The Cost of Large-Scale Healthcare AI

Large-scale artificial intelligence in healthcare is attracting growing interest as generative AI and foundation models move closer to clinical use. A 2026 publication in The Lancet Digital Health sets out sustainability concerns linked to these technologies, particularly their dependence on novel algorithms, large-scale datasets and hyperscale compute resources. Generative AI can produce apparently new text, images or videos from text or image prompts, while foundation models aim to address a broad range of tasks with limited task-specific training. Both approaches usually require large models, even larger datasets, extensive storage and major compute capacity. These requirements increase energy consumption and carbon emissions and also raise economic and social questions for health systems seeking to use AI responsibly. The central challenge concerns how clinical value, patient outcomes and resource costs can be balanced from the outset.

Environmental Costs of Scaling AI

Large-scale AI development carries material demands that extend beyond model performance. Llama 3.1 405B, an open-source large language model with 405 billion trainable parameters, was trained on more than 15 trillion data points over 30.84 million graphics processing unit hours. Training energy consumption produced an aggregate carbon footprint of 8930 tonnes of CO₂ equivalents. Resource demands at this scale are common in the development of large-scale AI models.

Must Read: Green Dialysis: Pathways to Sustainable Kidney Care

Additional environmental pressures arise from water needed to cool data centres, extractive mining for hardware manufacture and electronic waste at the end of hardware life. These factors expand the environmental footprint beyond electricity use alone. Hardware procurement and energy consumption also carry substantial financial costs, which can place pressure on low-income and middle-income countries and affect their capacity to develop and use these technologies. At healthcare scale, such costs become part of the broader sustainability assessment rather than a separate technical issue.

Healthcare uses of large-scale AI remain at an early stage, yet sustainability concerns already cover environmental, economic and social costs. The issue is not only whether AI can support care delivery, but whether the scale of resources required aligns with the expected value in clinical settings.

Medical Imaging and Resource Demand

AI can accelerate clinical workflows by reducing tedious analytical and data-entry tasks, allowing healthcare professionals to focus more on patient care. Task-specific AI models for image classification and image segmentation already show promise, especially in medical image applications. Specialised image-segmentation models using convolutional neural networks have reached human expert-level accuracy in dosage planning during radiotherapy and national breast cancer screening.

Even relatively targeted medical imaging workflows remain resource-intensive because they process very large volumes of data. Public medical imaging datasets used in single experiments can reach a scale comparable to major computer vision datasets derived from extensive internet scraping. Storing vast quantities of medical images and running comparatively simple AI models therefore carries a sizeable underlying carbon footprint. Publicly available medical datasets include resources associated with imaging, cancer, brain MRI, osteoarthritis and Alzheimer’s disease neuroimaging, underlining the breadth of data involved.

The use of large-scale AI models on medical imaging data can compound overall emissions from AI in healthcare. Enthusiasm for generative AI and foundation models in clinical settings reflects their broader success in areas such as large language model-based chatbots. Caution remains necessary because these models carry substantial material costs, and environmental degradation from AI infrastructure can negatively affect public health.

Equity, Metrics and Governance

Large-scale healthcare AI raises questions about economic viability and social impact, especially for low-income and middle-income countries. Expensive infrastructure, including data centres, reliable energy grids and fresh water for cooling, means only a few countries develop most large-scale AI models. Reliance on already strained resources limits the participation of lower-resource settings in building AI and using it to meet local patient care needs.

If large regions cannot build or use large-scale AI models for their own needs, existing biases in machine-learning systems could aggravate health inequities. Models developed mainly for and with data from the Global North may not address local requirements elsewhere. The additional value of large-scale AI in healthcare also remains unclear when compared with other task-specific AI methods, supporting the case for restraint.

AI development should define desired patient outcomes and expected resource costs from the outset. Composite metrics can combine task-specific performance with environmental costs, so model selection reflects both efficacy and resource efficiency. In image segmentation, precision should not be the only driver; carbon footprint from development and deployment also matters. Small gains in efficacy may carry resource costs that lack clinical relevance. Quality-adjusted life-years and disability-adjusted life-years could incorporate environmental impact, with lifecycle impact-assessment frameworks such as ReCiPe2016 offering a starting point. Funding bodies could gauge project support against expected carbon footprint, while deployment-stage oversight could mandate monitoring and reporting of resource costs.

Sustainable healthcare AI requires attention to clinical value, resource use and access. Low-power embedded devices can reduce resource costs and support deployment in remote regions where local hardware is limited or cloud connectivity is unavailable. Efficiency improvements in algorithms and hardware may lower the resources needed for some large-scale models, but wider uptake can increase total consumption through the rebound effect. Transparent and inclusive decision-making, with organisations such as WHO engaging stakeholders, can help embed sustainability in AI development while preserving the focus on healthcare usefulness and equitable access. The same balance applies when scale, efficiency and uptake move together across health systems.

Source: The Lancet Digital Health

Image Credit: iStock