This is an illustrative scenario, not an account of a specific client engagement. Consider a research group at a Nigerian federal university spending more on cloud compute per month than their department's entire equipment budget. They train NLP models on cloud GPUs because nothing local can handle the workload. That is not sustainable.
The requirement is specific: a single workstation capable of training mid-size transformer models (up to 7B parameters with quantisation), running overnight training jobs without cloud dependency, and supporting three concurrent researchers.
The Build
The spec should follow the group's actual training parameters rather than defaulting to "most powerful GPU available":
- AMD Ryzen 9 9950X (for data preprocessing, tokenisation, and parallel CPU tasks)
- 128GB DDR5 ECC-compatible RAM (large dataset loading without paging to disk)
- NVIDIA RTX 4090 24GB (24GB VRAM is the meaningful threshold for this workload — allows full 7B model fine-tuning without offloading)
- 2TB NVMe PCIe 5.0 (fast enough to feed training data without becoming the bottleneck)
- 4TB NVMe PCIe 4.0 (model checkpoints and datasets)
- 1200W 80+ Platinum (the 4090 can pull 450W sustained under training load)
The Impact
The 24GB of VRAM on the RTX 4090 is the load-bearing decision here — it is the threshold that allows full 7B model fine-tuning without offloading, and offloading is what turns a training run from long into impractical. Against rented cloud GPU time, a workstation is a fixed cost that stops accruing.
The bigger change is behavioural: researchers can iterate freely. They are not managing cloud budgets per experiment. The machine runs overnight. They check results in the morning. Whether the hardware pays back faster than cloud spend depends on your own utilisation — it is worth working out against your actual bills.