In the healthcare sector where precision and speed can directly impact patient outcomes, artificial intelligence has long been seen as a transformative force. Despite significant advances in models and algorithms, many AI initiatives still struggle to move beyond the pilot stage.
In an interview with iTNews Asia, Dr. Ilya Burkov, Global Head of Healthcare at NVIDIA-backed Nebius, shares insights on why many healthcare AI initiatives fail to scale and how infrastructure, often overlooked, is the critical missing link between promising prototypes and real-world clinical impact.
According to Burkov, a primary reason why AI projects stall in healthcare is not flawed models, but inadequate infrastructure. “The jump to production fails when systems designed for small-scale exploration are forced to handle real-world volatility. What works in controlled research settings often collapses under the demands of clinical environments,” he explained.
Early-stage AI projects can tolerate manual processes and limited datasets. However, real-world healthcare requires repeatability, parallel experimentation, and continuous validation capabilities that many existing infrastructures simply cannot support.
This mismatch, said Burkov, creates a bottleneck where innovation slows, not because the science is wrong, but because the systems cannot keep pace.
From sequential bottlenecks to parallel intelligence
Burkov pointed to GPU-native cloud architectures as a turning point in overcoming these limitations. “Medical teams historically constrained their research to fit computational limits, working with smaller datasets or accepting long processing delays. GPU-native systems eliminate this by enabling thousands of calculations simultaneously rather than sequentially.” he explained.
This shift allows hospitals to move beyond delayed, retrospective analysis. “Healthcare organisations can now generate insights from live patient data, transforming clinical decision-making from reactive to proactive,” he added.
Beyond speed: Enabling qualitative breakthroughs
Burkov emphasised that the value of advanced AI infrastructure is not just speed, but the ability to unlock entirely new categories of insight. “Infrastructure doesn’t just accelerate workflows, it changes what’s scientifically possible,” he said.
He pointed to epigenetics research, where trillions of chemical markers are analysed to understand how genes are regulated. With specialised GPU environments, researchers can train foundation models that reveal disease signals that are invisible to traditional approaches.
Similarly, in drug discovery, generative AI systems can analyse tens of millions of cells across thousands of genetic variables simultaneously. This enables precise identification of cellular failures, paving the way for highly personalised therapies.
The infrastructure gap
The impact of inadequate infrastructure becomes visible as datasets grow.
In drug discovery, researchers often reach a point where they can no longer test multiple hypotheses in parallel. What begins as rapid iteration turns into a sequential process dictated by hardware constraints.
- Dr. Ilya Burkov, Global Head of Healthcare at NVIDIA-backed Nebius
This forces teams into more conservative approaches, quietly slowing the pace of innovation. “The science remains sound, but the discovery process becomes limited by the system’s inability to scale,” he added.
As AI models grow in complexity, Burkov stressed the need for strategic infrastructure planning. He advised organisations to focus on value-driven deployment. “The most successful teams apply computational discipline, aligning resources with specific clinical outcomes rather than chasing model size,” he noted.
Balancing performance, cost, and security requires treating these factors as core design principles. When planned upfront, organisations can scale AI workloads predictably without triggering compliance risks or operational strain.
While hybrid and multi-cloud strategies are often promoted for flexibility, Burkov said their real-world implementation in healthcare is far from straightforward.
“The challenge is operational coordination. Without clear role definitions, hybrid environments can slow research instead of accelerating it,” he warned.
The overlooked readiness gap
Beyond infrastructure, organisations often underestimate the complexity of preparing data for AI. “High-volume data is useless without proper labelling and context. This requires specialised expertise that many teams lack,” Burkov noted.
He also highlighted the “assurance burden” - the need to continuously revalidate models as patient populations and clinical conditions evolve.
“Ultimately, success depends on workflow integration. Even the most accurate model will fail if it disrupts clinical practice,” he added.
Democratisation of healthcare AI
Advances in AI infrastructure are lowering barriers for smaller labs, enabling them to test ambitious ideas without owning large compute systems. This shift is reflected in funding trends, with AI-driven ventures accounting for a growing share of digital health investment across APAC. However, Burkov cautions against assuming a level playing field.
“Smaller teams are increasingly the source of breakthrough ideas, but large institutions still hold the advantage in clinical validation and deployment,” he added.
The result is an emerging ecosystem where innovation and scale are distributed across different players.
Where AI infrastructure will deliver the biggest gains
Over the next five years, Burkov expects the most significant advances in areas constrained by feedback speed. “Drug discovery will see immediate impact as infrastructure shortens the loop between hypothesis and validation,” he said.
Medical imaging is also set to evolve. “We’ll move from static snapshots to adaptive systems that update as clinical protocols change.”
Across healthcare, the defining factor will be the speed of learning. “The faster systems can learn from new data and be safely reviewed, the faster AI becomes an essential tool in care delivery,” he concluded.
