Across the Global South, artificial intelligence (AI) projects begin with similar patterns. A researcher decides to explore an idea.
The research idea is strong. The budget is limited. The internet connection is sometimes unstable. Access to high-performance GPU hardware is out of reach.
The question is no longer whether the research matters. The question is whether it can run.
This is the reality for many teams building AI systems outside the world's most compute-rich environments. The narrative of modern AI often centres on massive datasets and powerful infrastructure. For researchers and builders working on low-resource languages across the Global South, those assumptions rarely hold.
The result is what researchers have called the low-resource double-bind: the simultaneous constraint of limited data and limited compute. When both are scarce, experimentation slows, innovation narrows, and meaningful participation in global AI research becomes difficult.
In our ICLR 2023 AfricaNLP workshop paper, Adapting to the Low-Resource Double-Bind: Investigating Low-Compute Methods on Low-Resource African Languages, our research team set out to answer a practical question: can useful language models be built under real infrastructure constraints?
The Double-Bind Problem
The term low-resource double-bind captures a structural barrier. Low-resource languages account for a small fraction of available natural language processing (NLP) datasets, while access to graphics processing units (GPUs, specialised chips that accelerate machine learning training) and tensor processing units (TPUs, purpose-built hardware designed for large-scale AI workloads) remains limited for many researchers. Large pre-trained models dominate modern NLP, yet they are computationally expensive to train and fine-tune.
This creates a cycle: limited data restricts performance, while limited compute restricts experimentation. Instead of waiting for ideal conditions, our team embraced the constraints and investigated methods designed specifically for such environments.
A Compute-Aware Alternative: Language Adapters
Unlike traditional fine-tuning which updates every single parameter in a large model and requires significant computational resources, we explored a lighter approach called adapter-based tuning. Adapters are small add-ons placed inside an existing model, allowing it to adjust to new languages or tasks without rebuilding or retraining the whole system.
Adapters are lightweight, require minimal storage, and can be trained rapidly. Crucially, they enable experimentation without the need for high-end infrastructure.
Our experiments focused on training monolingual language adapters for 12 African languages using only free computational resources such as Google Colab. These adapters were then fine-tuned for Named Entity Recognition (a task that identifies names of people, organisations, locations, and other key terms in text) using the MasakhaNER datasets.
What We Found
The results were clear: adapter-based approaches achieved performance comparable to directly fine-tuning a large pre-trained model, despite relying solely on free compute resources.
Across languages, average F1 scores (a balanced measure of how accurately the model identifies and correctly labels entities) for adapter-based models were close to baseline fine-tuned models. This demonstrates that meaningful NLP experimentation is possible even under compute and data constraints.
Adapters also enabled rapid iteration. Because they are lightweight, researchers can train, retrain, and experiment more quickly, reducing the barrier to entry for innovation.
Why This Changes How We Design AI Systems
This research highlights system-level design choices that determine whether AI can perform under constraint and scale sustainably.
Resource-efficient AI is engineered through deliberate architectural choices that align data, models, and infrastructure for sustained performance under constraint. It prioritises approaches that:
- Respect infrastructure realities
- Minimise computational waste
- Enable experimentation in constrained settings
- Lower barriers for local researchers and institutions
The low-resource double-bind is not unique to Africa. It reflects conditions across many parts of the Global South and other compute-constrained environments.
By demonstrating that adapter-based methods can deliver comparable performance without expensive infrastructure, this work contributes to a broader shift: designing AI systems that scale responsibly.
Designing for Constraint as a Strategy
At Lelapa AI, we treat performance under constraint as a core design principle that shapes how we build, ensuring strong performance within real-world limits.
The double-bind forces clarity. It pushes us to ask:
- How can we maximise performance per unit of compute?
- How can we enable experimentation without high capital expenditure?
- How can architectural decisions reduce long-term system costs?
This paper lays the groundwork for those questions.
Looking Ahead
Future work will involve testing this approach on more models, exploring whether closely related languages can strengthen each other during training, and better understanding how data size and energy use affect performance.
These investigations deepen a central idea: scalable AI systems emerge from compute-aware design, not from brute-force scaling.
Designing AI for compute-constrained environments has strengthened our ambition and sharpened our strategic focus. At Lelapa AI, these methods power systems built to perform anywhere, reaching every corner of the world with disciplined, scalable design rather than relying on abundant compute or data.