Training a frontier model in 2025 takes a few hundred million dollars of compute, a few hundred to a few thousand engineers, and access to enough electricity that the build-out is now visible in national power-grid statistics. Inference — the recurring cost of serving the model — is what determines whether the unit economics of an application work.
The cost stack
- Training: $50M-$300M+ per frontier model, mostly GPU time at $2-$4/hour for an H100-class chip
- Inference: $0.0001-$0.10 per thousand tokens depending on model. Frontier flagships are 100x the cheapest open-weight alternatives
- Data: a few million dollars of human-feedback work to fine-tune a model. Quality of feedback labour matters
- Talent: a frontier-research engineer commands $500K-$5M total comp; a senior research scientist often $2M-$10M+
Labour-market impact
Eloundou et al. (2023) estimated that 80% of US workers have at least 10% of their tasks exposed to LLM automation. Brynjolfsson et al. (2023) found a 14% average productivity lift for customer-service agents using a generative-AI assistant, with the largest gains (35%) for the least experienced. The IMF (2024) estimated 60% of advanced-economy jobs face AI exposure, half positively (productivity-enhancing), half negatively (displacement-risk).
African economies have lower current AI exposure (because the labour mix is less white-collar) but face the second-order risk: outsourcing flows that have powered Kenya's BPO sector since the 2000s are exactly the work most exposed to AI automation. The strategic response — moving up the value chain into AI-augmented services — is happening but slowly.
Regulation
The EU AI Act (Regulation 2024/1689) entered into force August 2024 with phased application through 2026. It is the strictest comprehensive regime: bans certain uses (social scoring, real-time biometric surveillance), tightly regulates 'high-risk' uses (employment, credit, education, justice), and imposes transparency obligations on 'general-purpose AI' models.
The US is sectoral: SEC for algorithmic trading, Fed/OCC for banking via SR 11-7, FDA for medical AI, FTC for unfair-practice claims. The Trump administration rolled back the Biden 2023 executive order in early 2025; comprehensive federal legislation remains elusive.
Kenya: the Data Protection Act 2019 already constrains personal-data AI use. The Kenya AI Strategy 2025-2030 is in consultation. CBK has signalled prudential guidelines on AI in banking; CMA has consulted on algorithmic trading. The Office of the Data Protection Commissioner (ODPC) is the most active regulator on the ground today.
The chip supply chain
Frontier AI runs on GPUs. The cutting-edge GPUs are designed by Nvidia (US), manufactured at advanced nodes by TSMC (Taiwan), using lithography machines from ASML (Netherlands), with key chemicals from Japan. Three geographic chokepoints, all subject to geopolitics. The US export controls on advanced chips to China (October 2022, expanded 2023, 2024) are reshaping global AI infrastructure: companies are diversifying supply chains, China is investing tens of billions in domestic capacity, the global lead time on a new fab is five years.
The strategic question for African economies
Africa has no chip industry, modest AI research, but a large young workforce. The strategic choice is whether to be a price-taker (consume AI services priced and built elsewhere) or to invest in the layer where Africa can compete: localised applications, domestic data infrastructure, regulatory leadership, and the human capital to deploy and govern AI well. Kenya's window to position itself is narrow.
Closing the course
We've moved from what AI is, through how it works, to where it goes. The technology will keep changing; the discipline of using it well — knowing what it can do, what it cannot, where the risks lie, what to keep verified — is what compounds. Welcome to the field.
Exercise
A Kenyan fintech is evaluating whether to use OpenAI's API or run an open-weight model (Llama 3, Qwen, DeepSeek) on its own infrastructure for an internal document-processing pipeline that handles ~5 million documents per year. Build the unit economics comparison: what costs go into each option, and what factors beyond raw cost should drive the decision?