Summary
- The University of Granada is preparing two distinct AI computing deployments.
- Electrical supply, cooling, and technical areas at the Health Sciences Campus require modification.
- Final rack density, cooling architecture, procurement dates, and facility costs remain undisclosed.
The University of Granada is preparing its Health Sciences Campus for two artificial intelligence computing systems with a combined indicated value of €8.5 million, adding new electrical and thermal loads to an existing academic facility.
The University of Granada plans a system valued at approximately €5 million for AI training and research, alongside a €3.5 million platform linked to the European Eunomia programme for sovereign generative AI in public administration.
The two projects are moving through different funding and procurement routes. The larger installation remains dependent on institutional agreements and tendering, while the Eunomia system forms part of an established European programme. Both require modifications to the university’s electrical supply, cooling systems, and technical areas.
The facility has to arrive before the machine
Supercomputing projects are often described through processors, accelerators, and purchase values, although the enabling works can determine whether the hardware enters service on schedule. New systems may require higher-capacity switchgear, upgraded distribution, additional uninterruptible power, cooling plant, network fabric, controls, and structural changes.
Placing both machines within the same campus environment requires the university to consider their combined demand rather than assessing each rack separately. Shared cooling, electrical distribution, or network equipment can create common failure points even when the computing systems are procured under different projects.
AI training produces a sustained and concentrated load. Accelerator nodes may run close to full utilisation for long periods, placing more heat into fewer racks than conventional academic clusters. An existing machine room designed around lower-density CPU systems may have adequate building power but insufficient distribution or cooling at the rack.
The final hardware configuration will decide whether room-based air cooling can carry the load. Higher-density systems may require direct liquid cooling, rear-door heat exchangers, new chilled-water loops, or a combination that removes part of the heat at the rack while retaining conventional room cooling.
No rack density, total electrical demand, water temperature, cooling topology, or redundancy level has been published. The equipment budget therefore cannot be treated as the total project cost, since the electrical and mechanical work may require separate funding and construction packages.
Procurement and building design must move together
Public procurement creates a sequencing problem for the engineering teams. Facility designers need reliable power and heat data before specifying plant, while computing teams want to retain flexibility until accelerator pricing, availability, and performance are clearer.
Fixing the specification too early can leave the university buying an older platform. Delaying it can prevent engineers from finalising switchgear, pipework, cooling distribution, and commissioning plans. A range-based design may provide flexibility, but excessive spare capacity raises capital and operating costs.
The Eunomia deployment also brings security and governance requirements. A system supporting sovereign generative AI for public administrations will need controlled physical access, network separation, audit trails, data-governance controls, and incident-response arrangements alongside cooling and power resilience.
Commissioning will need to test the complete chain. Full-load runs should demonstrate electrical stability, cooling performance, alarm response, network throughput, controls behaviour, and recovery after equipment or utility failures. Where liquid cooling is used, leak detection, water quality, isolation, and maintenance procedures also form part of acceptance.
Academic facilities operate under different conditions from commercial data centres. Workloads can be uneven, research funding may be tied to fixed project periods, and specialist staff often support several systems at once. Spare parts, firmware, maintenance contracts, and access to vendor engineering can be as important as benchmark performance.
Granada’s climate will influence the heat-rejection design. High summer temperatures reduce the hours available for free cooling and can increase the energy used by mechanical systems, although the final effect depends on water temperatures, plant selection, and operating load.
The wider Spanish research-computing programme is placing increasingly industrial loads inside university buildings. Campuses are adding accelerators and storage faster than many legacy technical spaces were designed to support, turning electrical and mechanical refurbishment into a central part of the computing investment.
Granada’s next milestones are the institutional agreement for the larger system, release of the tenders, confirmation of the selected hardware, and a construction programme for the facility works. The machines may carry the headline performance, but the room around them will determine whether that performance can be delivered reliably.
