Karlsruhe Institute plans a €24 million AI data center for 2030 with a 10MW power option

The Karlsruhe Institute of Technology has announced an AI data center at its Campus North site in Baden-Württemberg, estimated to cost €24 million, or about $27.2 million, with construction due to finish by the end of 2029 and the site opening the following year.

The power envelope runs to 2MW, with an option to scale the supply to 10MW if demand for the facility's services remains high, but the compute side is looser: the project's IT capacity has not been determined, so the university has committed to a building, a power supply and a heat-recovery system ahead of the machine that will occupy them.

The facility is designed to serve other higher education institutions, government organizations and researchers — an arrangement closer to shared research infrastructure than to a commercial lease. The coverage names no tenant, no capacity price and no offtake agreement. The first customer is the institution paying for construction.

Dr. Martin Nußbaumer, a director at KIT's Scientific Computing Center, described a site that runs on two shifts: daytime capacity for AI applications when demand is particularly high, and the same infrastructure at night for resource-intensive simulations such as climate research. His aim, in his words, is a system that can 'breathe' with demand, avoiding idle capacity while strengthening digital sovereignty through AI services that make good use of existing resources.

Nothing in the disclosure covers procurement. There is no named compute vendor, no server count and no accelerator type, which suggests the constraint KIT is planning against is electrical rather than the supply chain for silicon: 2MW today, 10MW if demand materializes, with construction phased so the scale-up can follow.

Two workloads against one capital base is ordinary practice in research computing, and at this scale it is the detail in the announcement that matters most to anyone underwriting commercial compute. A 2MW machine amortized across a daytime AI curve and a night-time simulation curve does not need either workload to carry the full build cost. The €12 million per megawatt implied by the headline figure is a budget number, not a valuation, and the announcement does not say what the step to 10MW would cost.

KIT plans a system that siphons waste heat from the data center's servers into the district heating network, and the facility is to be built next to a combined heat and power plant on Campus North. The heat buyer is the campus itself, which makes the loop a cost reduction rather than a revenue line. Kora Kristof, KIT's vice-president for digitalization and sustainability, framed the design as heating campus buildings while running the data center against sustainability criteria, work she tied to the institution's decarbonization efforts. Whatever the loop is worth, it never leaves the balance sheet that funds the servers, so no third party has to price it.

On power, the announcement stops at the perimeter, with no description of the grid connection, the capacity of the neighboring plant, or how the 2MW will be sourced. The phased build, with the scale-up option held in reserve, is the shape a project takes when its load has to be earned before it is connected. Interconnection queues and load-class rules decide returns before a lease is signed, and a campus that holds an option on 10MW of its own load is holding a power right, whatever it decides to do with it.

Across these three German projects, the common thread is ownership. KIT is building its own facility and will run it; the Caritas installation is owned AI compute on site; Stuttgart is putting up a supercomputer center on its own campus. Institutions that could buy compute by the hour are instead carrying construction schedules, power supply and heat recovery, and the trade being made is control of the building and power against the flexibility of rented capacity.

KIT is not the state's only academic builder. Baden-Württemberg holds 32 data centers today, according to DataCenters.com, and the University of Stuttgart began constructing a supercomputer data center on its own campus in December 2025. In August 2026 a 78kW facility followed at the Caritas Hospital in Bad Mergentheim; BBT Group spent €3.6 million on it and owns the AI compute there. Against 78kW, KIT's 2MW design is roughly twenty-six times the rated load, the distance between a hospital's on-premise compute and a campus-scale research machine.

That leaves the capital question, which this project cannot answer for the market. Assets with a named anchor tenant or a sovereign buyer get infrastructure pricing, while everything else waits for one. KIT's project sits outside that frame: the owner and the user are the same institution, so there is no lease to price, no yield to compare and no merchant risk to place with anyone. It also explains the asymmetry in the disclosure. The €24 million is a budget line the institution controls, while the compute specification is a procurement it has not made.

Watch the gap between 2MW and 10MW. If KIT scales up, the campus becomes a load worth planning a regional grid around and the heat loop becomes a template for research sites with a power plant next door. If it stays at 2MW, the campus gets a machine it can schedule around two daily demand curves, and the 10MW option goes unused.