Essays on the economics of educational AI, the shifting architecture of the model landscape, and what it means to build for cost-sensitive, sovereignty-conscious environments. New piece every Tuesday.
Teacher workload has moved from a wellbeing footnote to a system-level political variable. Ministries that treat teacher cognition as national infrastructure will out-perform those that treat it as an HR problem.
Every conversation about AI in education assumes curriculum exists as usable data. It doesn't. Until curriculum becomes structured, machine-reasonable intelligence, no AI system can do more than generate around it.
Sovereign AI is becoming a marketing term. Real sovereignty is operational: data residency, weight custody, curriculum ownership, audit rights, exit provisions and offline operation. A procurement-grade definition.
The first phase of commercial AI was API-as-product. The next phase looks more like the shift from time-sharing to owned databases — buyers stop renting intelligence when the intelligence becomes strategic.
$30 billion a year of edtech spend, limited evidence of effectiveness. The diagnosis: the category optimised engagement and called it learning. The next category will be measured on outcomes against curriculum, not minutes on platform.
The frontier-model conversation dominates the headlines, but a quieter shift toward small specialised models, agent orchestration and local inference is reshaping the economics of AI — especially for education and emerging markets.