The Stack — Ep. 02: Sovereign AI and the residency moment

Episode two of The Stack goes deep on the intelligence layer, and specifically on the moment "sovereign AI" stopped being a conference slogan and turned into a set of real, shipping decisions — about where models get built, where data is allowed to live, and which language a system thinks in first. It is a subject wrapped in marketing, so we spend as much time on what is not true as on what is. These are the notes, with the caveats we keep on the record throughout: most of the compute figures below are announced and forward-looking, and every "leading model" banner is a vendor claim we attribute rather than endorse.

The through-line

In roughly eighteen months, the Gulf moved from importing artificial intelligence to producing it, and the shift shows up in three places at once: state-backed compute, frontier Arabic models, and a change in what buyers require — from "does it work" to "where does it run." We trace all three and try to separate the genuine structural change from the press-release theater.

What we cover

• The compute build-out, honestly. Saudi Arabia's PIF-backed HUMAIN and its NVIDIA partnership, projected at up to 500 megawatts over five years; the UAE's Stargate cluster inside a 5-gigawatt Abu Dhabi campus, with a first 200-megawatt phase expected live in 2026. And the number that actually matters, which is not headline capacity but utilization — how much of it is energized and doing paid work.
• Arabic models grow up. Jais 2, pretrained from scratch rather than adapted from an English base; Falcon-H1 Arabic; Saudi Arabia's ALLaM. Why "built as Arabic" matters far more than any single leaderboard ranking, and why we refuse to referee the "world's leading Arabic model" contest.
• The residency moment. How "where does the model run" shifted from a nice-to-have to a procurement requirement across regulated Gulf industries, and why domestic compute is what makes region-resident AI practical.
• What "sovereign" honestly means. The paradox of building sovereign AI on foreign chips under foreign export rules — which is why we treat the word as a direction of travel, not a finished state. We keep the geopolitics strictly factual.

Takeaways

1. The region now produces frontier Arabic models, not translations. That structural fact outlasts whichever model is momentarily on top.
2. Residency is becoming a buying requirement, which quietly favors builders who own the region-resident layer rather than renting a model that runs somewhere else.
3. Watch utilization, not megawatts. Announced capacity is easy. Energized, productive capacity is the honest signal, and it lags the headlines.

Region-grade AI is not English AI with an accent. It is built for the language, the culture, and the law — and increasingly, built where those things live.

Next in the series: the money layer, and why self-custody has to become the default rather than the enthusiast's choice.

Compute and model references are as announced by the named entities (NVIDIA/HUMAIN, G42/SoftBank/Stargate, MBZUAI/Inception, TII, SDAIA) and are frequently forward-looking; superlatives are attributed to their sources, not endorsed.