In June 2026, a team at Boston Children's Hospital solved 18 cases of children with rare diseases that had gone undiagnosed for years. They did it with the help of an AI system developed by OpenAI, and the findings were published in the scientific journal NEJM AI. This is neither an isolated case nor a futuristic scenario, it's a snapshot of a change already unfolding in hospitals around the world.
At Johns Hopkins, predictive models alert clinicians to sepsis risk before symptoms appear, cross-referencing medical history, genetics, and continuous monitoring data. In early 2026, Google launched MedGemma 1.5, a model capable of outperforming closed systems on medical licensing exams that hospitals can run on their own servers, without sending patient data to the cloud. In radiology, algorithms already automatically prioritize studies showing signs of bleeding, stroke, or tumors so specialists review them first.
The result: more than 900 medical AI devices and algorithms now hold FDA approval in the United States, and the figure in Europe (EMA) exceeds 200. In Latin America, recent surveys report that 81% of physicians already use some form of AI tool in their daily practice.
The part that doesn't make it into the picture: the data center
All that "instant" diagnosis comes with a physical cost rarely mentioned in the same conversation: electricity, chips, and cooling at a scale that is beginning to concern governments and power utilities alike.
According to the International Energy Agency, global electricity consumption by data centers was around 415 TWh in 2024 and could nearly double to 945 TWh by 2030, a figure already comparable to Japan's total electricity consumption. Goldman Sachs Research projects 165% growth in data center energy demand by that same date, driven almost entirely by generative AI. In the United States alone, data centers consumed 183 TWh in 2024, 4.4% of the country's total electricity, with estimates pointing to more than 400 TWh by 2030, and that´s before agent AI enters the equation.
A clinical diagnostic model doesn't run on a single query: it needs to be trained on massive volumes of images, records, and genomic data, and then respond in real time, 24 hours a day, to thousands of simultaneous requests across dozens of hospitals.
Training a large model can consume around 1,300 MWh. Every medical inference, reading an MRI, cross-referencing genetic variants, generating a report, runs on high-consumption GPUs that require constant cooling. And that cooling is often as energy-intensive as the computation itself.
What does this mean for a hospital that wants to seriously adopt diagnostic AI?
This is not a minor technical detail. Serious adoption of clinical AI requires answering four infrastructure questions:
Dedicated compute capacity: clinical-grade GPUs with sufficient redundancy to avoid interrupting emergency services.
Stable energy infrastructure: demand spikes that many hospital buildings, designed for a different era, were never sized to handle.
Efficient cooling: Germany already requires new data centers to meet a maximum PUE (Power Usage Effectiveness) of 1.2 from July 2026 onward.
Where the model lives: running AI locally, as MedGemma 1.5 enables, reduces cloud dependency and strengthens clinical data privacy, but shifts the responsibility and cost of maintaining that infrastructure to the hospital itself.
The figure that changes the calculation: small, specialized models
Not everything points to an energy dead end. UNESCO notes that smaller models, trained for a specific task rather than designed to be generalists, can reduce energy consumption by up to 90% without sacrificing clinical performance.
This hints at where the next generation of hospital AI may be headed: not one giant model doing everything, but smaller, specialized systems, one for imaging, one for genetic variants, one for triage, running efficiently and locally.
A concrete case: could Madrid handle it?
The Community of Madrid has nearly 8 million inhabitants, an aging population, a network of more than 35 public hospitals, and the Senior Plus program specifically for people over 70. Does it have the infrastructure to support diagnostic AI at scale?
What works in its favor. Madrid is already, by a wide margin, Spain's largest data center hub. It concentrates around 70% of the country's total installed capacity, with roughly 613 MW operational in 2026 and projections to multiply several times over by 2030, backed by more than €6 billion in announced investment. Add to that a regional health budget that in 2026 exceeded €11 billion for the first time, with funds already earmarked for telemedicine and healthcare digitalization.
What doesn't add up as well. That near-70% of Madrid's data capacity was not built for clinical diagnostics, it serves banks, telecoms, e-commerce, and increasingly, generic AI model training. Allocating a significant, stable share of that capacity at the availability and redundancy levels a hospital demands, where a system outage is not an inconvenience but a risk to a patient, would require a very different kind of resource negotiation than any corporate data center client makes today.
Moreover, the real bottleneck is no longer investment: it's the electrical connection permit. Securing 50–100 MW of dedicated, guaranteed capacity in the short term has become the chokepoint, even for projects that are already funded.
And an aging population doesn't reduce the compute load, it multiplies it. More comorbidities, more imaging studies, more continuous monitoring of chronic patients, more readmission and sepsis prediction algorithms running in parallel across thousands of patients over 70. That is precisely the profile of high-volume, time-sensitive, constant demand that consumes the most electricity and cooling.
The honest conclusion
No public study precisely calculates the exact MW that diagnostic AI deployed at scale across Madrid's entire hospital network would require. But with the numbers on the table, the reasonable answer is: technically feasible, economically demanding, and only workable if planned as critical infrastructure, not as just another cloud service.
The question the sector can no longer ignore
AI is already diagnosing rare diseases, prioritizing CT scans, and anticipating sepsis in real hospitals, today. That is no longer up for debate. What remains open is whether we are scaling energy infrastructure at the same pace we are scaling clinical expectations.
A hospital can have the best model in the world, but if it lacks the energy, hardware, and cooling to sustain it reliably around the clock, that model is useless in the emergency room at three in the morning.
The conversation about AI in healthcare cannot remain separate from the conversation about energy. Increasingly, they are the same conversation.