For a decade, NVIDIA GPUs have powered the AI boom. But Dr. Eric Balki, an early AI and Health informatics visionary, argues that the next phase of medicine, ranging from rural clinics to space-based data centres, may be powered by a different kind of architecture.
If you walk into any high-tech hospital or AI research lab today, you will likely find NVIDIA’s graphics processing units (GPUs) humming in the server room. They are the undisputed heavyweight champions of the artificial intelligence revolution. However, according to Dr. Eric Balki, CEO of TriPyramid Group and Honorary Research Fellow at Lancaster University, the industry is approaching a “hardware cliff.” As medical AI moves from experimental pilots to mass deployment, the priority is shifting from raw power to extreme efficiency; a shift that could open the door for Google’s Tensor Processing Units (TPUs).
Balki’s thesis is controversial but compelling: The future of healthcare isn’t just about training massive models; it’s about running them cheaply and reliably everywhere, including places where terrestrial infrastructure fails.
Why Dr. Balki’s Opinion Matters
Dr. Balki’s perspective is shaped by a unique hybrid career that spans hardcore computer engineering and frontline healthcare management, along with degrees in Physics, Finance, Management Science, Law, and Health.
His technical roots run deep. Long before the current AI hype cycle, Balki was architecting high-performance compute environments, including setting up a chip design supercomputer based on a Beowulf Cluster using Sun Microsystems hardware in the North East of England. This early work in distributed computing gives him a fundamental understanding of how computer architecture can make or break a system at scale.
Balki also developed “Alternodes,” an early natural-language programming model that powered autonomous web-based agents, foreshadowing the agentic AI systems we see today. Moreover, he tackled the challenge of managing large-scale academic-led commercial projects by establishing the Knowledge House Information System (KHIS), one of the largest databases dedicated to managing third-sector activities.
Crucially, however, Balki isn’t just a technologist. He also holds a PhD in Aging and AI, and has over a decade of experience running health centers.
“In front-line healthcare, we’re not just chasing speed,” Dr. Balki shares, drawing on his operational experience. “We’re chasing precision at scale, consistent performance across thousands of patients, and ensuring the quality-of-service provision remains high despite resource issues.”
The “Inference” Problem
To understand the shift, one must understand the difference between training an AI and using it. GPUs are incredibly good at training, such as teaching a model to recognize a tumor. But once the model is trained, it needs to be deployed to diagnose patients day in and day out. This is called inference.
For hospitals operating on tight margins, the cost of running these models (inference) is the real bottleneck. Balki argues that TPUs, which are custom-built specifically for the math of deep learning, offer a decisive advantage here:
- Cost Efficiency: They often deliver more processing power per watt of electricity.
- Predictability: They are designed to handle the “noisy” nature of clinical data—years of erratic scans, handwritten notes, and lab results, without the massive energy overhead of general-purpose GPUs.
The Ultimate Stress Test: Datacentres in Space and Globalized Healthcare
Perhaps the most radical part of Balki’s argument takes people off-planet. He envisions a future of space-based data centres that keep healthcare systems online even during natural disasters, wars, or cyberattacks on Earth and provide globalised AI-enabled access to healthcare using satellite technology, much like the GPS system.
However, in space, hardware faces a brutal enemy: Radiation.
High-radiation environments cause “bit flips”, random errors in computation that can be fatal in a medical context. Because TPUs have a more specialized, controlled architecture than general-purpose GPUs, Balki suggests they could be better suited for the fault-tolerant designs required in orbit.
“Space forces you to think differently. Radiation, thermal cycling, limited power… If healthcare is going to become truly global and always-on, we’ll need architecture that can work reliably regardless of the time it has spent in space,” adds Dr. Balki.
Why This Matters for Patients
This isn’t just a battle of billionaires and chip specs. If compute costs drop and resilience improves, the “democratization” of world-class healthcare becomes possible. Currently, top-tier AI diagnostics are often trapped in elite medical centers. Balki predicts that affordable, resilient inference chips could push these tools to the edge:
- Remote Clinics: Access to specialist-grade diagnostics in rural areas.
- Disaster Zones: AI triage centers that work even when the local power grid is down.
- Global Records: Patient data that travels securely across borders, crucial for refugees and migrants.
The Verdict
NVIDIA isn’t going anywhere. For training new models and handling complex, multi-purpose workloads, GPUs remain the gold standard. But the “NVIDIA Era” of total dominance may be evolving into a split ecosystem. As healthcare AI matures from a science experiment into a global utility, the hardware that powers it is getting a second look. If Balki is right, the chip that wins the future of medicine won’t just be the fastest—it will be the one that keeps running when the lights go out.
“All I can say is that the future is definitely bright for AI-enabled healthcare,” says Dr. Balki. “I am not sure how quickly we will get there, but space has an important role to play.”
