Healthcare’s push into artificial intelligence has misplaced its focus. For years, discussions centered on selecting models and purchasing graphics processing units. Yet the core issue remains overlooked: whether organizations possess the foundational systems to operate AI effectively. Most providers lack infrastructure designed to support AI as a primary function rather than an auxiliary tool. The challenge extends beyond hardware to encompass seamless integration of local computing, network capacity, and security protocols.
Jeremy Support, senior vice president and general manager of compute at Cisco, highlights a critical question: ensuring the right workloads rely on AI. Healthcare leaders frequently approach AI through a procurement lens, debating model selection, training locations, and cost-performance trade-offs. However, AI in healthcare is not optional. It must function reliably across diagnostics for rare diseases, clinical documentation automation, and other critical applications. The primary concern should shift from model choice to deployment strategy: where data resides, how workflows integrate with existing systems, and the costs of data transfers between locations.
In five years, the specific AI model a hospital adopts will matter less than whether the system can handle its workload without failure.
Legacy systems hinder AI progress
Outdated infrastructure remains a major obstacle to AI adoption, but providers do not need a complete system overhaul. A targeted refresh—replacing aging hardware at branches or edge sites—can serve as a stepping stone. Hospitals already possess much of the necessary infrastructure; the task is optimizing upgrades to enhance operations incrementally.
Consider cardiac MRI scans. While imaging occurs locally, processing often shifts to the cloud, creating inefficiencies: high data transfer costs, delayed results, and heightened security risks from unnecessary exposure. The solution is not always increased cloud reliance but designing workloads for edge environments, processing data at its source to minimize transfers.
IT leaders often hesitate due to concerns about compliance and security, fearing that flexible compute models conflict with strict regulations. This tension shows why workload placement is essential. Cloud scaling is not the sole answer; local processing can sometimes be more cost-effective, faster, and secure. For instance, edge-based cardiac MRI analysis keeps data within the hospital’s firewall while maintaining performance.
Infrastructure planning must address current design flaws to avoid future stagnation.
AI agents will transform network requirements
The next generation of healthcare AI will not consist of chatbots but of agentic systems, autonomous tools operating continuously. Cisco’s data shows these agents generate up to 25 times more network traffic than traditional chatbots, shifting demand from sporadic spikes to steady, high-volume usage. For hospitals operating across multiple states, managing this traffic will hinge on network bandwidth, wide-area connectivity, and edge compute capacity. Success will depend not on GPU power alone but on architecture capable of handling unpredictable, sustained workloads.
The overlooked discussion on AI readiness
For example, a radiology department using AI for image analysis needs not only powerful hardware but also seamless integration with picture archiving systems. Without alignment, even advanced models risk becoming operational bottlenecks. The focus often defaults to procurement debates, cloud versus on-premises solutions, but the critical question is workload placement. Local cardiac MRI processing reduces latency and costs while maintaining compliance, whereas centralized server transfers introduce delays and security vulnerabilities. The shift toward edge computing involves optimizing workload locations rather than abandoning cloud services entirely.
Agentic AI and the impending bandwidth crisis
The healthcare AI environment will soon transition from occasional chatbot interactions to continuous agentic operations, autonomous systems running around the clock. For multi-state hospital networks, managing this traffic will require robust wide-area connectivity, edge compute resources, and efficient bandwidth use. The distinction between success and failure will not stem from GPU capabilities but from whether infrastructure can handle persistent, high-volume workloads without performance degradation.
