Closing the divide in access to care
Did you know that half of the world's population – some 4.5 billion people – still lack access to essential health services? Although spending on healthcare takes up 10% of global GDP, there are only 17.2 doctors per 10,000 people globally, and in the hardest hit of low- and middle-income countries, there is fewer than one trained surgeon per 100,000 people – or ten for every million inhabitants. Globally, the World Health Organization estimates a shortfall of 11 million health workers by 2030. These figures tell a story of barriers to healthcare, especially in remote areas. And of barriers to training the next generation of doctors and surgeons.
But there’s hope: Investments in generative AI alone specifically for healthcare use cases are forecast to grow from $2.79 billion in 2025 to $50.79 billion in 2035. And with such a boost, AI can start to level the playing field.
Late last year, I had the opportunity to speak in Lisbon on this, a topic very close to my heart. At the second edition of the Atlantic Convergence – bringing together representatives of digital infrastructure, business, and public policy from all corners of the Atlantic – I spoke of the exciting developments in the realm of medical AI. A topic which demonstrates the potential of AI and the criticality of digital infrastructure for solving some of our planet’s most pressing problems. It is my hope that 2026 will bring greater access to and more general roll-out of these innovations. Innovations designed to protect the most precious asset each of us has: our life.
How AI is breaking down the access barriers to healthcare
AI cannot replace doctors and surgeons, but it can complement existing health services and broaden their geographical scope. From training doctors to performing remote consultations, and on to robot-assisted remote surgery, AI not only speeds up diagnosis, treatment, and healing, but also makes these accessible to people well outside the normal catchment area of centralized urban health services.
AI-supported Virtual and Augmented Reality (VR/AR) is one example: It can give students access to immersive training environments. For inexperienced doctors in rural settings, it provides remote access to AI support and mentoring from specialists. These technologies can also be used to provide virtual consultations with distant patients, as well as contributing to remote diagnostics and patient monitoring.
And then there’s robots. Surgical robots that operate on a patient when controlled from afar by a surgeon – a thousand kilometers away, or even halfway around the world. These have been used with remarkable success in cardiology, tumor removal (including brain tumors), joint replacements, and gynecological procedures.
But breaking down barriers doesn’t end there: AI is also accelerating new drug discovery and new treatment development, leading to breakthrough cancer therapies and novel treatments for degenerative conditions – faster and more efficiently. I truly believe that, with the help of AI-supported healthcare and medical research, the world will finally take a big step towards eliminating scourges like cancer and Alzheimer’s disease – and reduce suffering tremendously.
Training, supporting, discovering, and innovating: AI is opening the way to more options and greater access to medical services the world over. And the need for broad rollout of such innovations and systems is urgent: It is estimated that around the world 300 million surgical procedures are performed every year, but that another 143 million are required annually. So, more nurses, doctors, and surgeons need to be trained than ever before. More patients need to be seen, more afflictions cured. But to achieve this, AI needs our support. That is, the support of the infrastructure community. Because without infrastructure, without the flowing, exchanging, and processing of data, there can be no AI. Digital infrastructure will be central to beating cancer.
What infrastructure is needed for AI training?
Let’s have a look at what the infrastructure requirements are for AI in the medical industry, starting with AI training. We are seeing a proliferation of medical foundation models emerging. These, often based on general foundation models, are pre-trained on medical data – from biomedical research articles to clinical notes, and on to medical images – with a specific use case in mind. And once pre-trained, they are still not ready for use: They still need to be fine-tuned for the individual tasks they are destined to carry out and periodically re-trained to ensure knowledge of the most recent relevant medical insights.
Let me give you a couple of examples. Firstly, BERT is a large language model (LLM) which was originally trained by Google on a massive 3.3 billion words. This foundation model was then used as a basis for training medical models like BioBERT (for biomedical research & clinical data extraction) ClinicalBERT (for analyzing electronic health records (EHRs), clinical decision support, patient data management, patient outcome prediction), and MedBERT (for structured data analysis in EHRs, predictive modeling, clinical data mining, treatment response analysis, and optimizing treatment plans). Another example is SAM, the generalized image segmentation model, trained on one billion masks spread over 11 million carefully curated images. The pre-trained MedSAM is used for image diagnosis, personalized treatment plans, and efficient monitoring of diseases.
AI models can be trained anywhere – as long as there are enough processing units available and electricity to power them. Very often, they are trained in the cloud, in massive centralized hyperscaler data centers. Another option is emerging currently with the evolution of connectivity technology: creating distributed GPU as a Service communities. Companies that don’t want to depend too heavily on one single vendor can then spread the risk, with smaller GPU clusters in multiple data center facilities spread across a city, all interconnected seamlessly using an AI Exchange (AI-IX).
But wherever a company chooses to train and fine-tune their model, they need access to data, and that data can be, frankly, everywhere: in clouds, on-premise, housed by a partner company, or coming in from IoT devices or customers. And this data needs to be transferred to the location of the AI hardware. This will require a sophisticated multi-cloud and hybrid cloud strategy – with scalable cloud routing technology – to source the most up-to-date data. Additionally, it needs high bandwidth, optimized and secure connectivity between clouds and other infrastructure to move the data. Because a network that is not dimensioned for the immense amount of data needed for AI training will become a bottleneck. Here, highly scalable, virtualized cloud and AI routing technology provides a part of the solution.
What are the infrastructure requirements for AI inference?
Although a large amount of AI hardware is needed for training purposes, it isn’t needed for very long. In fact, it is said that in the coming year, 70% of AI processing will be inference, not training. To see how the magic of AI inference unfolds, let’s now look in more detail at some of the use cases I mentioned above. AI inference also requires access to computing power, but not as much as training. More importantly, the processing of inference needs to happen as close to the location of the use case as possible, in order to reduce the latency (delay) between the AI and its user. Unlike training, AI inference can even be carried out in edge data centers as long as these are well connected to the overall ecosystem. Inference needs access to the clouds and data centers where the trained model is stored for the long term. It also needs access to data: data about the patient’s medical background, data about treatment options, all the way through to data about the condition of networks through which critical medical data should travel and roads along which patients will be transported. And it needs access to users – be they people or IoT devices.
To enable access to data and devices in the lowest latency possible, cloud and AI inference as a Service routing is a central component of AI infrastructure, as is the interconnection of different network technologies, including fiber, 5G, and LEO satellite networks. Network stability and resilience – achieved through redundancy and optimized interconnection – are also essential, especially for critical use cases. All of this combined is possible through DE-CIX’s AI-IX concept – which brings together the interconnection technologies that allow seamless AI inference. In addition, enabling the exchange of highly sensitive medical data also demands the highest level of security of the dataflows, for example through creating a private interconnection environment, a Closed User Group, on the interconnection platform.
AI inference in the medical sector – training doctors and saving lives
For the training of surgeons in remote locations, latency is an especially important part of the equation. In an immersive environment for live mentoring or AI support, the latency of VR/AR data, for example, should not exceed 20 milliseconds (ms), or the surgeon wielding the scalpel may experience disabling disorientation. 20 ms translates into a distance of no more than 1,800 km that the data should travel – and then, only when the networks are fully optimized and interconnected in the most efficient manner possible. However, with the appropriate network technology in place, such remote mentoring and AI support has shown success rates equal to those of an experienced surgeon carrying out the same procedure.
Remarkably, remote surgery has been carried out over extremely long distances, across state and national borders, and even between continents. One real-life example was performed in 2025 between a surgeon in Florida and a patient in Angola, in which a cancerous prostate was successfully removed. The connectivity used existing network infrastructure – 11,000 kilometers of fiber-optic cable – with adaptive network optimization to ensure latency across the network of 140 ms. Such procedures are at the cutting edge of medical innovation, used for planned procedures on non-critical patients and with AI-supported adaptive control of the robot to ensure precision and minimize risks, despite the significant latency. However, for immediately life-threatening situations with critically injured patients, a solution like this is not likely to suffice.
What is known as “haptic-enabled force feedback” telesurgery is a different story: It allows the surgeon to truly feel what the robot is doing. This is a much more latency sensitive endeavor, but the results speak for themselves: Force feedback technology offers greater precision and shorter operating times than non-haptic remote surgery. It allows surgeons to deal in real time with complications such as needing to cut through scar tissue, and with its intelligent support is as good as – or even better than – the surgeon operating in person. Studies have shown a 25% reduction in operative time, a 30% decrease in intraoperative complications, and an increase in surgical precision of 40% for telesurgery. However, such technology works ideally at less than 10 ms latency, translating to no more than 1000 km distance.
These wonderful technologies can be bundled and mobilized in the form of the 5G connected ambulance. Mini surgeries on wheels, these are even more latency sensitive, operating at around 3 ms, within a radius of 250-300 km max. from the hospital and AI infrastructure. Connected ambulances provide AI-powered support for paramedics. The ambulance can connect to patient wearables to monitor condition even before the ambulance arrives on site, allowing paramedics to prepare effectively. High-resolution imaging and AR links support with rapid diagnosis, while haptic-enabled robotic arms bring remote surgery to accident sites, preparing critically ill patients for transport to the hospital more efficiently than a mobile doctor. As part of a smart city ecosystem, the ambulances can then be routed using intelligent traffic control systems to ensure the fastest route to the hospital. The results: a 30% improvement in survival rates for critical patients, a 40% reduction in certain medical errors by paramedics, 20 minutes faster time to definitive treatment, and a 50% reduction in response times in congested urban areas.
These are exciting use cases, and we await the broader roll-out of the technologies that will enable widespread adoption in 2026. The need is urgent. Because, according to the World Health Organization, between 5.7 and 8.4 million deaths are attributed to poor quality care each year in low- and middle-income countries, representing as much as 15% of overall deaths in these countries. Especially with the growing use of AI agents expected in 2026, AI-powered innovations in healthcare can turn the tide, enabling access to healthcare and life-saving procedures, wherever people are. But only with data centers, networks, and interconnection technology closely interwoven to enable high performance training and low latency inference. Only then will AI be able to solve any of our planet’s most pressing problems.
