Engineering & Architecture: Build Decisions This Week
2026‑09‑21
The past week has delivered a mix of lessons from the banking sector’s mobile playbook, a wave of AI infrastructure IPOs, and an inside look at OpenAI’s new agentic software factory. For CTOs juggling South African and UK/EU ecosystems, the key question is: Which of these signals should prompt an immediate architectural rethink, and which can be shelved for now? Below are three concrete build decisions worth evaluating today.
---
OpenAI’s deep dive into its agentic software factory shows a paradigm where the large language model (LLM) orchestrates sub‑agents, maintains state, and invokes external APIs in real time. The architecture collapses dozens of micro‑services into a single token‑driven loop.
| Pros | Cons |
|----------|----------|
| Rapid prototype cycles – internal bots such as PR reviewers or code linters can be spun up in minutes rather than weeks. | Compute cost per token – unlimited token usage scales linearly with demand, potentially eroding savings from traditional micro‑service billing. |
| Centralised observability – all calls flow through one trace, simplifying debugging. | PII risk – if tokens contain personal data, unsanitised storage or transmission can breach POPIA (SA) and GDPR (UK/EU). |
| Reduced operational overhead – a single entry point replaces many CI/CD pipelines. | Outbound traffic constraints – data‑protection regulations demand tight control over external calls; a universal gateway may need complex policy enforcement. |
Action: Conduct a cost–benefit exercise on a low‑risk feature (e.g., automated documentation generation). Compare the token‑based approach against a conventional micro‑service in terms of latency, total cost of ownership, and compliance audit effort.
---
Capitec Connect and FNB Connect are following the same mobile virtual network operator (MVNO) playbook, offering mobile services directly to customers. The lesson is not that MVNOs are a good fit, but that copying a bank’s identical strategy may be too derivative.
| Pros | Cons |
|----------|----------|
| Unified customer experience – customers receive banking and connectivity in one subscription. | Competitive duplication – Capitec and FNB are essentially selling the same product, reducing differentiation. |
| Data monetisation opportunities – network usage can feed predictive models for credit decisions. | Regulatory scrutiny – bundling financial services with telecom can attract oversight from both the CIPC (SA) and UK’s Competition & Markets Authority. |
Action: Prioritise niche verticals where connectivity data provides unique business insight—e.g., logistics, agritech—rather than a blanket MVNO strategy. This keeps bandwidth constraints in mind while avoiding direct head‑to‑head competition.
---
London-based Nscale’s IPO filing illustrates the appetite for managed AI infrastructure that can elastically serve GPU‑heavy workloads. For projects such as Unit1’s hyper‑realistic digital avatars, scaling graphics rendering and real‑time simulation is a major cost driver.
| Pros | Cons |
|----------|----------|
| GPU auto‑scaling – automatically add compute when user demand spikes (useful for live event streaming). | Vendor lock‑in – reliance on Nscale’s proprietary orchestration could hinder migration to other providers. |
| Reduced operational overhead – abstracts low‑level GPU management and driver tuning. | Cost transparency – pay‑per‑usage models may expose hidden fees once workloads become steady. |
Action: Run a proof‑of‑concept with a small subset of avatar rendering jobs on Nscale, measuring both latency and cost against an in‑house GPU cluster. Include compliance checks: ensure data leaving the SA data centre is GDPR‑compliant if user content crosses borders.
---
---
By focusing on these three build decisions—agentic orchestration, differentiated MVNO strategies, and managed AI infra—you’ll align your architecture with the most impactful shifts observed this week, while keeping security and compliance front‑of‑mind.
---
Sources
The recommendation to adopt an agentic software factory hinges on the assumption that the organization has a token‑budgeting model and robust PII sanitisation procedures in place. Validation of cost modeling against actual token usage will be required, as will confirmation that outbound traffic controls can enforce jurisdictional compliance at scale.