Gottfried Wilhelm Leibniz University Hannover is announcing the procurement of GPU and CPU computing power for the project "EU-SAI" to train a large language model (Large Language Model, LLM) with at least 100 billion parameters and at least 12 trillion tokens.

LOT-0001
Leibniz University Hannover - Procurement of GPU and CPU Computing Power for Training a LLM.
As part of a state-funded project running until June 30, 2026, the Gottfried Wilhelm Leibniz University Hannover (LUH, client) is issuing this Europe-wide tender for the procurement of GPU and CPU computing power for training a Large Language Model (LLM) with at least 100 billion parameters and at least 12 trillion tokens, including additional tasks as outlined in the tender documents.

The binding minimum requirement for GPU computing power is 3.133 × 10²⁵ FLOPs. This corresponds to approximately 8.8 million GPU hours on an NVIDIA H100 (H100 80 GB SXM5). In addition to GPU capacity, the client requires 7.5 million CPU core hours for supplementary preprocessing, orchestration, and supporting workloads.

FLOPs are defined as the product of (number of delivered GPU hours) x (non-sparse BF16 tensor core throughput per GPU as specified by the manufacturer).

All requirements defined in this tender arise directly from the client's goal of training a model with at least 100 billion parameters on at least 12 trillion tokens within 8-10 months.

Only NVIDIA GPUs are authorized for the training of the model and the tasks to be fulfilled (H100, H200, or later generations such as B100, B200, or GB200). This restriction is a technical necessity and does not represent a preference for any particular vendor. The training framework and software stack of the contracting authority are exclusively validated and supported on NVIDIA hardware. In particular, the existing model training pipeline (including CUDA, NCCL, PyTorch/Megatron-LM, and associated optimizations) is closely integrated with NVIDIA toolchains. For alternative GPU vendors, there is no equivalently validated environment available within the timeframe of this procurement. Porting, testing, and validation on non-NVIDIA hardware would require extensive development efforts that are incompatible with the available timeline. Therefore, the restriction of this procurement to NVIDIA GPUs is essential to ensure feasibility, stability, and timely delivery.

The performance period is 8-10 months starting from November 2025. All services must be provided within the EU/EEA or in countries with an adequacy decision by the EU Commission.

For further information, please refer to the service description