Leading hospitals are building serious AI compute of their own. In 2025 Mayo Clinic deployed a Blackwell B200 system to power its generative-AI and digital-pathology programs; the system trains foundation models on Mayo’s archive of some 20 million digitized pathology slides paired with millions of patient records. Its Atlas pathology foundation model, developed with Aignostics, was trained on more than 1.2 million whole-slide images, and Mayo says the Blackwell infrastructure can compress four weeks of analysis into one — with, as its own leaders stress, the physician still in the loop making the final call. Pharma is doing the same at scale: Eli Lilly is among the enterprises building large on-premises AI infrastructure for drug discovery.
The clinical AI stack around this compute is maturing fast. Ambient clinical documentation — Abridge, Microsoft’s Dragon Copilot / Nuance DAX, and Ambience — now drafts notes from doctor-patient conversations across many health systems, saving clinician time. FDA has authorized a large and growing number of AI-enabled medical devices, the bulk of them in radiology, and hospitals are piloting clinical large language models for summarization, triage, and decision support. Genomics, imaging, and pathology are all being reshaped by models that demand both heavy training compute and fast, private inference.
But medical AI runs headlong into a hard constraint: the data is the most regulated and most sensitive in the enterprise. HIPAA and its international equivalents make every transfer of protected health information a liability, and patients increasingly expect their records not to be training fodder for someone else’s model. Cloud AI can be made compliant, but it multiplies the number of parties touching the data and puts the control loop outside the institution’s walls.
This is why data sovereignty has become a board-level issue in healthcare, and why the on-premises argument is winning for the most sensitive workloads. When the model trains and runs on hardware the hospital owns, protected data never has to leave the building; latency for real-time clinical use is minimized; auditability is straightforward because there is one physical system to inspect; and the institution keeps the power switch. For a hospital, owning the infrastructure is not a niche preference but the option that best matches the obligations it already carries.
Safeera’s Centurion packages that argument into a single object: a roughly 100–120 kW self-contained AI appliance — GPUs, networking, power distribution, and liquid cooling in one enclosure — that installs on a hospital roof or stands on the grounds, needing no dedicated data-center room. Shipping on B300 Blackwell Ultra GPUs with an announced upgrade path to Vera Rubin, it gives a hospital its own private clinic inside the unit: a place to train pathology and imaging models, run ambient documentation and clinical LLMs, and keep custody of patient data, the model, and the off switch.