NVIDIA survey finds AI moving from lab to clinic with real ROI

For years, healthcare treated AI the way an old-school doctor treats experimental medicine: interesting on paper, but way too risky for actual patients.

That hesitation is officially gone. NVIDIA just released its second annual State of AI in Healthcare and Life Sciences report, and the headline is simple: the industry has stopped tinkering and started deploying.

According to the survey, 70% of healthcare and life sciences organizations are now actively using AI, up from 63% in 2024. More importantly, it is actually paying off. A whopping 85% of executives said AI is driving revenue, 80% reported cost reductions, and 85% plan to increase their AI budgets this year, with nearly half bumping spending by more than 10%.

Imaging and drug discovery lead the money trail

Generative AI and large language models are now the primary workload across the industry, used by 69% of respondents compared to 54% previously. Data analytics and predictive tools took the next two spots. But the real surprise was agentic AI: in its first year in the survey, 47% of organizations said they are already deploying or evaluating AI agents, mostly to automate literature review and knowledge retrieval.

Adoption isn't uniform, but it is heavily concentrated where the financial return is clearest. Digital health providers lead overall adoption at 78%, followed by medical technology at 74%.

When it comes to return on investment, 57% of medtech respondents pointed to medical imaging, like algorithms flagging anomalies on scans so radiologists can move faster. In biotech and pharma, 46% cited drug discovery as their highest-ROI use case. Meanwhile, hospitals and insurers are finding quick wins in back-office grunt work, with 39% citing administrative workflow optimization as their biggest financial win.

As Dr. Annabelle Painter from Visiba U.K. put it, the organizations seeing real impact are the ones embedding AI directly into daily workflows, rather than slapping another standalone dashboard on top of clinicians' desks.

Open source for research, proprietary for the clinic

Healthcare is also leaning hard into open source. An eye-opening 82% of respondents called open source software and models moderately to extremely important to their AI strategy.

Running open models lets organizations tailor systems to narrow medical domains without sending proprietary patient data outside their walls. But as John Nosta, president of think tank NostaLab, noted, there is a clear boundary.

Open models are ideal for early discovery, research paper parsing, and pipeline flexibility. But once an algorithm touches clinical decision-making where liability, safety, and regulatory compliance are on the line, proprietary, tightly governed systems will remain the default. Basically: innovate out in the open, but keep clinical deployment under strict lock and key.

Why it matters

Healthcare is notoriously conservative, and for good reason: mistakes cost lives. Seeing 85% of executives confirm real revenue impact proves that healthcare AI has crossed the chasm from experimental PR stunt to defensive operational necessity.

For patients and doctors, the immediate upside over the next 12 to 18 months won't be an omniscient robot doctor. It will be the boring stuff getting radically faster: prior authorizations cleared in minutes instead of weeks, scheduling automated, and radiology queues moving significantly quicker. For AI founders, the bar has shifted. No hospital CEO is buying an AI wrapper anymore; if your tool doesn't integrate natively into an existing EHR or directly reduce administrative overhead, you won't even get a meeting.

The widespread embrace of open source is also a shot across the bow for frontier model providers. Healthcare enterprises want domain-specific control and private deployment, not dependency on external API black boxes. The real bottleneck ahead isn't model capability, it is liability: as agentic workflows take on more administrative and clinical tasks, the legal framework for when an agent hallucinates a billing code or misses a nodule is still virtually nonexistent.