalable compute that modern AI workloads need, especially with IT leaders who are concerned about moving off-premises?

FOSTER: That tension is absolutely real, and it’s why questions about workload placement are imperative. There’s an underlying assumption that, to some extent, scale comes only from the cloud, and the cloud is a really important tool here. But at the same time, a lot of data that's generated at a hospital or a clinic stays local, so there's a high cost or operational expense to move all this data around. A cardiac MRI, for example, is imaging done locally that then gets processed in the cloud and returns a bit of data. It’s the kind of workload that works at the edge. And that’s what organizations need to consider when creating a solution for a workflow end to end, because you can’t govern your way out of a design flaw. Designing at the edge allows organizations to be more agile in the coming years.

HEALTHTECH: What do you think will be the top healthcare IT priorities for 2027?

FOSTER: I think the big shift in healthcare will be around AI doing more inferencing to support operations and leveraging agents. That will generate a tremendous amount of traffic on the network. Our data says agentic AI queries generate up to 25 times more network traffic than your average chatbot, and that changes the pattern of what that traffic looks like. Instead of spikes from a single request, AI agents are working 24/7, so it's a much higher level of bandwidth that will be consistently used up. For a healthcare organization distributed across several states, how it manages that WAN and all the connected pieces will be a top consideration. 

When it comes to compute, power and space, everyone can plan for and have access to GPUs, but how they pull together the architecture is what's going to ultimately make them successful. The winners in healthcare AI won't be the ones who buy the most powerful GPUs or leverage the smartest models, it will be the ones who build the right foundation, the right architecture that's ready to take on whatever workflow.

Over the past 10 to 15 years, people have talked about pushing toward edge infrastructure, but AI has now been a driving force in actually making it happen, because that’s where the data is being generated. That's where we're seeing companies trying to replatform their operations. For partners that have already been designing at the edge from a routing and switching perspective, we’re poised to take healthcare providers to the next level. Although many organizations may not think they need a GPU at the edge today, they expect to need one in the next decade, but they don’t yet have a plan in place. We want providers to start thinking about those infrastructure needs now so they don’t have to stall their strategic goals.