The Client
Fatimah Karimi is a data scientist and AI researcher working independently in Abuja. She consults for Nigerian financial institutions on credit scoring and risk model development, and she runs her own research agenda in parallel — publishing work on applying machine learning to Nigerian economic data that has attracted attention from academic institutions in Africa and abroad. She is prolific, technically rigorous, and deeply conscious of the economics of her research infrastructure.
The economics, by late 2025, had become a problem. Fatimah was training her models on cloud GPU instances — primarily Google Colab Pro and AWS p3.2xlarge. Between the two, she was spending approximately ₦400,000 per month on compute. For a solo practitioner, that's a significant recurring cost that does not generate direct revenue — it's research overhead that erodes margins on every consulting engagement.
The Challenge
The cloud compute dependency had practical problems beyond cost. Network latency to cloud instances made interactive model development — iterating on architecture, debugging training loops, running quick explorations — slow and frustrating. A model that failed after 10 minutes of training on a cloud instance had wasted 10 minutes plus queue time plus upload time plus download time. Local compute removes the network layer entirely.
Fatimah's training workloads were primarily PyTorch — transformer models, gradient boosting pipelines for tabular financial data, and occasional computer vision work for document processing applications. The GPU memory requirements ranged from 8GB for smaller experiments to 24GB for her largest models. Her CPU workloads — data preprocessing, feature engineering, cross-validation — benefited from high core counts.
She also had a specific requirement: the machine needed to run with no user intervention when she initiated training jobs, including through overnight runs, NEPA cuts, and generator transitions. She could not be monitoring hardware all night when a 12-hour training job was running.
The Consultation
We worked through the ROI calculation with Fatimah explicitly. At ₦400,000/month in cloud compute, a local workstation paying for itself in under 12 months required a total build cost under ₦4.8 million. We knew we could build a machine that exceeded her cloud instance performance for significantly less than that.
The key specification decisions:
- GPU: NVIDIA RTX 4090 24GB — the 24GB VRAM handles her largest model architectures; CUDA performance exceeds a p3.2xlarge's V100 16GB in most PyTorch benchmarks
- CPU: AMD Ryzen 9 7950X — 16 cores for data preprocessing pipelines and cross-validation parallelism
- RAM: 128GB DDR5 — holding large financial datasets in memory eliminates constant re-reads from storage during training loops
- Power: Online UPS with automated machine shutdown/resume capability — we configured unattended-upgrades and UPS management software (NUT) so the machine would safely hibernate on low battery and resume when power returned
The Build
AI Research Workstation — ₦4.2 million:
- CPU: AMD Ryzen 9 7950X — 16 cores, 5.7 GHz boost; excellent for pandas, scikit-learn, and data preprocessing
- RAM: 128GB DDR5-5600 ECC — error-correcting memory for long unattended runs; data integrity in extended training jobs
- GPU: NVIDIA RTX 4090 24GB — 16,384 CUDA cores, 1,008 GB/s memory bandwidth; significantly outperforms V100 in PyTorch
- Storage: 4TB Samsung 990 Pro NVMe — training data, model checkpoints, research datasets, all on NVMe speed
- Storage (backup): 8TB HDD in RAID configuration with cloud sync to Backblaze B2
- UPS: APC Smart-UPS 3000VA online double-conversion with USB management port
- Software: NUT (Network UPS Tools) configured for automatic safe shutdown at 20% battery, with resume on power restoration
- CUDA/PyTorch: Installed and benchmarked on delivery; confirmed working with her existing research codebase
- Remote access: Configured Tailscale VPN so she can SSH into training jobs from anywhere
The Result
Fatimah's first benchmark comparison: a transformer model she regularly trained on AWS p3.2xlarge (V100 16GB, ~45 minutes per epoch) trained in 28 minutes per epoch on the RTX 4090. Faster than her cloud instance, at no marginal cost per run.
Her cloud compute spend in the first month after installation: approximately ₦35,000 — for tasks that genuinely required cloud-scale or distributed compute. A 91% reduction. At that rate, the machine pays for itself against avoided cloud costs in 3.9 months. She reached cost neutrality 17 weeks after delivery.
The automated UPS management has worked correctly through six NEPA events during training runs. On each occasion, the system safely checkpointed and hibernated before the UPS battery depleted, then resumed automatically when power returned. She has woken up to completed training jobs — not failed ones — every time.
"I didn't realise how much the cloud friction was slowing down my research," she told us. "When iteration is fast and free, you just try more things. The research output has genuinely increased."
Key Takeaway
For data scientists and ML researchers who run regular model training, the economics of a local GPU workstation versus cloud compute often favour local hardware within the first year. The RTX 4090's performance per naira — compared to cloud GPU instance pricing in Nigeria — is compelling. The non-monetary benefits are equally significant: no network latency, no queue time, no per-experiment cost friction that discourages iteration. If you are spending more than ₦200,000 per month on cloud compute for training workloads, it is worth modelling the build cost against your current spend before renewing your cloud commitment.
Are you an AI practitioner or data scientist paying for cloud compute monthly? See the AI Series or talk to our team about the economics of a local workstation for your training workloads.