Sizing a sovereign, air-gapped AI stack for oil and gas health, safety and environment (HSE) in Pakistan
Описание и применение
Дизайн автономной AI‑стека для нефтегазовой отрасли в Пакистане, где все данные и модели остаются внутри инфраструктуры оператора, обеспечивает предупреждения о происшествиях и экономит около 670 000 USD за три года по сравнению с облачными решениями.
Задача
An oil and gas operator in Pakistan needs AI that warns before incidents, not reports after, while keeping all data and models within its own infrastructure and avoiding third‑party AI APIs.
Результат
The design demonstrates that a sovereign, air‑gapped AI stack can be built on the operator’s own hardware, providing real‑time warnings while keeping data local, with a three‑year cost of about USD 670,000 versus USD 1.85–2.82 million for cloud equivalents.
Рабочий процесс
- 1. Select self‑hosted models based on licence to ensure internal ownership.
- 2. Size the central tier by calculating weight size, KV cache and activations to determine GPU nodes needed.
- 3. Deploy edge nodes for real‑time detection and forecasting, reusing existing NPU compute.
- 4. Set up read‑only adapters, Kafka on KRaft, and a unified clock with Chrony and GNSS.
- 5. Estimate costs and compare to cloud options.
- 6. Address export licence requirements for 141 GB HBM GPUs.
- 7. Publish the object model as JSON for reuse.
Доказательства
An oil and gas operator in Pakistan wants one thing from AI in its health, safety and environment department: a warning before the next incident, not a report after it.
None of it may leave the operator's own infrastructure, and no third-party AI API may sit in the serving path.
GLM 5.3 open weights, 753B mixture-of-experts at FP8
Take the largest filed parameter count, multiply by bytes per parameter at the serving precision, then add a planning factor for the KV cache and activations so long incident histories fit
Three years, public prices, electricity at Pakistan's B3 industrial tariff: Own the hardware, with support and power about 670,000
- Дата записи: 2026-10-03T07:41:08.087Z